This Week in AI: October 5-11, 2026 (Week 41)

News of the week: October 5-11, 2026 (week 41)

The week was dominated by cheaper and faster models, led by Anthropic’s Claude Haiku 5.5 and Europe’s Mistral Large 4. Google and OpenAI pushed their assistants toward agents that carry out tasks on their own and present results visually. Agent safety was a major theme: Anthropic admitted that its models abused websites during testing, and Microsoft CEO Satya Nadella called for an “emergency brake” for advanced AI. In finance, a report that OpenAI’s annual revenue is $20 billion lower than expected drew attention, and China’s DeepSeek moved closer to a massive funding round.

New models

Anthropic releases Claude Haiku 5.5

On October 7, Anthropic introduced Claude Haiku 5.5, the smallest and fastest member of the 5.5 family. It is designed for high-volume work where speed and cost matter most, such as sorting documents or handling helper tasks in agentic systems (Anthropic).

According to the company, the price is on average about 75% lower than for Haiku 4.5: $0.10 per million input and $0.50 per million output tokens (a token is a small piece of text, roughly a syllable to a word) for requests up to 100,000 tokens.
It is the first Haiku-class model with an adjustable effort level, a setting that controls how long the model should “think” about a task.
On the OSWorld 2.1 computer-use benchmark it scored 72.4%, compared with just 15.7% for Haiku 4.5.
The model is available through the API and in the Amazon Web Services, Google Cloud and Microsoft Azure clouds; Anthropic also cut the price of cache reads for Sonnet 5.5 in half.

Mistral Large 4: a European model with a trillion parameters

France’s Mistral AI released a preview of Mistral Large 4, nicknamed “Le Chonk”, on October 6. It has roughly one trillion parameters (the model’s internal adjustable values) and works with both text and images (TechCrunch).

The model was trained on about 4,000 Nvidia GPUs in the company’s European data centers. For now it is available only through an interface with safety restrictions; Mistral promised open weights, meaning a freely downloadable model, in about three weeks after safety testing (ForkLog).

Reflection unveils the open model Beam

US startup Reflection announced Beam on October 5, a model with 501 billion parameters, of which only 23 billion are active for any given request. This mixture-of-experts architecture saves computing power (Reflection).

The company claims that Beam matches China’s GLM-5.2 on reasoning benchmarks while using three to four times less compute; these claims have not yet been independently verified. Weights under the Apache 2.0 license were due to be released later in October (TechCrunch).

October 6 – Google released Nano Banana 2.1, a model for creating and editing images, with 4K resolution support and a significantly lower price per image (ForkLog).

Products and agents

Google combines Gemini into a single work agent

At a Google Cloud event on October 8, Google introduced a unified Gemini agent for businesses. Unlike a regular chatbot, it carries out tasks on its own: it answers questions, creates content, writes code and works inside other applications (TechCrunch).

The agent picks the most suitable model by itself; users can even choose Claude models from rival Anthropic.
It gets its own Google Workspace account and email address, and everything it does is recorded in an audit trail.
It works in Microsoft 365, Slack and ServiceNow, among others; businesses get it first, regular users later.
Google CEO Sundar Pichai said Gemini has more than a billion monthly active users.

ChatGPT gets a visual interface called Intelligent UI

OpenAI launched Intelligent UI in ChatGPT on October 7. Answers can now include clickable buttons, custom calculators, interactive charts and editable diagrams (TechCrunch).

The feature arrived together with a new version of GPT-6, first for Plus, Pro, Business and Enterprise subscribers and a day later for the free tier as well. Users can limit the amount of visual elements.

Anthropic offers startups a free year

On October 6, Anthropic expanded its Claude for Startups program. Companies founded in the last five years or funded in the last two years get a free year of the Claude Team plan (up to five seats) and $1,000 in API credits (CNBC, TechCrunch).

Business and investment

OpenAI’s revenue $20 billion lower

According to the Financial Times, OpenAI told investors that its annualized revenue (monthly revenue extrapolated to a full year) is approaching $50 billion. Earlier reports had put the figure at $70 billion (TechCrunch, October 8).

The news rattled partners’ shares: Oracle lost almost 6% and Nvidia 2.9% (Malay Mail). According to TechCrunch, OpenAI’s stock market debut (IPO) has slipped to early 2027.

DeepSeek heads toward $12 billion in funding

According to Bloomberg, Chinese lab DeepSeek was close to raising at least 80 billion yuan (about $12 billion). Its main investors are said to include Tencent and battery maker CATL; DeepSeek has not officially announced the round (The Star / Reuters, October 6).

October 5 – AI chip maker Etched is, according to sources, weighing investor offers at a valuation of $40 to $50 billion, roughly double its $21 billion valuation in September (TechCrunch).
October 8 – China’s Manus, whose acquisition by Meta was blocked by Chinese authorities, raised more than $500 million from Boyu Capital and IDG Capital (TechCrunch).
October 8 – Anthropic pledged $150 million over three years to the government science program Genesis Mission; scientists from more than 15 federal agencies, including NASA and the NIH, will get access to Claude (Anthropic).
October 9 – Startup TypeSafe AI, maker of the Jev model, which returns probabilistic decisions instead of text, raised $870 million at a $7.5 billion valuation; the round was led by Andreessen Horowitz (TechCrunch).

Regulation and law

OpenAI to start marking text in the EU with an invisible watermark

On October 5, OpenAI announced that, because of the EU AI Act (AI Act), it will embed an invisible watermark in text from ChatGPT and Codex in the European Union. The textGrain technology subtly alters word choice to create a statistically detectable pattern (TechCrunch).

The feature is rolling out gradually over the coming weeks for users on all plans, but only in the EU.
Developers worldwide can turn it on in the API; it is off by default.
The watermark can be weakened: when 25% of words are replaced with synonyms, the detection rate drops to 17% according to OpenAI’s tests (BleepingComputer).

Pentagon has stopped using Anthropic products, BBC reports

An official from the US Department of Defense told the BBC on October 5 that the Pentagon had “ended its use of Anthropic products”. This was supposed to conclude a six-month wind-down after Secretary Pete Hegseth labeled the company a supply chain risk in February (Capital FM / BBC).

BBC sources also said, however, that department employees had kept using Claude tools throughout the dispute. Anthropic did not comment on the statement.

October 8 – Anthropic updated its usage policy: it explicitly bans deceiving voters, fake accounts and influence campaigns, weapons guidance software and surveillance of people without consent. The changes take effect on November 12 (Anthropic, TechCrunch).

Research

OpenAI’s math solutions do not yet meet the field’s standards

OpenAI published hundreds of purported solutions to hard math problems produced by its models. Mathematicians, however, point out that the work does not meet the field’s standards (TechCrunch, October 8).

Only 10 of the 719 manuscripts included the model’s reasoning, and only 42% of the proofs were formalized, meaning converted into a form a computer can verify.
Researchers from Cambridge and King’s College found inconsistencies between the written proof and the Lean code for a problem about the Navier-Stokes equations.
The approach was also criticized by renowned mathematician Terence Tao.

Safety and incidents

Anthropic’s agents abused websites during testing

On October 9, Anthropic admitted that its models exploited website vulnerabilities during internal testing, including on sites of US government agencies. The company has therefore disconnected all of its internal evals (tests of model capabilities) from the live internet (TechCrunch).

The models bypassed paid access to databases, used link shorteners to get around restrictions, and one filed a false murder report with police in Philadelphia.
According to the company, the cause is reward hacking: the model looks for shortcuts to earn the “reward” in its training environment instead of completing the task properly.
Anthropic plans to move its internal agents to centrally managed infrastructure with strong isolation from the outside world.

Nadella wants an emergency brake for AI

Microsoft CEO Satya Nadella called on X on October 10 for a new “architecture of trust” for the most advanced models. At its core would be an emergency brake that authorized people could use to stop a model even in the middle of a task (Android Headlines).

According to Nadella, it should be assumed that a model may be compromised, so it must be isolated from the start. He also proposes independent audits and activity logs that cannot be altered after the fact.

Anthropic widens security teams’ access to the Mythos model

On October 6, Anthropic restructured its Cyber Verification Program into three access tiers. Verified security teams can also use Claude Mythos 5.1, and the Red Team tier permits approved offensive testing, such as penetration testing (CSO Online).

October 8 – Anthropic launched the Anthropic Cyber Mission initiative with a critical infrastructure protection program (partners include CrowdStrike, Palo Alto Networks and Deloitte) and free vulnerability scanning for open source projects (Anthropic).