Items in this briefing come from OpenAI, Simon Willison and arXiv.org.
The Signal
Cheaper AI doesn't always mean a cheaper job. Simon Willison notes Haiku 5.5's price jumps fivefold once a prompt passes 100,000 tokens, and a new paper found most AI-built cheaper solutions fell short of the model's own quality, though some saved a lot. Test it on your own workload.
Why this matters: Teams building agents that browse the web can weigh the authors' case for training against attackers who adapt, rather than a fixed list of attacks, as a way to resist planted instructions.
Summary: This is a research paper on web agents, which read pages written by third parties and can be redirected by instructions planted on those pages. The authors say current defenses train against attacks fixed in advance, and attackers who adapt can get past them. Their method, AdvSim2Real, trains a task curriculum, an attacker and the agent together inside a frozen web world model, a simulator the agent trains in. They report that a 4B agent completes more tasks with and without attacks, and that the capability gain carries over to a real browser. In the simulator, on 150 web tasks, they report a 33.6% relative rise in model-judged completion against an attacker the agent never trained against.
Why this matters: Anyone planning to use an AI agent to cut the cost of a large repeated workload should test the cheaper result directly instead of assuming a strong model will build a good one.
Summary: This is a research paper that tests whether AI agents can turn their general skills into cheaper, task-specific solutions, which the authors call bottling. The agents get a whole unlabelled workload and must finish it within fixed time, compute and API budgets. Across ten models and three tasks, the authors report that strong direct performance does not reliably predict strong bottling, and 48 of 60 runs scored below the lower bound of their model's direct-performance range. They also report real savings, with Opus 5 keeping about 82% of its direct score at roughly 657 times lower reported cost on one classification task.
Via: Simon Willison (blog post) · Published: 2026-10-07
Why this matters: Anyone picking a cheap model for API work can compare their usual prompt length with Haiku's 100,000-token price step before choosing between Haiku and Luna, and note that Haiku 5.5 counts about 1.25 times as many tokens as the old Haiku.
Summary: This is a blog post by Simon Willison on Anthropic's new low-cost Haiku model, which he tested himself. He reports it costs $0.10 per million input tokens and $0.50 per million output tokens up to 100,000 tokens, matching OpenAI's GPT-6 Luna, and the price rises fivefold beyond that. His token counter shows a new tokenizer uses about 1.25 times as many tokens as the old Haiku for the same prompt. In his pelican drawing test, a low-effort run cost 0.0936 cents, while the max-effort run took over five minutes and cost 3.3826 cents. He also notes new monthly API credits for Max and Team plan users that do not roll over.
Why this matters: Mathematicians and researchers can read the Lean proofs and research details that OpenAI says it shared on GitHub to see how the results were obtained.
Summary: OpenAI says it is publishing a broad range of new mathematical results that came from an internal frontier model. The company also says it is sharing Lean proof formalizations and research details on GitHub.
Sources:
OpenAI, Simon Willison, arXiv.org
Daily Signals currently covers one topic: Artificial Intelligence. More topics may follow.