Fri, Oct 30, 2026Friday, October 30, 2026 · 20 stories · 8 min read
With Large 4 weights out, Mistral bets on control over raw scores
Ramit KoulFounder, Software Engineer & Innovator · Published 5:03 AM ET
Anyone can now download and run Mistral Large 4, a 1 trillion-parameter model that Mistral says leads open models in agentic coding and cyber defense.1, 2 The release comes as Artificial Analysis's new v5 index widens the measured gap between the best closed and open models to 12 points, so for now self-hosting buys control more than top scores.3 OpenAI also rebuilt its Model Spec around five values, and Hollywood unions asked a federal judge to reject blanket fair use for AI training.4, 5 Here are the 5 stories that matter most, then 15 briefs. Numbers in the text link to the references at the end.
1Models
Mistral releases the weights for Large 4 after three more weeks of training
Image: The Next Web
Mistral has published the weights of Mistral Large 4, nicknamed Le Chonk, on Hugging Face and made the model available through its API.1, 6 The model was first released as a limited preview on October 6.7 It has about 1.05 trillion parameters, activates around 52 billion per task, and takes text and images as input.8, 2 A quantised version runs on four Nvidia B200 or B300 GPUs.2
Mistral says three more weeks of reinforcement learning raised its Terminal-Bench 4.0 score from 28% to 41%, roughly level with GLM-5.3.2, 7 The company also says the model uses 47% fewer tokens per task than GLM-5.3 on the Artificial Analysis Coding Agent Index, and that it scored 79% on the CyberGym-E2E cyber defense test.2 Closed models are still ahead: Anthropic's Mythos 5.1 scored 60.9% on the same terminal benchmark.7
At testers' request, Mistral added low and medium reasoning modes alongside high, and made the high mode 50% more efficient.2, 9 The model was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centers, and CEO Arthur Mensch says the company has raised about $6 billion in three years.9 Smaller distilled versions, including one that runs on a laptop, are planned, but Mistral gave no dates.2
Why it matters for builders
Teams that need on-premises or sovereign deployment now have a frontier-class open model with reasoning depth they can tune per task.9, 8 Most of the benchmark figures come from Mistral, so test it on your own workloads before you switch.7
2Models
OpenAI rebuilds its Model Spec around five values and renames the model GPT
OpenAI has reorganized its Model Spec, the public document that describes how its models should behave, around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.4 OpenAI calls faithfulness the foundation, meaning the model should work out what people actually want instead of following instructions word for word.4
The update adds guidance on context, oversight and conflicting instructions, and drops most of the detailed style rules.10 The document now calls the model "GPT" instead of "the assistant".4 Most rules remain defaults that developers and users can change, with firm limits only where they are needed to protect people.4
OpenAI says its models do not yet fully reflect the Spec and that it is testing new training methods to teach these values.4, 10 It is also asking the public where they disagree with the balance the new Spec strikes.4
Why it matters for builders: The Spec spells out how GPT models are meant to weigh developer instructions against user intent and context, which shapes how your system prompts get read.4 OpenAI says actual behavior still lags the document, so test edge cases yourself.10
3Policy & legal
Hollywood unions ask a court to reject blanket fair use for AI training in the OpenAI case
SAG-AFTRA, the American Federation of Musicians and both Writers Guilds filed a joint amicus brief on Thursday in the consolidated New York copyright case brought against OpenAI by news outlets and authors.5 The 34-page brief argues that fair use should be decided case by case, not granted as a categorical exemption for training.5
The unions warn that a blanket ruling would allow a "potentially enormous transfer of value" from creators to AI companies.5 They also point to their 2026 contracts, which give writers and actors a notice-and-bargaining right when covered work is licensed to train commercial generative AI.5 A ruling that training is categorically fair use, they argue, would make those negotiated terms moot.5
Plaintiffs in the case include The New York Times, The Intercept, the Authors Guild, and writers such as George R.R. Martin, John Grisham and Jodi Picoult.5
Why it matters for builders: How this court rules on fair use will affect whether training data needs licensing, which feeds directly into data costs and legal risk for anyone training models.5
4Models
Artificial Analysis v5 moves most of its index to private tests and widens the open-weight gap
On Tuesday, Artificial Analysis released version 5 of its Intelligence Index, its largest overhaul since January.3, 11 For the first time, 60% of the score comes from evaluations with private test sets or held-out answers, a change meant to reduce benchmark contamination.11, 3 The index now has four equally weighted areas: knowledge work, coding, science and general capability. FrontierCode 2.0 and Terminal-Bench-Science are new, and GDPval-AA, SciCode and AA-LCR were dropped.3
The gap between the best proprietary and best open-weight models grew from 3 points under v4.3 to 12 points.11 Kimi K3.1 Pro, whose weights are not yet released, leads the open models at 46.11 At launch, Claude Opus 5.5 led at 58 and GPT-6 Astra followed at 57.3 No model leads every test, and Kimi K3.1 Pro tops Terminal-Bench 4.0 at 67%.3
The new tasks demand more work: median output tokens per task rose from 35,000 to 59,000.11 Effort settings also matter more. Claude Opus 5.5 scores 35 at low effort and 58 at max, at about ten times the cost per task.11
Why it matters for builders: v5 scores are on a new scale and cannot be compared with v4 numbers.3 When you choose a model, read the score alongside its effort setting and cost per task, because those can change the ranking.11
5Models
Genie 4 shifts the multimodal focus from what models read to what they generate
A Forbes analysis argues that Google's Genie 4 marks progress in multimodal output, not multimodal input, where most attention has gone so far.12 Google says Genie 4 generates interactive worlds in real time with native audio, including ambient sound, music and dialogue.13
Users can steer a scene while it runs by typing commands, such as moving the camera, changing the weather or adding characters.13 The model draws on Gemini Omni for realistic appearance and on other Gemini models for the reasoning that keeps a simulation consistent.13
Genie 4 is rolling out to a limited group of trusted testers.13 Google says it is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and AI Studio.13
Why it matters for builders: Once API access opens, Genie 4 could serve as a training environment for agents and robots or as a base for interactive media.13 Google has not published a release date for developers.13
In brief
Models
Kimi K3.1 Pro drops to ninth on the new index, and Gemini 4 Carbon takes first place. Under Artificial Analysis's Intelligence Index v5, Moonshot's Kimi K3.1 Pro fell from fourth to ninth, and Google's Gemini 4 Carbon, announced Wednesday, now ranks first at 59.14, 15 Moonshot has said it will release the Kimi model's weights on November 4.14
Dev tools
Mistral opens a Responses API preview so OpenAI-format agent code can run on its models. Mistral's new /v1/responses endpoint is in public preview and uses the OpenAI Responses format.16, 17 It is stateless and returns answers, reasoning and tool calls as separate typed items, and it supports text, image and file inputs, streaming and function calling across Mistral's hosted chat models, including Large 4.16, 17
Policy & legal
Anthropic tells California's governor that courts should be able to block risky AI models. In an October 15 letter to Governor Gavin Newsom, Anthropic proposed that the attorney general could ask a court to pause a model that poses a near-term risk of catastrophic harm, and that independent auditors should work inside leading AI labs.18 OpenAI's own recommendations backed shutdown controls and in-house auditors but not court orders, according to Politico.18
Products
A walkthrough describes a redesigned Claude app with nested sessions and cloud projects. According to a Geeky Gadgets guide, the update adds a session rail for managing several conversations at once, nested and master sessions, and context cards that show memory usage.19 The guide also describes cloud-based projects with shared custom memory and artifacts that can embed models and live data.19
Policy & legal
The Justice Department is reportedly examining Nvidia's planned $13 billion purchase of Hugging Face. According to Reuters, as cited by Tom's Hardware, the DOJ has asked how the deal could affect Nvidia's competitors, since Hugging Face is a major distribution channel for AI models.20 Nvidia has said it will keep supporting a range of chip and AI providers, and the deal is expected to close in the first half of 2027.20
Products
OpenAI drops a planned cut to ChatGPT Pro limits and doubles usage instead. OpenAI says usage limits for its $100, $200 and $500 Pro plans and for Business Premium are doubled, effective immediately, mainly for ChatGPT Work and Codex, after it had planned to reduce the Pro 200 allowance.21 Usage of Dots, its always-on assistants, will continue to count separately, but OpenAI does not publish exact limits.21
Models
Gnani AI releases Prisma v3.0, a speech recognition model for code-mixed Indian languages. The Bengaluru company reports a macro word error rate of 16.93 on the Vistaar benchmark across nine Indian languages and English, compared with 19.98 for ElevenLabs Scribe v2.22 Prisma v3.0 accepts up to 200 custom vocabulary terms per request and can be deployed in the cloud, hybrid or air-gapped on-premises environments.22
Products
The Gemini app now lets users generate, preview and edit files inside the chat. Users can type edits directly, highlight text to ask Gemini questions or have it revise documents, PDFs, slides, spreadsheets and code.23 PDF editing works on web and mobile, other formats are web only, and the feature is available in every language Gemini supports.23
Products
Meta's Muse agent now connects to common small business tools. Muse for Small Business, announced September 29, connects to Shopify, Slack, Stripe, QuickBooks, Notion and other apps, and Meta says nothing is published, sent or paid for without user approval.24 Muse runs in a dedicated cloud VM with an audit trail, and Meta says its data is not shared with Meta's ad systems.24
Industry
AI labs that called for a slowdown show little sign of decelerating. Six weeks after Anthropic CEO Dario Amodei and the heads of OpenAI and SpaceX said they wanted to curb rapid AI progress, The Washington Post finds little evidence that their companies have slowed, and reports that Amodei told investors he would race ahead.25
Infrastructure
Iluvatar CoreX says its Tiangai 300 GPU matches Nvidia's H200 on real workloads. In a live demonstration for Chinese media, the Shanghai chipmaker showed the Tiangai 300 beating the H200 on text-to-image and text-to-video tasks using models including Z.ai's GLM-5.3, according to a report cited by the South China Morning Post.26
Infrastructure
GitHub has added 4 million CPU cores this year to handle code pushed by agents. GitHub says monthly commits reached 7.4 billion in September, up from 1.4 billion in April, and that more than 40% of pull requests are now started by agents.27 The company also added 160 petabytes of storage and 50 Tb/s of network capacity, and is reworking its backend to separate storage from compute.27
Infrastructure
An opinion piece argues that electricity is the real limit on AI growth. Writing for Al Jazeera, the author notes that Amazon and Google have signed 20-year nuclear power deals that add output from existing reactors rather than new ones.28 The piece also points out that PJM's market monitor attributes about $29.4 billion in capacity costs across four auctions to data center demand.28
Industry
Valve says it will not ban AI-generated games from Steam. At a developer Q&A in Busan, Valve kept its policy of requiring developers to disclose AI use without banning it, and advised indie studios worried about AI clones to keep updating their games and building communities.29 Valve said it would look into reports of extortion emails threatening AI clones.29
Dev tools
A summary outlines the AI coding workflow promoted by Vercel's CEO. According to a Geeky Gadgets summary, the approach has developers test ideas early in isolated projects, track work against project goals instead of output volume, and set non-negotiable rules enforced by automated checks.30 It also calls for security quality gates and "breaker agents" that stress test applications before release.30
References
Every source behind this edition. Open one to read the full story.
