Three AI stories landed within 48 hours this week. Anthropic made its smallest model far cheaper, OpenAI pushed GPT-6 to every ChatGPT tier, and OpenAI also published a huge batch of mathematics from a model it has not released.
Anthropic cuts the price of its small model
Anthropic released Claude Haiku 5.5 on October 7, calling it the cheapest, fastest and most capable small model it has shipped. The company says an average workload costs about 75% less to run than on Haiku 4.5.
The per-token numbers are even starker. Haiku 4.5 was priced at $1 per million input tokens and $5 per million output tokens. For prompts up to 100,000 tokens, Haiku 5.5 starts at $0.10 and $0.50, a 90% cut. Longer prompts cost $0.50 and $2.50. Anthropic says about 90% of requests to the previous Haiku fell under the 100,000-token line, which is how the blended figure lands near 75%.
Price is not the only change. On the offline subset of OSWorld 2.1, a computer-use benchmark, Anthropic reports 72.4% for Haiku 5.5 against 15.7% for Haiku 4.5. It is also the first Haiku with adjustable “effort” levels, so developers can trade speed against reasoning depth. The target is high-volume work: summarization, classification and sub-agents.
For comparison, GPT-6 Luna, the model OpenAI now serves to ChatGPT’s free tier, costs the same $0.10 and $0.50 per million tokens through the API. That parity holds only for shorter prompts: Luna switches to a higher rate ($0.20 and $0.75) above 272,000 input tokens, while Haiku 5.5 does so above 100,000.
GPT-6 reaches every ChatGPT plan
OpenAI began rolling out GPT-6 in ChatGPT on October 7 for Plus, Pro, Business and Enterprise subscribers, who get a model called GPT-6 Sol. Free and Go users were next in line on October 8, receiving a lighter version named GPT-6 Luna. The rollout is phased, so not every account will see it immediately.
The launch comes with “Intelligent UI,” which lets ChatGPT answer with interactive charts, maps, calculators and forms instead of plain text. Users can dial the feature back in settings. OpenAI also claims GPT-6 Instant shows live web-search answers 44% faster than GPT-5.6 Instant.
372 families of math results, and a lot of questions
The most debated news is the least tangible. On October 6, OpenAI put 722 manuscripts, grouped into 372 families of results, in a public GitHub repository. All of it came from an unreleased internal model that OpenAI has not named. Many papers include Lean proofs, which a computer can verify, though not all do.
OpenAI says the model was given roughly 4,000 problems and that the average result used compute equal to about three hours of ChatGPT Pro thinking. A spokesperson told Scientific American that nearly all results came from a single prompt given to a single AI agent, though some needed several attempts. That is a sharp drop from the 10,000-agent swarm, costing millions of dollars, that OpenAI used for an earlier Navier-Stokes result.
Mathematicians are cautious. OpenAI admits some unformalized results may contain errors, and the warning proved timely. On October 7, a day after the release, it withdrew three manuscripts after a sign error invalidated one proof and two papers that depended on it, and revised 14 more. The catalogue now lists 719 manuscripts, still in 372 families, with withdrawn versions kept online alongside explanatory notices. An advisory group of mathematicians had asked labs to disclose the model, the prompts and per-problem compute, but OpenAI has released only averages. MIT’s Andrew Sutherland said claims of single-agent, one-shot success should be treated as unverified until the model is released and others can replicate them.
What it adds up to
Together, the stories show the market moving along two tracks. At the low end, strong models are getting cheap enough to run at scale. At the top, labs are claiming research-level results that outsiders cannot yet check. Verification, not generation, may become the bottleneck.