PacketNebula

Articles & guides

Explainers, how-to guides and the occasional news note. The answer comes first, always, then the numbers, diagrams and sources for going deeper.

Official Cursor key art for Origin, showing the Origin cube mark and wordmark over a blurred hillside at dusk, from the Origin product page.

Dev News

Cursor Origin: what the GitHub rival actually ships

Cursor opened its Origin git forge to all paid plans on 17 August, the day GitHub broke. What it does, what it is missing, and who holds your code now.

 Answer card stating that Google won a bankruptcy auction for the internal enterprise data of the defunct Spirit Airlines with a 10 million dollar bid, beating a 7.5 million dollar bid from Mercor, that the package covers about 100 million employee emails and 500 million Microsoft Teams messages, and that the approval hearing in the Southern District of New York moved to 9 September 2026 after the flight attendants union objected.

Dev News

Google's $10M Spirit Airlines Data Buy Slips to September 9

Google won a bankruptcy auction for Spirit Airlines' internal emails and Teams history at $10 million. The cabin crew union objected, so approval slipped to 9 September.

 Product render of the Cerebras CS-4 rack on a grey studio background, a black cabinet with the Cerebras logo and three horizontal Wafer-Scale Backpack modules slid partway out of the left side, each showing dense rows of orange-tipped connectors, with power and liquid cooling hardware on the right.

Dev News

Cerebras CS-4: 750 PFLOPS from an overclocked wafer

Cerebras' CS-4 packs three WSE-3 Turbo wafers into one rack for 750 PFLOPS. The die is unchanged. Here's what the 30x claim really measures, and what's missing.

 Answer card summarising Groq's $350 million Series A of 17 August 2026, reported at a $3.5 billion valuation, led by Disruptive with Nvidia planning to participate, five days after Groq became an Nvidia Cloud Partner.

Dev News

Groq's $350M reset: from Nvidia rival to Nvidia partner

Groq closed a $350M Series A on 17 August, reportedly at $3.5B, with Nvidia joining the round. The LPU challenger now sells Nvidia capacity. What that changes.

 Answer card summarising the reported Stripe acquisition of OpenRouter for more than $7 billion on 16 August 2026, unconfirmed by either company, and OpenRouter's published pricing of no inference markup with a 5.5% fee on credit top ups.

Dev News

Stripe's $7B OpenRouter deal: what changes for your keys

Bloomberg says Stripe finalised a deal to buy OpenRouter for over $7B. Nothing is confirmed and nothing changed on your API keys. What Stripe is really buying.

 Answer card: the 50 percent weekly usage limits promotion in Claude Code ends at 11:59 PM PT on 19 August 2026 for Pro, Max and Team plans, returning weekly limits to standard levels, which removes a third of the headroom rather than half, with no change to plans, billing or the 5 hour session window.

Dev News

Claude Code weekly limits drop a third after August 19

The +50% weekly limits promotion in Claude Code ends at 11:59 PM PT on 19 August 2026. That is a third off your headroom, not half. What changes, what does not.

 Anthropic announcement illustration for the Claude text watermarking post: an ornate quill pen resting across a detailed engraved hand, on a muted heather background.

Dev News

Claude's text watermark: no opt-out, and no detector yet

Anthropic marks every Claude output with a SynthID style watermark. Global, no opt-out, weakest on code and short text, and the public detector has not shipped.

 Official joint announcement card from SB Energy showing three photographs across the top, a groundbreaking ceremony with excavators holding a large US flag over a group in hard hats beside SoftBank and SB Energy banners, an aerial view of a large industrial campus, and two women at a community jobs event, above a dark navy panel carrying the NVIDIA, SB Energy and OpenAI logos.

Dev News

Nvidia's $105B OpenAI guarantee: a lease, not a loan

Nvidia disclosed residual value guaranties capped at $105 billion behind OpenAI's Ohio leases, next to $1.5 billion of actual equity. Here is what that obligates.

 Answer card: Nvidia disclosed a 63.44 billion dollar US equity portfolio for the quarter ended 30 June 2026, of which Intel at 29.99 billion and SpaceX at 20.98 billion make up 80.3 percent, with a large zero marking the number of shares bought or sold in the seven positions carried over from the previous quarter.

Dev News

Nvidia's Q2 13F: $30B Intel, $21B SpaceX, zero buys

Nvidia's 13F for the June quarter shows a $63.44 billion equity book, up 3.4x. Every carried share count is unchanged, so the growth is marks plus one conversion.

 Official Quantinuum announcement card for the Oracle partnership: the Quantinuum wordmark and logo above the word AND in mint green, with the Oracle logo below, set on a pale mint background with faint wireframe globe arcs in the corners.

Dev News

Quantinuum Helios on Oracle Cloud: 98 qubits, $8M revenue

Quantinuum is installing a 98 qubit Helios inside an Oracle Cloud AI data center. No price, no availability date, and $8.0 million of quarterly revenue behind it.

 Answer card: Alibaba published the Qwen3.8-27B weights on 14 August 2026 under Apache 2.0 as a dense native vision language model with 262,144 tokens of native context, while the 2.4 trillion parameter flagship weights published two days earlier carry a custom qwen3.8-max licence and are text only.

Dev News

Qwen3.8-27B ships Apache 2.0 with vision, the 2.4T doesn't

Alibaba's small open model got the permissive licence and a vision encoder. The 2.4T flagship weights are text only under a custom licence. What that changes.

 Answer card: Microsoft began merging the consumer Copilot app and Microsoft 365 Copilot into a single application on 13 August 2026, with mobile and web worldwide in mid August and Windows and Mac in mid September, and retires Copilot Group Chat, Podcasts and consumer Deep Research on 18 August 2026 while leaving work and school data boundaries unchanged.

Dev News

Copilot's one app merge: what goes away on August 18

Microsoft folded its two Copilot apps into one. The merge is automatic. Group Chat, Podcasts and consumer Deep Research retire on 18 August.

 Writer product screenshot from the Palmyra X6 announcement, showing a model picker with Palmyra X6 selected above Palmyra X5, next to an AI visibility audit playbook dashboard listing agent runs with status, latency in seconds, average cost and total spend per row.

Dev News

Palmyra X6 runs on GLM-5.2, priced $2 in and $8 out

Writer's new flagship is post-trained GLM-5.2 from Z.ai. It scores 0.87 against Opus 4.8's 0.86, and that number comes from Writer's own nine evaluations.

 Z.ai official benchmark chart for GLM-5.3, six grouped bar charts comparing GLM-5.3, GLM-5.2, Kimi K3, Fable 5 and GPT-5.6 Sol on Terminal Bench 3.0, DeepSWE, Agents Last Exam, AutomationBench, HLE with Tools and GDPval-AA v2.

Dev News

GLM-5.3: open weights in two weeks, breaking API change now

Z.ai shipped GLM-5.3 on 14 August. The weights land in about two weeks, and thinking can no longer be disabled, so swapping the model ID breaks non-thinking calls.

 Cerebras launch artwork for the Ultrafast tier, dark background with orange and blue light trails and the title Accelerating GPT-5.6 Sol Ultrafast with OpenAI.

Dev News

GPT-5.6 Sol Ultrafast: 14x on the chart, 5.6x end to end

OpenAI previewed Ultrafast on 13 August, a Cerebras backed API tier running GPT-5.6 Sol at up to 750 tokens a second. End to end it is 5.6x, and no price is published.

 A Verne-branded white robotaxi driving past the Croatian National Theatre in Zagreb, from the Verne press material for Europe first commercial robotaxi service, which runs Pony.ai autonomous driving technology.

Dev News

Uber and Pony.ai: 2,000 robotaxis, four cities unnamed

Pony.ai and Uber announced more than 2,000 robotaxis across five European cities on 14 August 2026. No city names, no dates, and the fleet it grows from is 10 cars.

 The official xAI announcement card for Grok 4.6, showing the words Grok 4.6 in white on a dark grey gradient with a pale swirl of light.

Dev News

Grok 4.6: same $2/$6, and a 200K pricing cliff

xAI shipped Grok 4.6 on August 12 at the same $2/$6 as 4.5, plus five index points. Two costs moved quietly: the cache rate, and a 200K repricing cliff.

 Answer card explaining that on 12 August 2026 Twitch added a Training for Generative AI setting, enabled by default, allowing Amazon to train generative AI content models on channel streams, videos on demand, clips, chat, text and images, found under Settings then Security and Privacy, with the opt-out applying only to future training.

Dev News

Twitch's AI training toggle is on, and it covers your chat

Twitch switched on Amazon genAI training by default on 12 August. The toggle hides in Security and Privacy, and your setting decides for everyone in your chat.

 Google's official announcement image for Gemini 3.7 Flash, showing the Gemini spark logo rendered in blue and purple gradient light on a dark background.

Dev News

Gemini 3.7 Flash: half price, and it doubles on 1 January

Google shipped Gemini 3.7 Flash on 13 August at $0.75/$3.75 per million tokens. The pricing page says that rate runs out on 31 December 2026.

 NVIDIA launch artwork for Nemotron 3.5 Lightning, the model name set over a dark green graphic with a stylised lightning motif, from the NVIDIA blog post announcing the release.

Dev News

Nemotron 3.5 Lightning: 3B active, 350 tokens a second

NVIDIA's 30B mixture-of-experts model runs on 3B active parameters, and the post-training data shipped too. The speed survives outside NVIDIA. The context claim does not.

 Answer card summarising the ChatGPT desktop app for Linux public preview of 11 August 2026: DEB and RPM packages for x64 and ARM64, five tested distributions, and one window holding ChatGPT, ChatGPT Work and Codex.

Sysadmin News

ChatGPT desktop for Linux: the .deb adds an OpenAI apt repo

OpenAI shipped a native Linux ChatGPT app on 11 August as DEB and RPM. The DEB writes an OpenAI repository into apt, and the bundled Codex runs commands.

 Answer card: DeepSeek replaced the weights behind deepseek-v4-pro with DeepSeek-V4-Pro-0813 on 12 August 2026, keeping the same endpoint and the same price of 0.435 dollars per million input tokens and 0.87 per million output, with no change log entry and no 0813 weights on Hugging Face.

Dev News

DeepSeek V4 Pro 0813 is live, at 3x the Flash price

DeepSeek pointed deepseek-v4-pro at the 0813 build on 12 August. Same endpoint, same price, no change log entry, and no 0813 weights on Hugging Face.

 Answer card: Intel priced 20 billion dollars of common stock on 10 August 2026, upsized from the 15 billion dollar offering proposed that same morning, its first public share sale since the 1971 listing, with a large zero marking the number of external 14A foundry customers named in the filing.

Dev News

Intel's $20B share sale: 14A still has no named customer

Intel priced a $20 billion stock sale on 10 August, its first since 1971. The filing names no 14A foundry customer, and that is the number that actually matters.

 Answer card: Anthropic, Macquarie Asset Management and GIC announced Theseus Infrastructure on 10 August 2026, a platform owned by Macquarie managed funds and GIC that will develop and lease purpose-built data centers to Anthropic as anchor tenant under long term agreements, with an initial United States focus and zero dollars, megawatts, sites or dates disclosed.

Dev News

Anthropic's Theseus: a data center deal with no numbers

Anthropic, Macquarie and GIC launched Theseus Infrastructure on 10 August. No capital figure, no megawatts, no sites. The structure is the actual news.

 Answer card: on 10 August 2026 NVIDIA signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish AI compute financing platforms targeting over 500 billion dollars of third-party capital, raised through special purpose entities secured on the compute, with NVIDIA offering residual value support on up to 25 percent of an opportunity.

Dev News

NVIDIA's $500B compute financing and the 25% backstop

NVIDIA signed MOUs with six asset managers on 10 August to mobilize over $500 billion for AI compute. It may also cover 25% of the residual value risk.

 Answer card: Bloomberg reported on 8 August 2026 that Situational Awareness invested 400 million dollars into Source Foundry, 500 million in total, at a reported 5 billion dollar valuation, for a stealth startup building lithography tools aimed at the extreme ultraviolet market ASML dominates.

Dev News

Source Foundry: $400M and a $5B bet against ASML

Situational Awareness put $400 million into Source Foundry at a $5 billion valuation. The startup wants to build EUV lithography tools. Here is what that money buys.

 Answer card: from 14 August 2026 auto mode becomes the default permission mode for new Claude Code sessions on Pro, Max and Team plans, replacing the per action approval prompt with a separate classifier model, with the classifier overhead no longer billed on those three plans while Enterprise and API accounts keep paying.

Dev News

Claude Code auto mode is the default from August 14

Auto mode becomes the default permission mode in Claude Code on 14 August for Pro, Max and Team. What the classifier blocks, what it misses, and who still pays.

 Meta announcement artwork for Muse Glimmer, an abstract network of blue lines and nodes on a dark background, from the Meta research blog post introducing the model.

Dev News

Meta Muse Glimmer: the 30B agent needs 24GB of VRAM

Meta released Muse Glimmer on 10 August: 30B dense, Apache 2.0, open weights. It does run locally, but the floor is a 24GB card and the headline speed is not yours.

 The Department of Energy seal next to the Arcee AI mark on a teal field, the announcement image Arcee published for the Genesis-Science-1 partnership with DOE.

Dev News

Genesis-Science-1: DOE Opened a Portal, Not a Model

The DOE launched its Genesis Open Models Initiative on 7 August, first model Genesis-Science-1. No size, no license, no weights yet. One real deadline.

 A Wayve autonomous Ford Mustang Mach-E in Uber livery on a London street, from the Wayve press release announcing that Transport for London granted private hire vehicle licences to its self-driving cars.

Dev News

Wayve and Uber cleared in London: 15 cars, driver inside

TfL granted private hire licences to Wayve self-driving Ford Mustang Mach-E cars on 5 August 2026. Up to 15 of them, one year each, human driver on board.

 Answer card: the Financial Times reported on 7 August 2026 that ByteDance is pretraining a mixture of experts model of up to 10 trillion total parameters on roughly 30,000 GPUs over three to six months, with no model name, no active parameter count, no chip and no release date disclosed.

Dev News

ByteDance's 10T model is pretraining, not a product

The Financial Times says ByteDance is pretraining a model of up to 10 trillion parameters on about 30,000 GPUs. No name, no active count, no date, no confirmation.

 Answer card: on 6 August 2026 AMD signed a definitive agreement to acquire Taalas, a Toronto startup whose HC1 chip etches Llama 3.1 8B weights into TSMC 6nm mask layers instead of holding them in HBM, with terms undisclosed and closing expected in the fourth quarter of 2026.

Dev News

AMD buys Taalas: the HC1 etches Llama 3.1 into silicon

AMD is acquiring Taalas, whose HC1 chip burns Llama 3.1 8B weights into TSMC 6nm mask layers instead of HBM. What that buys, and what it costs you.

 SpaceX render of the planned Terafab semiconductor complex in Grimes County, Texas, showing an extremely long low rectangular building stretching across flat open land under a clear sky.

Dev News

Terafab: what Tesla and SpaceX's $16.8B first phase buys

Tesla and SpaceX confirmed Terafab in Grimes County, Texas. $16.8 billion for phase one, 3,000 jobs, no process node and no date. What is real and what is a slide.

 NIST physicist Yicheng Shi, lead author of the study, working at the optical bench used to generate and measure the entangled photon pairs sent across 62 km of commercial fiber between Gaithersburg and College Park, Maryland.

Network News

Entanglement Ran 62 km of Aerial Fiber at 92.8% Uptime

NIST, UMD and Qunnect pushed entangled photons down 62 km of weather-exposed commercial fiber for a full day. The rate is 1,500 pairs a second. Read that twice.

 Answer card: IonQ announced a $28 million DARPA contract extension on 6 August 2026 under the It is About Time program, with a further unexercised $30 million option, together covering 125 Evergreen-05 optical atomic clocks, a 5-litre unit delivering 50 femtoseconds of stability at one second and nanosecond holdover over 10 days.

Dev News

IonQ Evergreen-05: DARPA funds 25 clocks, not 125

IonQ's updated DARPA release promises 125 Evergreen-05 atomic clocks. The signed $28M covers 25 of them. We read the structure, and the specs.

 Official OpenAI launch artwork for the ChatGPT update: a large white pill-shaped slider on a starfield, its blue fill stopping at the third of five stops, labelled 5.6 High.

Dev News

GPT-5.6 Sol in ChatGPT is not the Sol in your API calls

OpenAI retuned GPT-5.6 Sol for ChatGPT chat only on 6 August. Work, Codex and your API calls keep the old one. Free users move to Luna with unlimited chats.

 Answer card: on 5 August 2026 Anthropic confirmed it is building an in-house custom silicon team to co-design chips and Claude models, while keeping its multi-chip approach across AWS, Google, Nvidia and AMD, with no tape-out date, no foundry partner and no product announced.

Dev News

Anthropic's custom silicon team: no chip, no timeline

Anthropic confirmed an in-house silicon team on 5 August 2026. The roles are public, first silicon is future tense, and AWS, Google, Nvidia and AMD all stay.

 Official Meta launch artwork for Muse Code and Muse Spark 1.2: a fan of blue lines on a pale background converging into one line that ends inside a small ring.

Dev News

Meta Muse Code: Muse Spark 1.2 and the 21x cheaper tier

Meta shipped Muse Code in beta on 5 August, a terminal coding agent on Muse Spark 1.2. It tops none of Meta's own charts. The pitch is a 21x cheaper tier.

 Answer card: on 5 August 2026 Demis Hassabis became Chair of Google DeepMind and Chief Scientist of Alphabet, Koray Kavukcuoglu became SVP reporting to Sundar Pichai over models and the Gemini app and developer teams, and Chief Scientist Jeff Dean left after 27 years with three colleagues to found Discovery Loop.

Dev News

Jeff Dean leaves Google: what changes for Gemini devs

Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le are off to Discovery Loop, and Hassabis steps back. What it does to your Gemini integration: nothing yet.

 Official Unitree product render of a silver and black G1 humanoid robot mid stride against a plain grey backdrop, arms swinging and one foot lifted.

Dev News

Unitree IPO prices at $9B: 73.6% of humanoids go to labs

Unitree priced its Shanghai IPO at 150.80 yuan, valuing it near $9 billion. The prospectus says 73.6 percent of its humanoid revenue came from research and education.

 Official Mistral AI announcement thumbnail for Shieldstral, with the model name set over the Mistral gradient artwork.

Dev News

Shieldstral 1.0: the 3B guard model that ties a 20B

Mistral shipped Shieldstral 1.0 on 4 August, a 3B open-weights moderation model for text and images. It ties a 20B on text F1 and leads on multimodal.

 Official xAI announcement artwork for Grok Voice Think Fast 2.0, the model name set over the dark xAI release graphic.

Dev News

Grok Voice Think Fast 2.0: grok-voice-latest costs 60% more

xAI moved the grok-voice-latest alias to Think Fast 2.0 today. Same connection string, $0.05 to $0.08 a minute. What you get for it, and whether the speed pays.

 Answer card: Google begins removing Google Assistant from Android phones, tablets, Wear OS watches, compatible headphones and projected Android Auto on 4 September 2026, a rollout expected to take a few weeks, after which the device cannot switch back to Assistant.

Dev News

Google Assistant shuts down September 4: what you lose

Google starts removing Assistant from phones, watches, headphones and Android Auto on 4 September 2026. There is no switching back, and Gemini has gaps.

 Answer card: Qwen3.8-Max is not open source today; it is API only, the weights are promised next week on Hugging Face and ModelScope, and no licence has been named.

Dev News

Qwen 3.8 Max: is it open source? Not yet

Headlines call Qwen3.8-Max open source. Today it is API only, the weights are promised next week, and no license has been named. What that really means.

 Official MiniMax H3 announcement banner on a purple jellyfish background, with the model name and the subtitle Next-Generation Open-Weights General-Purpose Multimodal Video Model.

Dev News

MiniMax H3 open weights: not licensed in the EU, UK or US

MiniMax put the 33B H3 video weights on Hugging Face on August 3. The license names the EU, the UK, South Korea and the US as excluded territories.

 Answer card: Bending Spoons agreed on 4 August 2026 to buy Airtable in an all-cash deal at an enterprise value of 1.285 billion dollars, about 2.25 billion in equity value, closing expected before the end of 2026, with no announcement about pricing or the free plan.

Dev News

Bending Spoons buys Airtable for $1.285B: what changes

Bending Spoons is buying Airtable for $1.285 billion, all cash. The deal is confirmed. What happens to your bill is not, so here is the track record.

 Answer card: Bloomberg reported on 2 August 2026 that the M5 MacBook Air is in major shortage two months after Apple raised its price to $1,299, with US delivery estimates of two to six weeks and the higher memory configurations waiting longest.

Dev News

MacBook Air M5 shortage: $1,299 and a six week wait

Apple raised the MacBook Air to $1,299 in June and still cannot keep it in stock. US delivery runs two to six weeks, and the higher RAM configs wait longest.

 Answer card explaining that a Los Angeles County Superior Court judge granted Santa Monica a preliminary injunction on 27 July 2026 barring Waymo from operating its charging facilities at 1222 and 1310 Broadway between 11pm and 6am, on a public nuisance claim, with a status conference set for 23 October 2026.

Dev News

Waymo's Santa Monica depot ban: 11pm to 6am until trial

A judge barred Waymo from running its two Broadway charging lots between 11pm and 6am. The robotaxi bottleneck turns out to be the depot, not the driving.

 Answer card: Article 50 of the EU AI Act became applicable on 2 August 2026, covering AI interaction disclosure, machine-readable marking of generative output, deepfake labelling and AI-generated public interest text, with fines up to 15 million euros or 3 percent of worldwide turnover.

Dev News

EU AI Act Article 50 is live: what you must label

The AI Act transparency rules applied on 2 August 2026. The delay everyone read about covered different articles. What actually binds you now, and what still has no test.

 Answer card: Qualcomm confirmed on its Q3 FY2026 earnings call of 29 July 2026 that prices rise by a double-digit percentage on products shipping after 1 September 2026, broad across end markets, layering in gradually as contracts and product cycles expire, with no percentage or product list published.

Dev News

Qualcomm's double-digit chip price rise starts September 1

Qualcomm confirmed a double-digit price rise on chips shipping after September 1. It never named a percentage, and it is small next to the memory bill.

 Thinking Machines Lab announcement artwork for Inkling-Small, hand-drawn black loops winding across a cream graph-paper grid, carrying no text or product imagery.

Dev News

Inkling-Small: 276B that beats the 975B at coding

Thinking Machines shipped Inkling-Small on 30 July: 276B total, 12B active, Apache 2.0. It out-codes the 975B flagship, and forgets half of what that one knew.

 OpenAI announcement artwork for the publication Ten advances in mathematics and theoretical computer science, white text centred on a soft cyan and blue gradient.

Dev News

OpenAI Astra: what the ten Lean proofs actually prove

OpenAI's unreleased Astra produced ten results on decade-old maths problems for about $2,000. The Lean certificates are public and they compile. What that settles.

 Answer card: DeepSeek-V4-Flash-0731 replaced the April preview weights behind the same deepseek-v4-flash endpoint on 31 July 2026, keeping the same 284B parameter architecture with 13B active and the same $0.14 and $0.28 per million token pricing, while Terminal Bench 2.1 rose from 61.8 to 82.7.

Dev News

DeepSeek V4-Flash-0731: same API, new model

DeepSeek swapped the weights behind deepseek-v4-flash on 31 July. Same endpoint, same $0.14 price, Terminal Bench 2.1 up from 61.8 to 82.7.

 Google promotional image for Nano Banana 2 in Google Earth, showing generated aerial scenes alongside example prompts including rendering the Pompeii ruins as they looked in 78 A.D. and transforming the Google Mountain View campus into a futuristic sci-fi utopia.

Dev News

Google pulled Nano Banana 2 from Earth after one day

Google shipped AI image generation into Google Earth on 30 July and rolled it back on 31 July. The SynthID watermark only covers the pixels it generated.

 Zoox purpose-built robotaxi, a bidirectional four-passenger pod with no steering wheel and no pedals, photographed for the company announcement of its NHTSA Part 555 commercial exemption.

Dev News

Zoox wins Part 555: paid robotaxi, no steering wheel

NHTSA granted Zoox a two-year Part 555 exemption on 31 July, up to 2,500 vehicles a year. First time a passenger vehicle with no driver controls can charge fares.

 Google DeepMind announcement image for Gemini Robotics 2 showing an Apollo humanoid robot crouching to grasp a green watering can rendered with a point cloud overlay, beside the headline Gemini Robotics 2 brings whole body intelligence to robots.

Dev News

Gemini Robotics 2: what Google's own benchmarks say

Google DeepMind shipped Gemini Robotics 2 on 30 July with a demo reel and a score table. The table says a dustpan works 32 percent of the time. We read both.

 Answer card summarising Seagate fiscal Q4 2026: 50TB-class drives qualify in late 2027 and ship in 2028, nearline exabytes are allocated into calendar 2028, and price per exabyte rose about 10 percent year over year.

Sysadmin News

Seagate 50TB HAMR drives: 2028, not 2027

Seagate says 50TB-class drives qualify in late 2027 and ship after. The number that hits your budget first is price per exabyte, up around 10 percent.

 Answer card summarising the reported Amazon overrun: a Claude Sonnet author matching tool ran 860 percent over budget to $1.8 million across five months without detection, never shipped, and two further projects added roughly $675,000.

Dev News

Amazon's $1.8M Claude bill: 860% over, 5 months unseen

The FT reported a failed Amazon tool that burned $1.8 million in Claude Sonnet tokens. At list price that is 120 to 600 billion tokens, unnoticed.

 The official Cloudflare blog card for the announcement, titled Post-quantum authentication to origins is now supported, with an isometric illustration of a server under a shield.

Security News

Cloudflare origin PQC: ML-DSA-44, OpenSSL 3.5.0

Authenticated Origin Pulls and Custom Origin Trust Store now take ML-DSA certificates. The requirements: OpenSSL 3.5.0, seed-only keys, Full (strict).

 Answer card explaining that the FCC added foreign-produced advanced robotic devices to its Covered List on 28 July 2026, blocking new models from equipment authorization, with a definition starting at 4.4 pounds, one environmental sensor and a 200 kbps network link.

Dev News

FCC robot ban: the 4.4 lb rule that catches vacuums

The FCC put foreign-produced robots on its Covered List on 28 July. The definition starts at 4.4 lb with one sensor and 200 kbps, so it catches robot vacuums.

 Answer card: on 30 July 2026 OpenAI cut GPT-5.6 Luna from $1 and $6 to $0.20 and $1.20 per million tokens, an 80 percent reduction on both input and output, cut Terra from $2.50 and $15 to $2 and $12, and left Sol unchanged at $5 and $30, twenty one days after the family went generally available.

Dev News

GPT-5.6 price cut: Luna drops 80%, Sol stays at $5

OpenAI cut GPT-5.6 API prices on 30 July. Luna drops 80% to $0.20 / $1.20 and Terra 20%, while Sol does not move. What it does to your bill.

 The official Model Context Protocol wordmark, a white looping ribbon logo beside the words Model Context Protocol, on a near black background.

Dev News

MCP 2026-07-28: what the stateless core removes

MCP 2026-07-28 makes the protocol stateless and removes sessions outright. The twelve month deprecation policy in the same release does not cover that list.

 Official co-branded announcement image for the partnership, the ssi wordmark on a dark card beside the green Nvidia eye logo and wordmark, separated by a vertical rule on a black background.

Dev News

Nvidia SSI deal: $5B reported, 10x compute on Vera Rubin

Nvidia and Safe Superintelligence announced a partnership on 27 July. What the release actually says, what the $5 billion rests on, and what 10x leaves out.

 The official announcement image for ChatGPT for Academic Researchers, a soft pastel gradient card carrying the program name.

Dev News

OpenAI's free GPT-5.6 for researchers: what you get

OpenAI opened applications on 29 July for free frontier model access at universities. The real terms: five seats, twelve months, ChatGPT Pro rate limits.

 Answer card: Google VP of Devices and Services Shakil Barkat confirmed price adjustments across the Pixel family ahead of the Made by Google event on 12 August 2026, citing a supplier-driven RAM crisis and Morgan Stanley figures of $2.80 per gigabyte in 2025 against $12 in 2026, which is 4.3 times.

Dev News

Pixel 11 price increase: the RAM crisis, in numbers

Google confirmed Pixel 11 prices are going up and blamed a supplier-driven RAM crisis. The two numbers it cited divide to 4.3x, not the sixfold everyone printed.

 Poolside announcement graphic for Laguna S 2.1, purple isometric cubes descending across a grid on a cream background, with the words Introducing Laguna S 2.1 in black.

Dev News

Laguna S 2.1: 118B open weights on one DGX Spark

Poolside ships Laguna S 2.1, a 118B MoE with 8B active and a 1M context, under an open licence. The weights fit one DGX Spark at four bits. The 1M window does not.

 Answer card summarising that Tesla is installing Optimus production lines at Fremont for 2026 while the V3 robot itself remains unrevealed, with only one hardware figure ever published.

Dev News

Tesla Optimus V3: production starts, reveal still missing

Tesla is installing Optimus production lines at Fremont for a robot it has never shown. What the Q2 call actually confirmed, and why every V3 spec sheet is fiction.

 Official San Francisco AI Summit photograph dated 24 July 2026 showing eleven executives and officials applauding on stage in front of a blue backdrop.

Dev News

NVIDIA and SK's $500B deal: what the 2GW actually commits

SK Group and NVIDIA announced a $500 billion-plus partnership on July 24. The instrument is letters of intent. What is named is a 2GW AI factory for 2027, plus HBM4.

 False-colour electron micrograph of the HRL silicon quantum dot device, showing plunger gates P1 to P6 and S1 with exchange gates between them, a three-layer metal routing schematic, and a cross-section through the quantum well, with 200 nanometre scale bars.

Dev News

IBM buys HRL: 18 spin qubits and a 300mm fab

IBM is buying HRL Laboratories from Boeing and GM. The release has no price and no qubit count. We read the April paper that does: 54 dots, 18 qubits.

 Answer card: Apple released iOS 26.6 and iPadOS 26.6 on 27 July 2026 with a release note covering bug fixes, security updates and an optimized Spotlight index to prepare for iOS 27, the index the rebuilt Siri reads for personal context.

Dev News

iOS 26.6 is out: the Spotlight index it quietly builds

Apple shipped iOS 26.6 on July 27. One line of the release note matters: it rebuilds the Spotlight index that the new Siri in iOS 27 will read.

 Answer card: the European Commission fined Google 460 million euros for search self-preferencing and 430 million euros for Play anti-steering on 23 July 2026, with 60 days to comply before periodic penalty payments.

SEO News

Google's €890M DMA fine: what changes in 60 days

The EU fined Google €890M on July 23 in its first DMA decisions against the company. The money is trivial. The cease and desist order attached to it is not.

 Voxel Colosseum scene rendered in a browser demo generated by Kimi K3, showing a Roman amphitheatre packed with crowds inside a blocky city, with an FPS 120 overlay in the corner.

Dev News

Kimi K3 open weights land July 27: the 1.4TB catch

Moonshot ships the Kimi K3 weights by July 27. At 2.8 trillion parameters in MXFP4 that is about 1.4 TB, and Moonshot itself recommends 64 or more accelerators.

 Cursor bar chart titled Cost to rebuild SQLite by model mix, showing six stacked planner and worker bars from $1,339 for Opus 4.8 with Composer 2.5 up to $20,057 for Fable 5 alone, with a footnote marking the starred runs as informal.

Dev News

Cursor agent swarm: SQLite for $1,339 and the 15x catch

Cursor swarmed agents at the 835-page SQLite manual and got a working Rust database for $1,339. The 15x cost gap being quoted needs a run Cursor left out.

 AMD launch slide for Helios showing a row of AMD Helios racks with the EPYC, Instinct, Pensando and ROCm logos and the words In Production Today.

Dev News

AMD Helios vs Vera Rubin NVL72: the MI455X numbers

AMD launched Helios on July 23: 72 Instinct MI455X GPUs, 31TB of HBM4, and a claim of 30% more tokens per dollar than Vera Rubin NVL72. We read the footnotes.

 The signature page of the Open Weights and American AI Leadership letter, showing the logos of its 25 signatories including NVIDIA, Microsoft, Meta, IBM, Dell, Hugging Face, Mistral, Mozilla and Y Combinator.

Dev News

Open Weights letter: 25 signers, and what it changes

On July 24, 25 companies signed Open Weights and American AI Leadership. Who signed, who pointedly did not, and what it changes if you ship on downloadable models.

 Etched announcement image showing the underside of one of its liquid-cooled inference racks, dense with black coolant hoses and copper cold plates, captioned Frontier Inference Clusters.

Dev News

Etched raises $300M at $10.3B: where did Sohu go?

Etched raised $300M at a $10.3B valuation on July 23. Its own site no longer says transformer, or Sohu. What it is really shipping, and what stays unproven.

 Answer card: Black Forest Labs put FLUX 3 into early access on July 23 2026, a single multimodal model for video with native audio, image editing and robot action prediction, but only Video and Action are live for partners, Image is coming weeks, and the open-weight FLUX 3 Dev lands later in 2026.

Dev News

FLUX 3 is here: video, audio, open weights later

Black Forest Labs launched FLUX 3 on July 23: one model for video with audio, image and robot actions. Early access only, and the open weights land later in 2026.

 Anthropic Claude Opus 5 launch graphic, the new flagship model announced July 24 2026.

Dev News

Claude Opus 5 is here: same price, thinking on by default

Anthropic shipped Claude Opus 5 on July 24 at the same $5/$25 price as Opus 4.8, pitched as Fable 5 intelligence at half the cost, with thinking on by default.

 Answer card: PsiQuantum signed a $125M performance-based agreement with DARPA on July 22 under the Quantum Benchmarking Initiative, funding verification and validation of its pathway to a utility-scale quantum computer, building on a $31.8M Stage C award from September 2025, with no qubit count or error rate disclosed.

Dev News

PsiQuantum's $125M DARPA deal: what it validates

PsiQuantum signed a $125M DARPA agreement on July 22. It funds verification of a quantum pathway, not a working machine. We read what it actually buys.

 Answer card: on July 22, 2026 OpenAI launched Presence, a managed platform for governed voice and chat agents running on the GPT-5.6 series, in limited availability deployed with OpenAI engineers, with a 75% resolution figure that comes from OpenAI its own support line.

Dev News

OpenAI Presence: the enterprise AI agent platform

OpenAI launched Presence on July 22, a managed platform for governed voice and chat agents. What it is, who it is for, and what the 75% figure really means.

 NVIDIA Spectrum-X Ethernet switch hardware for AI factories, the platform that the Spectrum-6 102.4 Tb/s switch silicon plugs into, shown in NVIDIA corporate branding.

Network News

NVIDIA Spectrum-6: what 102.4 Tb/s Ethernet changes

NVIDIA's Spectrum-6 is a 102.4 Tb/s per-chip Ethernet switch for the Vera Rubin platform, with co-packaged optics. What's real, what's marketing, and who it's for.

 Answer card: a San Francisco judge gave final approval to Anthropic's $1.5 billion author settlement on July 20, 2026, about $3,000 per book across roughly 500,000 works, but it settles the piracy, not the fair-use question.

Dev News

Anthropic's $1.5B author settlement: what it changes

A judge gave final approval to Anthropic's $1.5B author settlement on July 20. What it actually settles, the fair-use question it leaves open, and what changes for you.

 Answer card: on July 22, 2026 OpenAI opened a self-serve Ads Manager for ChatGPT, with ads shown only to Free and Go users in a Sponsored block below the answer, targeting conversational context, and the old $50,000 minimum spend removed.

Dev News

ChatGPT ads are live: OpenAI's self-serve Ads Manager

OpenAI opened a self-serve Ads Manager for ChatGPT on July 22. Who sees the ads, where they sit, the budget confusion, and what it actually means for you.

 Answer card: on July 21, 2026 Anthropic shipped Record a skill in Claude Cowork, the workspace in the Claude desktop app, letting you screen-record a task with voice narration so Claude turns it into a reusable skill, available on Pro, Max and Team, with the caveat that it captures everything on screen.

Dev News

Claude Cowork Record a Skill: teach agents by demo

Anthropic shipped Record a Skill in Claude Cowork on July 21. What it does, which plans get it, and the screen-capture catch nobody is flagging.

 Answer card: on July 20 2026 Bristol Myers Squibb said it will add a second NVIDIA DGX SuperPOD built on eight DGX Vera Rubin NVL72 systems, Nvidia post-Blackwell generation, with an up-to-10x performance-per-megawatt claim measured against the infrastructure it replaces, not against Blackwell, and with nothing running yet.

Dev News

NVIDIA Vera Rubin: what BMS's first AI factory really is

Bristol Myers Squibb is deploying an NVIDIA DGX SuperPOD on eight Vera Rubin NVL72 systems. It's a build commitment, not a running cluster. What Rubin really is.

 Answer card: Alibaba put Qwen-Image-3.0 live in Qwen Chat on July 21 2026 claiming text legible to 10 pixels, native rendering in 12 languages and infographics from live data, while shipping no open weights, no benchmark score and no model card.

Dev News

Qwen-Image-3.0 is here: no weights, no benchmarks

Alibaba's Qwen-Image-3.0 went live July 21 claiming 10px text and live-data infographics, but with no weights, no benchmark and no model card. The honest read.

 Answer card: Google shipped Gemini 3.6 Flash on July 21, 2026 at $1.50 per million input tokens and $7.50 per million output tokens, down from nine dollars output on 3.5 Flash, using 17 percent fewer output tokens and scoring higher on Google-reported benchmarks, while Gemini 3.5 Pro is still only in testing with partners.

Dev News

Gemini 3.6 Flash: cheaper than 3.5, better on paper

Google shipped Gemini 3.6 Flash on July 21 at $1.50/$7.50 per million tokens, cheaper than 3.5 Flash and higher on its own benchmarks. The Pro is still in testing.

 Answer card: on July 7 2026 Mistral released Robostral Navigate, an 8B vision-language model that drives a robot from a single RGB camera and a plain-language instruction with no LiDAR or depth sensor, scoring 76.6 percent on the R2R-CE validation-unseen benchmark and beating the best depth or multi-camera systems by 4.5 points.

Dev News

Robostral Navigate: Mistral robots on one camera

Mistral's first robotics model steers robots from one RGB camera and plain-English commands. Here's what the 8B Robostral Navigate really does, and where it stops.

 Answer card: on Sunday July 19, Anthropic number theorist Levent Alpoge posted an explicit polynomial map on X, built with Claude Fable 5, that he says is a counterexample to the Jacobian Conjecture, a problem open since 1939.

Dev News

Did Claude Fable 5 disprove the Jacobian Conjecture?

An Anthropic number theorist used Claude Fable 5 to post a public counterexample to the 85-year-old Jacobian Conjecture. What is real, and what still needs checking.

 Answer card: Claude Code now runs on Bun 1.4.0 rewritten in Rust, shipping since version 2.1.181 as a roughly 100MB native binary with no Node.js, with startup about 10 percent faster on Linux.

Dev News

Claude Code drops Node for Bun 1.4.0 in Rust

Since v2.1.181, Claude Code ships as a native binary running Bun, ported from Zig to Rust with Claude. What changed for you, the real numbers, and the slop row.

 The HMND 01 Alpha humanoid robot standing in an industrial facility in a black vest branded HMND, SAP and Martur Fompak, with two arms ending in parallel grippers and grey logistics totes stacked around it.

Dev News

Humanoid HMND 01: what the robot unicorn really ships

London's Humanoid just hit a $1.2B valuation. Here's what its HMND 01 Alpha robot actually does, the specs that are real, and what's still just a claim.

 Answer card: on July 19 2026 Alibaba announced Qwen3.8 and put Qwen3.8-Max-Preview live on its Token Plan and Qoder tools, while the 2.4T parameter count and open weights are promised soon with no date, no model card and no benchmark table.

Dev News

Qwen3.8 preview is live: the 2.4T open weights are not

Alibaba's Qwen3.8-Max-Preview went up on July 19 with a 2.4T-parameter claim and open weights promised soon. Here's what you can run today, and what's still a tweet.

 Answer card: Fireworks closed a 1.505 billion dollar Series D on July 16 2026 at a 17.5 billion dollar valuation, led by Atreides Management, Index Ventures and TCV with Nvidia among the backers, as an AI inference platform for serving and fine-tuning open-weight models.

Dev News

Fireworks AI $1.5B Series D: what it means for open models

Fireworks closed a $1.505B Series D at a $17.5B valuation, Nvidia among the backers. A raise ships no code. Here is what it signals for teams serving open weights.

 Answer card: Classic McEliece is now in ISO/IEC 18033-2, the first post-quantum algorithm to reach full ISO standardization, built on a 1978 cryptosystem, with a public key that can exceed one megabyte.

Security News

Classic McEliece is the first ISO post-quantum standard

ISO just made Classic McEliece the first post-quantum algorithm in ISO/IEC 18033-2. Here is what that milestone means, and why it is not your TLS default.

 Answer card: Google let its Gemini 3.5 Pro target of July 17 slip past, a third miss, with no model card and nothing in the public API, while the 2M context window, Deep Think mode and roughly fifteen dollar per million pricing are all reporting rather than an official Google announcement.

Dev News

Gemini 3.5 Pro missed July 17: launch vs leak

Google's Gemini 3.5 Pro slipped its reported July 17 target, a third miss. The 2M context, Deep Think and pricing are all leaks. Here is what Google actually said.

 Answer card: from July 20, 2026 Claude Fable 5 is included in every Max and Team Premium plan at 50 percent of normal limits, while Pro and Team Standard users get only metered access at ten dollars per million input and fifty dollars per million output tokens plus a one-time hundred dollar credit.

Dev News

Claude Fable 5 goes permanent July 20: the fine print

Anthropic makes Claude Fable 5 permanent from July 20, but only Max and Team Premium include it, at 50% of limits. Pro and Team Standard pay per token.

 Answer card: NVIDIA published its Ising Decoder ColorCode 1 Fast color code pre-decoder on July 13 under Apache 2.0, claiming 347.7x better logical error rate and 7.3x faster runtime than Chromobius, though the figures come from a d=31 simulation and the speedup compares a DGX GB300 GPU against a Grace CPU.

Dev News

NVIDIA Ising Decoder ColorCode 1: what 347x really means

NVIDIA's open color code pre-decoder claims 347.7x better logical error rates. We read the chart, and the headline point at d=31 is extrapolated, not measured.

 Answer card: the Cyberspace Administration of China published seven registered on-device generative AI services on July 15 2026, Apple Intelligence among them with Alibaba Qwen handling text and images, but no launch date was given.

Dev News

Apple Intelligence China: approved, not shipped

China cleared Apple Intelligence on July 15 with Alibaba Qwen inside. No launch date came with it. What the filing actually covers, and what it does not change.

 Answer card: NVIDIA unveiled two new Jetson Thor modules on July 16, the T3000 at 865 FP4 teraflops with 32GB and the T2000 at 400 FP4 teraflops with 16GB, same Blackwell architecture as the T5000 at about half the size and power, both shipping in the first quarter of 2027 with no price posted.

Dev News

NVIDIA Jetson T3000 and T2000: Thor for mainstream robots

NVIDIA's new Jetson T3000 and T2000 bring Thor-class AI to smaller, cheaper robots. Here's the real spec gap, the Q1 2027 catch, and what's still hype.

 Answer card: Thinking Machines released Inkling on July 15 under Apache 2.0, a 975B mixture-of-experts model with 41B active parameters per token and a 1M token context window, which the lab says is not the strongest model available today.

Dev News

Thinking Machines Inkling: what the 975B open model is for

Thinking Machines released Inkling, a 975B open-weights MoE with 41B active params. It loses plenty of benchmarks on purpose. Here is what it is actually built for.

 Answer card: Apple opened the iOS 27 public beta on July 13 2026 with a rebuilt Siri that holds a conversation, reads on-screen content and takes actions in apps, though the iPhone Siri AI needs an iPhone 15 Pro or newer.

Dev News

iOS 27 public beta: the new Siri AI is finally here

Apple opened the iOS 27 public beta with a rebuilt Siri that holds a conversation and acts in your apps. Here is who can actually use it, and the real catches.

 Answer card: Moonshot shipped Kimi K3 to the Kimi app and API on the evening of July 16 in Max and Cluster Max tiers, with a confirmed 1M token context, while the 2.8T parameter count, benchmark table and API prices are still circulating rather than officially posted.

Dev News

Kimi K3 is live: the confirmed specs and the leak noise

Moonshot's Kimi K3 hit the Kimi app and API on July 16 with a 1M context. The 2.8T params, benchmarks and prices are still leak-grade. What to trust today.

 Answer card: Apple sued OpenAI, its hardware chief Tang Tan and former engineer Chang Liu in a California federal court on July 10, alleging trade secret theft at every level. OpenAI denies it and nothing is proven.

Dev News

Apple v. OpenAI: what the trade secrets suit changes

Apple sued OpenAI on July 10 over alleged trade secret theft. What the complaint actually claims, what stays unproven, and what it changes for your work.

 Answer card: Nvidia now sells AI chips in Singapore, Malaysia and Japan only to companies on a vetted white list, and more than half its former customers there failed the first review.

Dev News

Nvidia halves its Asia AI chip buyer list: who's cut

Nvidia cut more than half its authorised AI chip buyers in Singapore, Malaysia and Japan. What the white list check involves, who failed it, and if it reaches you.

 Answer card: New York froze state environmental permits for new data centers drawing 50MW or more, for up to a year, while it writes an environmental review.

Sysadmin News

New York pauses 50MW+ data centers: what changes

New York is the first US state to pause new hyperscale data centers. Where the 50MW line falls, what is carved out, and whether any of it reaches your bill.

 Answer card: DeepSeek V4 graduates from preview this month, and the deepseek-chat and deepseek-reasoner aliases stop responding after July 24, 2026, with reasoner mapping to v4-flash thinking rather than the larger v4-pro.

Dev

DeepSeek V4 is here: deepseek-chat breaks July 24

DeepSeek V4 is graduating from preview, and the deepseek-chat and deepseek-reasoner API aliases stop working after July 24, 2026. What to change, and the one trap.

 Answer card: ByteDance shipped Seedream 5.0 Pro on July 8, and it splits a finished poster into 10-plus editable transparent layers while rendering in-image text in more than ten languages.

Dev

Seedream 5.0 Pro: the image model that ships you layers

Seedream 5.0 Pro (ByteDance, July 8) splits an image into 10+ editable transparent layers and renders in-image text in ten-plus languages. Where it wins, and slips.

 Answer card: Meta is building its own AI chip, Iris, the newest in its in-house MTIA line, made with Broadcom on TSMC, targeting ranking and app inference with production starting in September.

Dev News

Meta Iris: what its new AI chip changes for you

Meta's new AI chip, Iris, enters production in September, built with Broadcom on TSMC. Here's what it really changes for you, and what's still just hype.

 Answer card: Meta Muse Spark 1.1 is a multimodal agentic model with a self-managed 1M-token context, launched July 9 at $1.25 input and $4.25 output per million tokens.

Dev News

Meta Muse Spark 1.1: cheap agentic coding, real caveats

Meta shipped Muse Spark 1.1 on July 9, its first paid model API: a 1M-token agentic coder at $1.25/$4.25. What it wins, where it trails, and the catch.

 Answer card: NIST has nine newer post-quantum signatures in progress, but Cloudflare says deploy ML-DSA now because it is the only standardized and ready option, at the cost of a signature growing from 64 bytes to 2,420.

Security News

Post-quantum TLS: why ML-DSA is the signature to ship

NIST has nine newer post-quantum signatures cooking, but Cloudflare says ship ML-DSA now. Here is what its size and speed really cost your TLS.

 Answer card: TypeScript 7.0 shipped July 8, the compiler ported to Go, about 10x faster on full builds, with the VS Code codebase type-checking in 10.6 seconds down from 125.7.

Dev

TypeScript 7.0 is here: the Go rewrite is 10x faster

TypeScript 7.0 shipped July 8, the compiler rewritten in Go and about 10x faster on full builds. What really changed, and the one catch for Astro and Svelte users.

 Answer card: GPT-5.6 goes public July 9 after a US government safety review, in three tiers, Sol at $5/$30, Terra at $2.50/$15, Luna at $1/$6, with Terra the default pick.

Dev News

GPT-5.6 goes public July 9: Sol, Terra or Luna

OpenAI is releasing GPT-5.6 to everyone from July 9 after a US safety review. Sol, Terra and Luna tiers, the prices, and which one to actually use.

 Answer card: Tencent open-sourced Hy3, an Apache 2.0 licensed 295B mixture-of-experts model that runs only 21B active parameters per token, with 256K context, free on OpenRouter until July 21.

Dev News

Tencent Hy3: what a free 295B open model really gives you

Tencent's Hy3 is an open 295B MoE under Apache 2.0, free on OpenRouter until July 21. Here's what it really is and the hardware you need to run it.

 Answer card: Fable 5 wins the hard coding benchmarks by 15 to 17 points; Grok 4.5 costs about 8x less on output and uses far fewer tokens.

Dev

Claude Fable 5 vs Grok 4.5: is 8x cheaper worth it?

Fable 5 wins the hard coding benchmarks; Grok 4.5 costs about 8x less on output and burns far fewer tokens. The gap, the price, and which one to actually run.

 Answer card: Grok 4.5 launched at $2 input and $6 output per million tokens, the cheapest serious coding model, trailing Fable 5 on hard benchmarks but far more token-efficient.

Dev

xAI Grok 4.5 is here: cheap, fast, coding-first

xAI shipped Grok 4.5 on July 8 at $2/$6 per million tokens, the cheapest serious coding model. What it wins, where it trails Fable 5, and the EU catch.

 Answer card: OpenAI GPT-Realtime-2.1 and a mini reasoning model for the Realtime API, with p95 latency down at least 25% from better caching.

Dev News

OpenAI GPT-Realtime-2.1: what changed for voice agents

OpenAI shipped gpt-realtime-2.1 and a mini reasoning model on the Realtime API. What actually changed, the 25% latency cut, and when to pick each.

 Answer card: OpenAI built its own inference chip, Jalapeño, a custom accelerator for running models built with Broadcom in about nine months, still an engineering sample.

Dev News

OpenAI Jalapeño: what its first chip changes for you

OpenAI's first custom chip, Jalapeño, is an inference accelerator built with Broadcom. Here's what it really changes for you, and what's still just hype.

 Answer card: Claude Fable 5 is back worldwide; its per-token rate is fixed but the effort dial moves the real bill by about 7x between low and xhigh.

Dev

Claude Fable 5 is back: what each effort level costs

Claude Fable 5 is back worldwide. Its rate is fixed at $10/$50, but effort (low to max) multiplies real cost by 7x. The numbers per level, and when each is worth it.

 Answer card: Fable 5 costs five times Sonnet 5 during the intro window and buys a 17-point SWE-bench Pro lead; Sonnet 5 remains the right daily driver.

Dev

Claude Sonnet 5 vs Fable 5: is the top model worth 5x?

Fable 5 costs five times Sonnet 5 right now. It buys a 17-point SWE-bench Pro lead and long-horizon depth. Where the top model earns it, and where Sonnet 5 is enough.

 Answer card: Claude Sonnet 5 closes the gap to Opus 4.8, wins terminal and agent work and ties knowledge tasks at 40 to 60 percent less cost, while Opus 4.8 still leads deep coding and math.

Dev

Claude Sonnet 5 vs Opus 4.8: cheaper, nearly as good

Claude Sonnet 5 costs 40 to 60 percent less than Opus 4.8 and closes most of the gap, even winning agentic coding. Where Opus still wins, and which to pick.

 Answer card: Proton Lumo 2.0 is private by policy, not by locality. Saved history is locked so even Proton cannot read it, but the prompt is decrypted on a Proton EU server to answer it, then forgotten.

Security

Proton Lumo 2.0 review: how private is it, really?

Proton Lumo 2.0 is a big jump for the private ChatGPT alternative. What its zero-access encryption actually protects, what the server still sees, and who should switch.

 Answer card: Qwen 3.7 Max and GLM-5.2 are tied on coding benchmarks, but GLM is cheaper and open-weight while Qwen Max wins reasoning and long autonomous runs.

Dev

Qwen 3.7 Max vs GLM-5.2: benchmarks, price and the catch

Qwen 3.7 Max vs GLM-5.2, fact-checked: they tie on coding benchmarks, but GLM is cheaper and open-weight while Qwen Max wins math, reasoning and long autonomous runs.

 Answer card: Microsoft 2011 Secure Boot certificates expire in 2026, but existing Linux systems keep booting; the fix is enrolling the 2023 key with fwupd.

Security Guide

Secure Boot certificates expire in 2026: the Linux fix

Microsoft's 2011 Secure Boot certificates expire in 2026. Your Linux or dual-boot PC keeps booting; the real fix is enrolling the 2023 key with fwupd. Here's how.

 Answer card: Claude Mythos did not breach the NSA; it ran an authorized red-team drill on test networks, but its real capability (a 17-year-old bug exploited in hours) is the actual story.

Security

No, Claude Mythos didn't hack the NSA: what really happened

An authorized AI red-team drill became the viral claim that Claude Mythos hacked the NSA. What Anthropic's eval really showed, and what it means for defenders.

 Answer card: Qwen 3.7 Max is API-only and cannot run locally yet; the open Qwen models (Qwen 3.6 27B, qwen3:8b to 32b) run offline via Ollama.

Dev Guide

Qwen 3.7 local: what you can actually run offline

Qwen 3.7 Max is API-only, so you can't run it locally yet. Here's how to run Qwen offline today with Ollama and the open Qwen models, sized to your hardware.

 Answer card: GPT-5.6 is not one model but three tiers (Sol, Terra, Luna); Terra is half the price of GPT-5.5, Sol tops Terminal-Bench, and there is a new ultra subagent mode.

Dev

GPT-5.6 vs GPT-5.5: what actually changed

GPT-5.6 is not one model but a new Sol, Terra and Luna tier system. What changed from GPT-5.5: prices, the coding SOTA, the ultra subagent mode, and whether to switch.

 Answer card: a DMARC aggregate report is a daily XML summary of who sent mail as your domain, how many messages, and whether DKIM and SPF passed and aligned.

Email/DNS

How to read a DMARC report: the aggregate XML, decoded

How to read a DMARC aggregate report: what the XML fields mean (source IP, count, disposition, DKIM and SPF alignment) and how to spot real senders versus spoofing.

 Answer card: every file has three permission sets (owner, group, other), each with read write execute, written as letters like rwxr-xr-x or as octal like 755.

Sysadmin

Linux file permissions explained: rwx and 755

How Linux file permissions work: the rwx triads, owner group and other, octal modes like 644 and 755, and the execute bit on directories that quietly breaks things.

 Answer card showing MB as 1,000,000 bytes (decimal) next to MiB as 1,048,576 bytes (binary), with the note that they drift apart as numbers grow.

Sysadmin

MB vs MiB: why your 1 TB drive shows up as 931 GB

MB vs MiB, KB vs KiB: the difference between decimal (1000) and binary (1024) units, why a 1 TB drive reads as 931 GiB, and where Mbps and bytes part ways.

 Answer card: the June 2026 nginx CVEs are critical but only affect HTTP/3 or HTTP/2-upstream/gRPC configs; a default static or HTTP/1.1 nginx is not exposed.

Security Guide

Is your nginx exposed? How to check and harden it

The June 2026 critical nginx CVEs only hit specific configs. How to check if you are exposed, which version patches them, and how to harden nginx for the long run.

 Answer card: a sudo user instead of root, SSH keys with passwords off, a default-deny ufw firewall, fail2ban, and automatic updates. Set up the key before disabling passwords.

Security Guide

How to secure a new Ubuntu VPS

Secure a fresh Ubuntu VPS: a sudo user instead of root, SSH keys with passwords off, a default-deny ufw firewall, fail2ban, and automatic updates. The order that matters.

 Answer card: nslookup on any OS, dig on Linux and macOS, dig +short for just the answer, and @8.8.8.8 to query a specific resolver.

Network Guide

How to do a DNS lookup from the command line

DNS lookup from the command line: nslookup on any OS, dig on Linux and macOS, dig +short for just the answer, and @8.8.8.8 to query a specific resolver. The commands.

 Answer card: ipconfig on Windows, ip addr on Linux, ifconfig on macOS for the local IP, and curl ifconfig.me for the public IP.

Network Guide

How to show your IP address from the command line

Show your IP from the command line: ipconfig on Windows, ip addr on Linux, ifconfig on macOS for the local address, and curl for the public one. The commands per OS.

 Answer card: chmod 644 for files, chmod 755 for scripts and folders, chmod +x to make a file executable, chmod -R to apply down a tree.

Sysadmin Guide

How to change file permissions with chmod

Change file permissions with chmod: chmod 644 for files, chmod 755 for scripts and folders, chmod +x to make a file runnable. The numbers explained, with examples.

 Answer card: GLM-5.2 beats GPT-5.5 on several coding benchmarks and costs about 6x less than Opus 4.8, but Opus still leads the hardest tasks and is the only one that reads images.

Dev

GLM-5.2 vs GPT-5.5 and Opus 4.8: benchmarks and price

GLM-5.2 vs GPT-5.5 and Claude Opus 4.8, fact-checked: it beats GPT-5.5 on coding and costs about 6x less than Opus, but Opus still leads the hardest tasks.

 Answer card: a static site wins on speed and security, WordPress wins on convenience for non-technical editors; the deciding question is who edits the site.

SEO

WordPress vs a static site: which to run in 2026

WordPress vs a static site, honestly: static is faster and safer, WordPress is easier for non-coders. How they really differ on speed, security, cost and upkeep.

 Answer card: change an OVH DNS zone by editing the A record in the DNS Zone tab to point the domain at your server IP.

Network Guide

How to change your DNS zone at OVH

Change your DNS zone at OVH in the DNS Zone tab: edit the A record to your server IP. The steps, the leftover-record trap, and how to check the change took.

 Answer card: manage a Linux service with systemctl status, restart and enable --now, and read logs with journalctl.

Sysadmin Guide

How to manage services with systemctl

Manage Linux services with systemctl: status, start, stop, restart, enable at boot, and journalctl for logs. The commands you need and what enable --now does.

 Answer card: find files on Linux with find . -name, filtering by type, size or mtime and acting with -delete.

Sysadmin Guide

How to find files with find on Linux

Find files on Linux with find: by name, type, size or age, then act with -delete or -exec. The syntax, real examples, and how to delete matches safely.

 Answer card: search text in files with grep -r, adding -i to ignore case, -n for line numbers and -l for filenames.

Dev Guide

How to search text in files with grep

Search text in files with grep: grep -r for a whole tree, -i to ignore case, -n for line numbers, -l for filenames. The flags that matter, with a real example.

 Answer card: download a file with curl -O and test an API with curl -i, adding -L to follow redirects.

Dev Guide

How to download a file and test an API with curl

Use curl to download files (-O, -L) and test APIs (-i, -X, -d) from the command line. The flags that matter, a POST with a JSON body, and how to read the response.

 Answer card: extract a tar.gz with tar -xzf and create one with tar -czf, reading the flags as extract or create, gzip, file.

Sysadmin Guide

How to extract and create tar.gz archives (tar)

Extract a tar.gz with tar -xzf and create one with tar -czf. What the x, z, f, c flags mean, how to target a folder with -C, and the bz2 and xz variants.

 Answer card: generate an SSH key with ssh-keygen -t ed25519, which creates a private key and a .pub public key.

Security Guide

How to generate an SSH key (ssh-keygen)

Generate an SSH key with ssh-keygen -t ed25519, set a passphrase, and copy the public key to a server. Same steps on Linux, macOS and Windows, with the RSA fallback.

 Answer card: free a busy port on Linux by finding the PID with sudo lsof -i :8080 then running kill, or kill -9 as a last resort.

Sysadmin Guide

How to kill the process using a port on Linux

Address already in use on Linux? Find the PID with lsof -i or ss -ltnp, then kill it (kill -9 last). Steps, a fuser one-liner, and SIGTERM vs SIGKILL.

 Answer card: free a busy port on Windows by finding the PID with netstat -ano then running taskkill /PID number /F.

Sysadmin Guide

How to kill the process using a port on Windows

Port already in use on Windows? Find the PID with netstat -ano, then taskkill /PID /F to free it. Step by step, with the PowerShell one-liner and the gotchas.

 Answer card: flush the DNS cache on Linux with sudo resolvectl flush-caches for systemd-resolved, or restart nscd or dnsmasq.

Network Guide

How to flush the DNS cache on Linux

Flush the DNS cache on Linux the right way: resolvectl for systemd-resolved, plus nscd and dnsmasq. Find which resolver you run and confirm the cache cleared.

 Answer card: flush the DNS cache on Windows by running ipconfig /flushdns in Command Prompt or PowerShell, no reboot needed.

Network Guide

How to flush the DNS cache on Windows

Flush the Windows DNS cache in one command: ipconfig /flushdns. Why and when to do it, how to confirm the cache is empty, and what to try when it does not help.

 Answer card: reveal a saved WiFi password on Windows with netsh wlan show profile name=NetworkName key=clear, reading the Key Content line.

Network Guide

How to see a saved WiFi password on Windows

Forgot a WiFi password your PC already knows? One netsh wlan command shows any saved WiFi password on Windows in clear text. Steps, a real example, and the gotchas.

 Answer card: JWTs are not encrypted, anyone can read them; the signature proves who issued the token, not who may read it.

Security

Are JWTs encrypted? No, and the difference will bite you

JWTs are encoded and signed, not encrypted: anyone holding a token can read every claim. What the signature really protects and what never belongs in a payload.

 Answer card: three DNS records decide if your mail lands or bounces; SPF lists allowed senders, DKIM signs messages, DMARC sets the failure policy.

Email/DNS

SPF, DKIM and DMARC explained: the records your email needs

SPF lists who may send for your domain, DKIM signs each message, DMARC sets the policy when checks fail. How the trio works and how to deploy it safely.

 Answer card: a random 8 character password falls in under 2 hours offline, while 16 random characters hold for 1.4 trillion years at the same speed.

Security

How long does it take to crack a password in 2026?

Real crack times for 2026: an 8 character password falls in under 2 hours offline, 16 characters holds for ages. The math, the hardware and what to do.

 Answer card: DNS TTL is the expiry date on a DNS answer, the seconds a resolver may cache it; propagation is mostly waiting for old TTLs to run out.

Network

What is DNS TTL? Cache, propagation and the values to use

TTL is the number of seconds resolvers may cache a DNS answer. How it really works, why propagation is mostly waiting, and the right TTL per situation.

 Answer card: a /24 is 256 addresses with 254 usable, because the first 24 bits name the network and the last 8 bits identify hosts.

Network

What is a /24? CIDR notation and subnet masks, finally clear

A /24 is a block of 256 IPv4 addresses, 254 usable. What the slash number means, how masks work bit by bit, and how to size real networks.

 Answer card: TLS 1.3 is one round trip faster than TLS 1.2 and removed the broken options; both are secure when configured well.

Security

TLS 1.2 vs TLS 1.3: what changed and what to run in 2026

TLS 1.3 handshakes in one round trip instead of two, drops every broken cipher and encrypts more of the handshake. What changed, why, and how to migrate.