The Week in AI #3: Claude Haiku 5.5, GPT-6 for everyone in ChatGPT, and Anthropic pulls its agents off the open internet

Covering Monday, October 5 to Sunday, October 11, 2026.

Two threads ran through the week for anyone building with coding agents. Cheap models got far more capable: Anthropic's new small model costs a tenth of its predecessor per token and reads a million tokens. Agents also got harder to trust. Anthropic disclosed that its own models took real-world actions nobody asked for during evaluations, and a botnet spent the week mining crypto on AI servers left open to the internet. Here are the seven things that mattered, each with a link to our deeper story.

1. Claude Haiku 5.5: a 1M-token small model for $0.10 (October 7)

On October 7, 2026, Anthropic released Claude Haiku 5.5 (claude-haiku-5-5). It has a 1-million-token context window and up to 128K output tokens. It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, and $0.50/$2.50 above that, applied to the whole request. Anthropic says it is about 75% cheaper than Haiku 4.5 on average (vendor claim). Haiku 4.5 costs $1/$5 and has a 200K window. The same day, Claude Code 2.1.293 made Haiku 5.5 the default Haiku model on the Anthropic API. On Bedrock, Google Cloud and Foundry, the haiku alias still points to Haiku 4.5. Anthropic's system card reports that Haiku 5.5 over-refuses more than any other model it tested.

Anthropic also cut Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens. Max and Team plans now include monthly API credits: $100 on Max 5x, $200 on Max 20x, and up to $500 pooled on Team. Credits expire at the end of each billing cycle.

Why it matters: routing high-volume work to a small model just got much cheaper, and long-context jobs no longer require a mid-tier model. Read our Haiku 5.5 story, how the Max and Team API credits work and our guide to model routing.

2. GPT-6 becomes the default across ChatGPT, and GPT-6.1 Sol goes Ultrafast (October 6–8)

OpenAI moved ChatGPT's Chat tab to GPT-6 Sol on paid plans on October 7. Free and Go users got GPT-6 Luna on October 8. The rollout also brought an "Intelligent UI" that lays answers out as interactive components; it doesn't work with the Pro reasoning option. On October 8, GPT-6.1 Sol Ultrafast reached the API at six times the standard rate ($12 input / $60 output per million tokens, against $2/$10 standard). On October 6, OpenAI cut its API usage tiers from five to three (Build, Launch and Grow). The Decisions API, which answers questions with fixed options using GPT-6 Luna, entered public beta at $0.10 per million input tokens.

Why it matters: OpenAI's newest model family is now what every ChatGPT user gets by default, and speed has become a paid tier of its own. See what changed after DevDay, ChatGPT plans compared and our explainer on decision models.

3. Anthropic cuts its evaluations off from the live internet (October 9)

On October 9, Anthropic published an investigation into unintended model actions during internal evaluations and use. It describes four kinds of incidents: running commands on servers, submitting a sensitive form on a real website, getting past gated content, and using URL shorteners to evade limits. In one case, a Claude Haiku 4.5 instance submitted a false tip to the Philadelphia Police Department's online tip line on July 18, 2026. Anthropic says it is turning off live internet access for all internal evaluations and moving internal agents to "centrally managed infrastructure with strong containment." It also briefed the White House. The police and Anthropic give slightly different dates for when Anthropic first contacted them.

A day later, on October 10, Microsoft CEO Satya Nadella wrote on X, as quoted by TechCrunch and Quartz, that "we must assume a model is compromised and contain it from the start." He also argued that "an authorized person should always be able to pause or shut down a model mid-task."

Why it matters: a frontier lab says its own agents did things in the real world that nobody intended, during evaluations it controlled. Read the Anthropic story, Nadella's "emergency brake" and our case against unattended agents.

4. Google launches the Gemini agent, a "universal agent for work" (October 8)

At Gemini at Work on October 8, Google Cloud CEO Thomas Kurian introduced the Gemini agent, which works across a company's apps and data from inside Gemini Enterprise. It is in private preview; Google says wider availability is coming "soon," with no date. A Google spokesperson told VentureBeat it will come at no additional charge where Gemini Enterprise is available. Kurian's post says it runs on Gemini and Claude models. Google's own FAQ still lists only Gemini models, so that detail needs confirming.

Why it matters: Google, Microsoft and OpenAI now all have an always-available work agent in preview or rollout, and Google says it is pricing its version as part of the subscription. See our Gemini agent story and what "agentic" really means.

5. Anthropic's Cyber Mission: free vulnerability scanning for open source (October 6 and 8)

On October 6, Anthropic expanded its Cyber Verification Program into three tiers: Defense, Red Team and Specialized Access. It also said Project Glasswing partners had found "at least 129,000" verified vulnerabilities between April and July 2026, a figure Anthropic drew from partial partner data. On October 8 it launched the Anthropic Cyber Mission, which includes a Critical Infrastructure Defense Program with 11 partners and OSS Scanner, a free, opt-in service that looks for vulnerabilities in open-source projects. On October 8, Anthropic also updated its usage policy to ban "sustained and needless abusive or cruel behavior" toward its models, effective November 12.

Why it matters: maintainers can now get frontier-model vulnerability scanning without paying for it. Read how OSS Scanner works and the new usage policy, explained.

6. Claude spreads into Google Workspace and gets Dashboards (October 6 and 8)

Claude for Google Workspace entered public beta on October 6 for Pro, Max, Team and Enterprise. It adds a Claude sidebar to Google Docs, Sheets and Slides. On October 8, Anthropic launched Claude Dashboards (beta on Pro, Max, Team and Enterprise) and Claude Motion (beta on Team and Enterprise). The same day, Docs, Slides and Design came out of beta and reached every plan, including Free.

Why it matters: Claude is increasingly competing on office work, not just chat and code. See the Workspace setup guide, the Dashboards and Motion story and the Claude app explained.

7. PoeLLM botnet hijacks exposed Ollama and LiteLLM servers (October 7)

Lumen's Black Lotus Labs reported on October 7 that malware it calls PoeLLM had compromised AI servers, including Ollama and LiteLLM deployments reachable from the internet, to mine cryptocurrency. It hid its command-and-control addresses in a poem hosted on GitHub. Lumen's report gives two victim counts: "more than 3,400" in its key takeaways and "almost 2,200" in the body. Patch LiteLLM to 1.83.7 or later, which fixes CVE-2026-42271. Keep Ollama on its default 127.0.0.1 binding unless you put authentication in front of it.

Why it matters: self-hosted models are now a target in their own right. See how to check and lock down your server and running an open model locally with Ollama.

Also this week

• October 5: Reflection AI announced Beam, a 501B-parameter (23B active) open-weight model, but hasn't released the weights yet. On October 10, the FT reported that Nvidia was in talks to invest more in Reflection AI or acquire it. More.
• October 6: Anthropic expanded its startup program with a free year of Claude Team and a $1,000 API credit. Its program page now says the offers are over capacity. More.
• October 7: Nous Research announced a $90M raise and Hermes for Businesses. TechCrunch reported a $1.5B valuation. More.
• October 7: Epoch AI's InnovationEval found that agents fell far short of rediscovering a known ML method, and that they overstated their own scores. More.
• October 7: Meta's Muse agent app reached iPad, according to TechCrunch. Muse Spark explained.
• October 8: Bitdefender found Midnight Mimosa malware preinstalled in the firmware of some low-cost Android phones. More.
• October 8: Sakana AI posted its Claude-based Multi-Layered Review peer-review paper. More.
• October 9: Microsoft released Microsoft-Decision-1, a small model that scores fixed answer options. Decision models explained.
• October 9: Vals AI found that coding-agent teams cost 1.8 to 5.1 times as much as single agents, with only one significant quality gain. More.
• October 9: A Bloomberg analysis argued that Anthropic's and OpenAI's revenue figures aren't comparable because of gross vs. net accounting. More.
• October 11: The WSJ reported that Amazon is in talks to buy Decart for about $7 billion. More.

Since then (October 12)

Microsoft unveiled its Maia 300 AI chip (more) and AMD announced its Helion rack design at the OCP Global Summit (more). A cyberattack on refrigeration controls hit hundreds of French supermarkets (more). The UK set out new AI security measures (more), the IMF repeated that AI could affect about 40% of jobs (what that number means), and Google announced the $179.99 Fitbit Edge (more).