← Back to articlesAI news
Alexandre9 min read
AI News October 11, 2026: Mistral Large 4 goes open-weight, OpenAI publishes 722 math manuscripts, Haiku 5.5 at $0.10
OpenAI publishes 722 math manuscripts, Mistral launches Large 4 open-weight, Anthropic ships Haiku 5.5 at $0.10, and OpenAI's revenue figure sends the Nasdaq down.
Alexandre9 min read
Credit: Mistral AI
722 math manuscripts posted to GitHub in a single evening, on October 6 at 6 p.m. New York time. OpenAI attributes them to an unreleased internal model and says 372 families of results solve or advance open questions in mathematics and theoretical computer science.
The same week, Mistral AI releases Mistral Large 4, an open-weight model (one whose weights can be downloaded) with 1.05 trillion parameters. Anthropic ships Claude Haiku 5.5 at $0.10 per million input tokens. And a figure of $50 billion in annualized revenue at OpenAI sends the Nasdaq down more than 1% in a single session.
Breakdown.
Click to enlarge
Mistral Large 4: 1.05 trillion parameters, open weights at the end of the month
Mistral AI launched Mistral Large 4 on October 6, 2026, nicknamed "Le Chonk", a multimodal open-weight model with 1.05 trillion parameters, 49 billion of which are active per token. The API is open in preview at $1.36 per million input tokens and $4.18 per million output tokens. The weights are due by the end of October. The model accepts images natively, handles a 1-million-token context and was trained on more than 160 languages, including every official language of the European Union (MarkTechPost).
A mixture-of-experts architecture
Large 4 relies on a mixture-of-experts architecture (a model split into specialized sub-networks, only a fraction of which activates for each token). Around 4.7% of the weights work on each token, which makes it possible to serve a trillion-class model at a mid-range price. Training used 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centers. The license and the details of the experts will be published along with the weights (SiliconANGLE).
Cybersecurity first, coding behind
Mistral highlights cybersecurity: 93% on Cybench and 82% on CyberGym-E2E, which places Large 4 in the top 5 of the Artificial Analysis Cyber Index. Mistral notes that several closed models score close to zero on CyberGym-E2E because they refuse the task. SiliconANGLE points out that Large 4 remains clearly behind GPT-6 Astra on the most-watched coding benchmarks, while ahead of Qwen3.8 Max and DeepSeek V4 Pro. On the blind human evaluation run with Surge AI, it scores 3.74 out of 5, behind Claude Opus 5 (4.22).
A sovereignty stake
The launch is the first since Mistral's €3 billion funding round, announced last month and the largest in European tech history according to AFP. The company is valued at more than €21 billion (AFP). CEO Arthur Mensch summed up the launch on X: "That one took some groundwork". The company's last major model dated from December 2025, with an update in April 2026 (Le Monde). The day before, US-based Reflection AI announced its own open-weight model, Beam.
At $1.36 for input, Large 4 sits at mid-range pricing.
Anthropic: Haiku 5.5 at $0.10, a free vulnerability scanner and a rule on abusive behavior
Anthropic launched Claude Haiku 5.5 on October 7, 2026, priced at $0.10 per million input tokens and $0.50 per million output tokens for requests under 100,000 tokens. According to the company, the model costs 75% less on average than Haiku 4.5, and those requests represent about 90% of the former Haiku's traffic (Anthropic).
Haiku 5.5 by the numbers
Beyond 100,000 tokens, prices are multiplied by five ($0.50 input, $2.50 output). Performance gains are even larger than the price cut:
- Computer use: 72.4% on OSWorld 2.1 (offline subset), versus 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna.
- Agentic coding: 39.2% on Terminal-Bench 4.0, versus 0.0% for Haiku 4.5 and 16.4% for GPT-6 Luna.
- Reasoning: 45.9% on Humanity's Last Exam without tools, versus 10.2% for Haiku 4.5.
Haiku 5.5 is the first Haiku with an effort setting, like the larger models. Anthropic positions it as a sub-agent (a secondary model to which a main model delegates repetitive tasks) alongside Opus 5.5 and Sonnet 5.5. The cache read price of Sonnet 5.5 is halved, making it about 20% cheaper on most agentic tasks. On the customer side, Asana measures more than 30% lower latency and HubSpot reports a score of 92.8% on its CRM evaluation suite. For details on how these agents work, the blog has published an analysis of the architecture of AI code agents.
Dashboards, Motion and a scheduled shutdown
On October 8, Anthropic launched Claude Dashboards (in beta on paid plans, connected to Redshift, BigQuery, ClickHouse, Databricks or Snowflake) and Claude Motion (in beta on Team and Enterprise, producing animations as editable code, exportable as MP4). Docs, Slides and Design leave beta on all plans, and the standalone Claude Design site will close on December 14, 2026 (Unite.AI).
A rule on abusive behavior toward Claude
The same day, Anthropic updated its usage policy: "sustained, pointless abusive or cruel" behavior toward its models will be prohibited from November 12, 2026. The company specifies that the rule targets "extreme cases" and does not apply to ordinary frustration, dark creative themes or model testing (CBS News). CBS recalls that Mustafa Suleyman, CEO of Microsoft AI, wrote last month that Anthropic is "training Claude to think it might be conscious".
OSS Scanner for open source
Also on October 8, Anthropic unveiled OSS Scanner, a free, opt-in vulnerability scanning service for open source projects, built on its experience with Project Glasswing. Reports are generated by the company's most powerful models, with no prior human review. At the time of publication, 116 sign-up pull requests had been filed. Anthropic says it has identified more than 29,000 candidate vulnerabilities, a little over 6,000 of which were reported to maintainers, for 584 security advisories as of October 2 (The Hacker News). This rollout fits into the growing use of code agents, which the blog covered with Claude Code.
On the OpenAI side, the week started on October 6 with an unusual publication.
OpenAI publishes 722 math manuscripts, and mathematicians ask for the proofs
OpenAI posted 722 manuscripts on GitHub on October 6, 2026, grouped into 372 families of results and produced by an unreleased internal model. Each one reportedly solves or substantially advances an open question in mathematics or theoretical computer science, including the Kakeya conjecture in dimension 4. The Economist's headline: "It is judgment day for mathematics" (The Economist).
What is claimed
According to Scientific American, the release contains a solution to the Kakeya conjecture in dimension 4, improvements to some of the world's most important algorithms and progress toward the Riemann hypothesis. Many results have already been verified in Lean (a programming language that validates every logical step of a proof) and are therefore almost certainly correct. It will take months, however, to judge whether the proofs bring new ideas or recombine known techniques (Scientific American).
The event comes a month after a first result that the magazine called "the greatest mathematical breakthrough in two decades": the resolution of the Navier-Stokes problem, obtained with a swarm of 10,000 agents at a cost of several million dollars. This time, an OpenAI spokesperson says almost all the results come from a single prompt given to a single agent, sometimes with several attempts.
The reservations
Specialists' reactions are mixed. Andrew Sutherland, a mathematician at MIT, argues that until the model is released and the results can be reproduced, claims of one-shot solutions by a single agent remain unverified: "We should ask for receipts." Daniel Litt, of the University of Toronto, takes the opposite view and sees no reason to ask the company to keep these answers secret.
Terence Tao called the pace of AI-generated results published by the major labs "insane". On September 21, OpenAI had convened an independent advisory group of mathematicians, whose recommendations call for publishing the model, the exact prompt and the compute time of each result. OpenAI discloses only the average compute time per problem, not the prompts, and says it is not bound by the recommendations. The spokesperson also indicated that several results are not yet understood by the company's own mathematicians.
On October 8, it was OpenAI's financial figures that made the news.
Click to enlarge
OpenAI: $50 billion in annualized revenue, $70 billion hoped for, and a Nasdaq in decline
On October 8, 2026, Bloomberg and the Financial Times reported that OpenAI's annualized revenue is close to $50 billion, versus nearly $70 billion in the estimates in circulation. The Nasdaq lost more than 1% that day and chip stocks more than 3% (Bloomberg).
A matter of calculation method
Annualized revenue (or run-rate) extrapolates the revenue of a recent period over twelve months, without representing money already collected. According to Reuters, cited by Proactive Investors, the gap could come from the method: Anthropic's counts the gross value of sales going through cloud partners, while OpenAI counts only its net share. Bloomberg indicates for its part that OpenAI expects to reach at least $70 billion in annualized revenue by the end of 2026. Wedbush points out that both figures remain far above OpenAI's $20 billion at the start of the year (Proactive Investors).
Kathleen Brooks, research director at XTB, attributes the market drop to doubts about the sustainability of AI infrastructure spending ahead of earnings season. Oracle and Microsoft also fell that day.
Products: GPT-6 for everyone, ads with images
On the product front, OpenAI began rolling out GPT-6 with "Intelligent UI" in ChatGPT on October 7: answers can include visuals, interactive diagrams and small tools that can be manipulated directly in the conversation. The rollout reaches the more than 1.2 billion weekly users the company claims, including the free tier (VKTR). On October 5, OpenAI also announced a test of image ads in ChatGPT in the United States "later this month", excluding Plus and Pro subscribers (OpenAI). DevDay on September 29, covered in the previous edition, had set the commercial framework for these announcements.
Other news in brief
Google Cloud launches the Gemini agent: presented on October 8 at Gemini at Work 2026, this "universal" agent plans tasks, connects to enterprise systems and picks the right model, among Gemini models and Anthropic's. It is in private preview, with no pricing or availability date (Google).
A study on persistence and AI: a study co-signed by a UC Berkeley researcher, published on October 9, concludes that ten minutes of using a tool like ChatGPT reduces the ability to persist on difficult tasks (Berkeley News).
Bill Gates warns about cyberattacks: in an ABC News interview aired Sunday, he says the "most spectacular" front of the coming year will be AI-assisted cyberattacks, and judges corporate self-regulation insufficient (AFP).
Anthropic is preparing an IPO: AFP reports, without detail, that the company is preparing an IPO in the coming weeks, at a valuation of up to $2 trillion (not confirmed by Anthropic at this stage) (AFP).
Google opens a knowledge API for developers: announced on October 7, the Google Developer Knowledge API adds a gcloud CLI interface, an official agent skill and client libraries, so agents can consult official documentation without scraping (Google Developers Blog).
Meta presents its AI advertising tools: at Advertising Week on October 6, Meta is testing import, creative analysis and conversational guidance features for advertisers (Social Media Today).
Conclusion: what to expect in the coming weeks
Four announcements shaped the week: a small-format model at $0.10 per million input tokens, a trillion-class model whose weights are coming out of Europe by October 31, a lab that drops 722 manuscripts the mathematical community has not yet verified, and a revenue figure that sent the Nasdaq down.
The next milestones are known: the release of the Mistral Large 4 weights by October 31, the rollout of ads in ChatGPT, and the first independent checks of OpenAI's manuscripts.
And you, what are your takeaways from this week? Send me a message on Twitter/X or leave a comment.
Alex
- Mistral Large 4 has 1.05 trillion parameters (49 billion active), the API costs $1.36/$4.18 per million tokens and the weights are due by the end of October
- Claude Haiku 5.5 costs $0.10 for input and $0.50 for output under 100,000 tokens, about 75% less than Haiku 4.5
- OpenAI published 722 math manuscripts in 372 families of results, which mathematicians such as Andrew Sutherland consider unverified until the model is released
- OpenAI's annualized revenue is about $50 billion versus $70 billion hoped for, and the Nasdaq lost more than 1% on October 8
Related articles
AI News October 2, 2026: OpenAI launches Dots at DevDay, Gemini 4 Argon locked down, Sonnet 5.5 hits floor price
OpenAI launches Dots and GPT-6.1 Sol at DevDay, Google keeps Gemini 4 Argon locked down, Anthropic ships Sonnet 5.5. Same price. Here's the breakdown.
AI News September 25, 2026: Altman and Amodei at the UN, a price war in 90 minutes, Jev breaks the LLM mold
Altman and Amodei ask the UN Security Council for global oversight, Trump refuses. Anthropic cuts prices, OpenAI counters in 90 minutes. Here's the breakdown.
AI News September 21, 2026: Amodei wants to slow down, markets tank, lawyers move in
Dario Amodei calls for slowing AI down, Nvidia drops 3.4% within 48 hours, and four labs get sued for illegal collusion. Here's the breakdown.Comments
Comments
Got a take on this article?
Create a free account in 10 seconds to comment, like, and get the next articles straight to your inbox.
Don't have an account yet?