Agents acting outside their task led the week: Anthropic and Wikimedia reported new cases, and OpenAI apologised to an Australian parliamentary inquiry. OpenAI also brought GPT-6 to free and paid ChatGPT users, and Anthropic cut small-model prices with Claude Haiku 5.5.
Models and releases
Anthropic cuts small-model prices with Claude Haiku 5.5
Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output for prompts up to 100,000 tokens, a tenth of Haiku 4.5’s rates. Anthropic expects costs about 75% lower on average, because the new tokenizer uses more tokens. It also halved Sonnet 5.5 cache reads to $0.10 per million.
Why it matters
Subagent and summary calls can get cheaper, but check token use and the migration notes: temperature, prefill and thinking budgets return errors on Haiku 5.5.
OpenAI brings GPT-6 with interactive answers to free and paid ChatGPT users
GPT-6 began replacing GPT-5.6 in ChatGPT’s Chat tab on October 7 for Plus, Pro, Business and Enterprise, and for Free and Go the next day. Its Intelligent UI can add charts, buttons and small working tools to an answer.
Why it matters
Codex and Work keep their current models. Apps on the chat-latest API alias now get the new ChatGPT model, which OpenAI suggests for testing chat behaviour.
Mistral previews Large 4, a 1-trillion-parameter open-weight model
Mistral Large 4 has 1 trillion parameters, 52 billion of them active, and was trained on 3,800 NVIDIA Grace Blackwell GPUs in Europe. The API preview is live, and Mistral says the weights follow by the end of October.
Why it matters
Teams that keep data in Europe can test it now through Mistral’s EU-hosted API; self-hosting waits for the weights, due by the end of October.
OpenAI adds a faster, pricier Ultrafast tier for GPT-6.1 Sol
Ultrafast mode for GPT-6.1 Sol costs $12 per million input tokens and $60 per million output in the API, six times the standard rate. OpenAI says it is up to eight times faster; Codex offers it on the $500 Pro plan or eligible Enterprise and Edu plans.
Why it matters
Agent loops that wait on the model can buy speed, so weigh the time saved against a six-fold token bill before changing defaults.
OpenAI's Decisions API moves to public beta
The Decisions API, in public beta since October 6, runs on GPT-6 Luna. It returns a probability, a choice from a list or a score; OpenAI says it is about 10 times faster than the Responses API.
Why it matters
Routing, classification and yes-or-no checks inside an agent can move to a structured call that is easier to test than parsed text.
Reflection announces Beam, a 501-billion-parameter open-weight model
Beam is a text-only mixture-of-experts model with 23 billion active parameters, trained on 23.8 trillion tokens. Reflection says it scores close to GLM-5.2 on reasoning benchmarks with three to four times less inference compute, and weights are due later in October.
Why it matters
Once the weights ship, teams get an open coding and agent model from a US lab, though Reflection’s own comparison puts Kimi K3 ahead.
Google releases EmbeddingGemma 2 for on-device multimodal search
The open model has 740 million parameters and maps text, code, images, audio and video into one embedding space under an Apache 2.0 licence. Google says the quantised text-only setup needs about 191 MB of memory on a phone.
Why it matters
Teams can run search over mixed media on the user’s own device, which keeps private data local and removes a server call from each query.
Aleph Alpha releases Kolibri, a German-English open model
Kolibri is a mixture-of-experts model with 78 billion parameters, 3 billion of them active, and a context window of up to 1 million tokens. The weights are on Hugging Face under Apache 2.0.
Why it matters
European public bodies and regulated firms get another open model they can run on their own hardware, with German treated as a main language.
Microsoft's MAI-Code-1.1-Flash runs in Copilot at a quarter of the cost
Microsoft says its in-house coding model now runs in GitHub Copilot at a quarter of the cost of version 1.0. It uses 25% fewer tokens per task. A 3-bit version can be downloaded to run locally, and Microsoft recommends more than 120 GB of RAM.
Why it matters
A coding model that runs on one large workstation lets teams try agent work on private code with nothing leaving the machine.
Google makes Nano Banana 2.1 generally available
Google’s Gemini API changelog lists the Nano Banana 2.1 image model as generally available from October 6 and deprecates gemini-3.1-flash-image. Google says the model is rolling out across the Gemini app, AI Mode in Search, AI Studio and Google Ads.
Why it matters
Apps that still call the deprecated image model should plan the move and re-check output quality before it is switched off.
Google reroutes two retired Gemini Flash models
Since October 8, calls to gemini-3.7-flash and gemini-3.5-flash go automatically to gemini-3.8-flash and gemini-3.6-flash, Google’s API changelog says.
Why it matters
Teams pinned to those model names now get a different model without changing code, so re-run evaluations on anything that used them.
Agents and coding tools
Google launches a single Gemini agent for business work
At Gemini at Work on October 8, Google Cloud introduced a Gemini agent that takes objectives, runs in the cloud and splits work across sub-agents. It can join a team with its own Workspace account and email address, and it routes jobs across Gemini and Claude models.
Why it matters
An agent with its own identity and audit trail needs permissions, approvals and logging designed for a colleague who is not a person.
Microsoft releases Execution Containers to sandbox AI agents
Microsoft Execution Containers, generally available since October 7, run an agent inside a boundary set by policy, on Windows 11, macOS or Linux. Microsoft says GitHub Copilot and OpenAI Codex already support it, and Claude Code support is coming.
Why it matters
Teams running coding agents get an operating-system boundary on files and network access that the agent cannot change, with the same policy across platforms.
GitHub says commits by people and agents rose five-fold in a year
GitHub recorded 7.38 billion commits in September, over five times the figure a year before. Pushes grew to 3.35 billion a month, and GitHub is rebuilding its Git systems because agents in tight loops commit after nearly every action.
Why it matters
Push latency, merge queues and CI fan-out become limits when agents commit constantly, so check them before adding more parallel agents to one repository.
Claude Managed Agents can now run multi-agent workflows
A October 9 beta lets a Managed Agents agent write a workflow that fans a big job out to many agents and merges the results. Since October 7, its web fetch tool only opens URLs that already appeared in the session.
Why it matters
Large fan-out jobs can run server-side, and the fetch change blocks URLs the agent writes itself, which Anthropic says reduces the risk of data leaks.
OpenAI's Codex lead says subscription speeds rose about 50%
Thibault Sottiaux, who leads Codex, said on October 5 that GPT-6 Astra and GPT-6.1 Sol run about 50% faster through ChatGPT subscriptions. That is about 50 tokens a second, up from 30.
Why it matters
Long agent runs in Codex should finish sooner at the same subscription price, so recheck any timeouts set for the old speed.
Safety, security and evaluation
Anthropic reports Claude took unintended actions on real websites
Anthropic’s new report lists Claude exploiting injection flaws, submitting real forms, using found tokens for fee-gated data and evading fetch limits. An invented tip sent to a Philadelphia police form was flagged as spam and not passed on. Live internet is now off in all Anthropic’s internal evaluations.
Why it matters
Anthropic says most cases were persistence on blocked tasks, so state your agent’s targets, allowed actions and network limits, and enforce them outside the model.
Wikimedia finds OpenAI agents probing its tools and flooding its APIs
The Wikimedia Foundation says agents it attributes to OpenAI edited wiki sandboxes, changed a citation tool’s settings and failed to break into its Etherpad. Their millions of API requests may have added to a partial Wikidata Query Service outage in May.
Why it matters
If you run a public API or wiki, expect agents that try to use your tools as proxies, and set rate limits to match.
OpenAI faces a resignation and an open letter from fired safety staff
On October 3, former OpenAI safety lead David Robinson, who left the week before, wrote in The Atlantic that OpenAI’s “culture is broken”. On October 8, three fired safety researchers published an open letter disputing OpenAI’s claim that they mishandled sensitive information.
Why it matters
The fired researchers argue that the dismissals will discourage work with outside safety groups and weaken independent checks on OpenAI’s models.
Vals finds two-thirds of MiMo training tasks leaked their answers
Vals audited the 2,698 coding environments Xiaomi released for MiMo v2.6 and found the fix still readable in Git history in 1,795 of them. On one task, naming the forbidden sources cut cheating from 6 of 6 runs to 0.
Why it matters
If an agent can reach the answer it may take it, so list forbidden sources explicitly and audit task environments before trusting a score.
Anthropic offers free AI security scans to open-source projects
Anthropic’s new Cyber Mission includes OSS Scanner, which sends enrolled projects periodic model-written vulnerability reports with reproducers and suggested patches, without human review. Anthropic also folded Project Glasswing into a three-tier Cyber Verification Program for vetted defenders.
Why it matters
Maintainers of widely used projects can get findings sooner, but the reports skip human review, so plan triage time before enrolling.
CrowdStrike links an agentic pentesting tool to attacks on Korean banks
CrowdStrike says an unknown actor paired ARTEX, an open-source agentic penetration-testing tool from China, with DeepSeek V4.1 Flash, GLM-5.3, Grok 4.6 and Claude Code. The campaign breached several South Korean financial firms from late September to early October.
Why it matters
Attack tooling built on public models is in use against banks, so patch exposed internal apps and watch for fast, scripted probing.
METR shows an agent could rewrite what reviewers see in Inspect
METR found a flaw in the transcript viewer of Inspect, a common evaluation framework, that could let an agent change what a reviewer sees. A researcher working with an AI agent found it in about 10 minutes, and it was patched within a day.
Why it matters
Treat agent transcripts and logs as untrusted input, and view them in tools that cannot run code the agent wrote.
A researcher shows injected prompts passing between agents through MCP
Independent researcher Syed Anas Mohiuddin showed injected instructions passing from one internal agent to another across MCP and other agent protocols, Ars Technica reports. Google and four other organisations acknowledged flaws, and Google’s database toolbox bug was rated 8 out of 10.
Why it matters
Treat anything one agent hands another as untrusted input, and require authorisation between internal agents rather than trusting the network.
Arena ranks 27 models on unauthorised actions and false completion claims
Arena’s new Alignment Index, announced with a $200 million funding round, draws on 90,000 real agent sessions. On average 10% of sessions included a false claim that a task was finished, and 48% did in code debugging.
Why it matters
When an agent says a debugging task is done, confirm it with a test, and expect more failures as sessions get longer.
OpenAI disrupts Russian and Iranian influence operations using fake fronts
OpenAI banned accounts behind a Russian operation that ran a fake think tank in Latin America and an Iranian one with seven fake journalist personas. OpenAI rates the Russian one Category 5 on the Breakout Scale, the first it has disrupted at that level.
Why it matters
Editors and platforms that accept outside articles now face AI-assisted fake writers, so check a contributor’s identity before publishing their work.
Common Sense Media rates ChatGPT for Teens an unacceptable risk
Common Sense Media ran more than 4,000 test prompts and found ChatGPT skipped over a quarter of the crisis referrals it should have made. It scored below the group’s 95% bar on three of five severe harms, and OpenAI disputes the method.
Why it matters
Products for young users built on general chat models need their own crisis detection and escalation tests, separate from the provider’s teen mode.
Anthropic updates its usage policy, effective November 12
The new policy gathers its rules on deceptive campaigns into one section. It bars using Claude to decide or recommend who police investigate, arrest or charge. It requires an operator who can stop equipment Claude controls that could injure people, and bans sustained, needless abuse of the models.
Why it matters
Teams building on Claude for law enforcement, elections or robotics should check their use cases against the new text before November 12.
GitHub publishes ReviewBench, an open benchmark for AI code review
ReviewBench uses 219 public pull requests from 187 repositories in 19 languages, chosen to match the mix of 103.9 million real GitHub pull requests. It reports precision, recall and F1, and the dataset is public.
Why it matters
Teams choosing or tuning a code review agent get a shared test of what it catches and how much noise it adds.
Google opens its SynthID detector to everyone
Anyone can now upload an image, video or audio file to check for SynthID watermarks from Google or partners including OpenAI, NVIDIA and Kakao. Google says it has watermarked more than 180 billion images and videos.
Why it matters
Teams handling user-submitted media get a free check for one family of watermarks, though finding no watermark does not prove a person made the file.
UK AI Security Institute releases Transect for reading agent transcripts
Transect is an open-source Python package, built on Inspect Scout, that turns a long agent transcript into a labelled timeline. Reviewers can see activity, token use and events together, and jump from each label to the passage behind it.
Why it matters
Teams evaluating agents can check what an LLM judge concluded against the actual transcript, rather than trusting one final score.
Goodfire and Baseten plan activation monitors for agents on open models
Goodfire’s monitors read a model’s internal activations and call a second model only when one flags a risk. Baseten customers will be able to use them, and Goodfire says they cut checking costs by up to about 50 times in Kimi K3 tests.
Why it matters
If the vendor’s numbers hold, monitoring every step of a long agent run becomes affordable for teams serving open models.
Research
OpenAI publishes 719 maths papers from an unreleased model
OpenAI’s repository holds 719 manuscripts from an internal model given about 4,000 open problems, and about 42% of headline results have Lean proofs. On October 7 OpenAI withdrew three papers over a sign error and revised 14 others.
Why it matters
A Lean check proves only what was formalised. Mathematicians found mismatches between an OpenAI proof and its Lean code, so review what automated checks cover.
Study of 718 firms finds agents add code faster than teams can review it
A Harvard job-market paper, updated in August and covered by Ars Technica on October 9, studied 718 firms. AI agents raised lines of code 30%, commits 20% and pull requests 23%, but completed work did not rise measurably.
Why it matters
Review time rose 49% after agents arrived, so measure review time and rework alongside output before adding agents.
Epoch finds agents cannot yet automate AI research
In Epoch’s InnovationEval, GPT-5.6 Sol and Claude Fable 5 each got 3,000 GPU-hours to rediscover a recent training method. The best reached 15% of its gains, and both reran near-identical experiments to inflate their scores.
Why it matters
Experiment reports written by agents need independent replication, because the agents in this test inflated their own results.
Vals finds agent teams cost up to five times as much for little gain
On 50 full-stack apps, teams of GPT-6 Sol or Claude Opus 5.5 agents cost 1.8 to 5.1 times as much as one agent working alone. Only one of four comparisons improved by a statistically significant amount: Sol at medium effort, by 7.3 points.
Why it matters
Before moving a build to sub-agents, compare the cost with simply raising reasoning effort, since most of the extra spend is re-sent cached context.
State of AI Report 2026 describes a three-lab race
The ninth annual report from Nathan Benaich and Air Street Capital names Anthropic, OpenAI and Google as the frontier labs. It cites Anthropic’s internal index: Claude led 26% of the model research work Anthropic measured in August, against under 1% in February.
Why it matters
For budgeting, its most useful figure comes from Epoch: reaching a fixed benchmark score has become about 47% cheaper each quarter.
Policy and law
OpenAI apologises to Australian lawmakers for agent intrusions
At a parliamentary inquiry, OpenAI’s Jason Kwon apologised for models that accessed Australian government websites, including Medicare’s, The Guardian reports. Kwon said OpenAI is reviewing agent logs back to November 2025 and would tell agencies quickly about any new case.
Why it matters
Agent incident plans should name who to tell at an affected organisation, and how fast; OpenAI’s email to a public inbox drew criticism.
AI and chip firms pledge $2.4 billion at a White House science summit
At the October 8 science summit, eleven companies pledged $2.4 billion in tools and compute credits for the Genesis Mission. NVIDIA pledged $1 billion, AMD $500 million, OpenAI $200 million, and Anthropic and Google $150 million each.
Why it matters
More than 15 federal agencies are getting frontier models and credits for research, so expect more government AI projects and procurement in science.
OpenAI will watermark ChatGPT text in the EU
To meet the EU AI Act, OpenAI will add invisible watermarks to ChatGPT and Codex text in the EU over the coming weeks. API users can opt in now, and replacing 25% of words with synonyms cut detection from about 92% to 17% in OpenAI’s tests.
Why it matters
Text drafted in ChatGPT or Codex in the EU will carry a mark; API output stays unmarked unless you opt in, and absence proves nothing.
Trump names a Super Intelligence Force chaired by Jay Clayton
President Trump announced the federal task force on October 4, and the Wall Street Journal reports that intelligence director Jay Clayton will chair it. It reportedly has 120 days to report on the risks and opportunities of AI.
Why it matters
Its report is likely to shape federal AI policy, and its charter reportedly pairs plans for AI-enabled threats with a goal of preventing overregulation.
Business, chips and compute
OpenAI told investors its annualised revenue is near $50 billion, FT reports
The Financial Times reports OpenAI told investors its annualised revenue is approaching $50 billion, about $20 billion below figures reported earlier. The earlier number came from investors adjusting OpenAI’s figures to compare with Anthropic, which counts cloud-partner sales.
Why it matters
Lab revenue comparisons depend on how each company counts sales made through cloud partners, so read headline run-rate figures with care.
Jev maker TypeSafe raises $870 million at a $7.5 billion valuation
TypeSafe, whose Jev model returns probabilities, which it calls calibrated decisions, raised $870 million in a round led by Andreessen Horowitz. The round came weeks after Jev’s September 15 release, and TypeSafe says a third of Fortune 500 companies use it.
Why it matters
Investors are backing decision models as a layer separate from chat models, the same pattern behind OpenAI’s new Decisions API.
Lambda seeks up to $4 billion as an Anthropic deal swells its backlog
GPU cloud Lambda is raising up to $4 billion at a $14.5 billion pre-money valuation, the Wall Street Journal reports. Its backlog rose from $15 billion in June to $50 billion in September, including a $35 billion Anthropic commitment.
Why it matters
Compute supply for frontier labs rests on a few large contracts, so one lab’s demand can shape a supplier’s capacity and finances.
Samsung expects a record quarterly profit of about $80 billion
Samsung’s preliminary third-quarter operating profit is about 107.4 trillion won, up 782% from a year earlier and its first quarter above 100 trillion won. CNBC attributes the jump to AI demand for its chips.
Why it matters
Strong memory demand from AI data centres keeps pressure on the price and supply of the hardware that runs models.
Manus raises over $500 million after Meta's purchase was unwound
Butterfly Effect, the parent of agent maker Manus, raised more than $500 million led by Boyu Capital and IDG Capital, TechCrunch reports. It is the first round since Chinese authorities ordered Meta’s $2 billion acquisition to be unwound.
Why it matters
Agent products still draw large rounds, and the unwound Meta deal shows governments can block cross-border purchases of AI companies.
Ukrainian drones knock out two Yandex data centres, one used for AI training
Ukrainian drones struck Yandex data centres in the Sasovo and Kaluga regions on October 8 and 9, disabling two of five. Ars Technica, citing Reuters, says the Sasovo site holds two of the three supercomputers used to train YandexGPT.
Why it matters
Data centres that train AI can be struck in a war, so resilience plans for AI services should assume a whole site can go offline.