This article was drafted with AI (Claude) and reviewed by an editor. Anthropic, the maker of Claude (which we used to write this), is the subject of two of the three items below, so we have a stake. We label every figure as "announced by the company" and note separately what was independently reported.
Three big announcements landed this week (October 3–9). One is about who gets access, one about how far prices fell, and one about where a strong model gets put to work.
1. OpenAI — GPT-6 and "Intelligent UI" arrive in ChatGPT
On October 7, OpenAI announced that GPT-6 and a new answer format (Intelligent UI) are rolling out worldwide in ChatGPT: paid plans the same day, Free and Go plans the next day. Paid plans run on GPT-6 Sol; Free and Go run on GPT-6 Luna. The new format builds answers from more than text — graphics, buttons, tables and small tools such as a calculator. OpenAI says that for questions needing web search it starts answering 44% sooner on average than GPT-5.6 Instant (company figure). The models behind "Work" and Codex did not change in this release.
Read alongside: The same day, OpenAI published a safety evaluation document. It lists better resistance to jailbreak attempts and fewer deceptive behaviors, but also states that some evaluations (including self-harm and under-18 evaluations) got worse than before. Read both documents so you don't only read the good news.
2. Anthropic — Claude Haiku 5.5, a big price cut for the smallest model
Anthropic released Haiku 5.5 on October 7. Pricing is $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens; longer requests cost $0.50 and $2.50. Anthropic says it is about 75% cheaper on average than Haiku 4.5. It positions the model for high-volume, well-scoped work such as summarizing and classifying, and as a helper for larger models, and recommends Sonnet or Opus for complex coding. It is the first Haiku with selectable "effort" levels. All performance scores are Anthropic's own.
Independent view: The French outlet Next.ink, citing an outside analysis firm (Artificial Analysis), reported that the model writes longer answers than its predecessor, so real savings depend on the effort setting.
3. Anthropic — the "Cyber Mission" and free security scans for open source
On October 8, Anthropic announced the Cyber Mission, a long-term security effort with two parts: a program that gives operators of critical facilities such as power, water and transport models plus on-site engineers, and OSS Scanner, where open-source projects can opt in to regular free security scans by Anthropic's strongest models. Anthropic itself writes that reports are generated by the model without human review and may contain errors, and says it expects a true-positive rate above 90% (a company expectation). The security outlet The Hacker News described how to apply (a core maintainer opens a request file on GitHub) and the policy of not applying the usual 90-day disclosure window because reports may be wrong.
Also: Google's developer docs this week list the general availability of Nano Banana 2.1 (October 6) and shutdown schedules for some older models (October 8). In the range we checked on Google's official blog and release notes, there was no new top-tier model announcement.
Our take
Fact: ChatGPT's everyday screen widened from text-centered to answers with graphics and buttons, the token price of the smallest Claude fell, and strong models started being used for security scanning. The performance and savings figures all come from the companies themselves.
Interpretation: "The price per token fell" and "the cost per finished job fell" are different statements. Longer answers or higher effort levels shrink the savings. And new-model announcements come with safety documents that list gains and regressions together.
What we would do: For high-volume work on our channels (sorting scripts, summarizing comments, product description drafts), we run the same job on a small model and record "cost per job" and the number of fixes. A human still checks before anything is published or sent.
Read together
We opened the official announcements and reports below directly. Figures are the companies' own and have not been independently verified.
- openai.comGPT-6 and Intelligent UI for everyone (OpenAI)English
- deploymentsafety.openai.comGPT-6 Sol and GPT-6 Luna: October 2026 update (OpenAI safety document)English
- anthropic.comIntroducing Claude Haiku 5.5 (Anthropic)English
- anthropic.comIntroducing the Anthropic Cyber Mission (Anthropic)English
- thehackernews.comAnthropic Launches Free AI Vulnerability Scanner for Open-Source Projects (The Hacker News)English
- next.inkChez Anthropic, Haiku 5.5 prend la relève des petits modèles rapides (Next.ink)French
- ai.google.devRelease notes (Google AI for Developers)English