This Week in AI, 13 Items — Three New Models Land at $2/$10, and the Contest Moves to Tokens per Task and Who Can Use Them

For work you run through the API, you can try near-Astra results at a fifth of the unit price. But OpenAI chose the benchmarks and reasoning settings, and it states that Astra still scores highest on scientific research at 68.1%. The split of Astra for the hardest reasoning and Sol for complex coding and computer use also appears in the 2 October model guide. If you try it in ChatGPT, pick it in "Work", not in Chat.

If you run Sonnet with thinking off, check before migrating. Anthropic’s migration guide says you must switch to the new `between_tools` setting before moving to Sonnet 5.5. On cyber, it is the first Sonnet to ship with Opus 5.5-style safeguards, so higher-risk cybersecurity tasks visibly fall back to Sonnet 5 (routine bug fixing is unaffected). It is also the first Sonnet with classifiers that block reasoning extraction, and its thinking can no longer be decoupled from the account that created it, which affects workflows that move conversations between accounts, including switching accounts mid-session in Claude Code.

You probably cannot use this model yet even if you apply. Access is limited to trusted cyber defenders and no developer date is given, so $2/$10 is not a price you can buy today but a limited-time preview price, doubling afterwards (the end date of the introductory period is not stated). The scores are on benchmarks Google chose, with no direct comparison to other vendors’ latest models. Putting frontier capability behind a vetted tier and then widening in phases is the same pattern as the three vendors in issue 3 (6 September).

The biggest caution is that costs cannot be forecast. The first dot is free, and the post only describes a future where you add more dots or scale their speed and monthly workload, so no price is published. On privacy, content from Business, Enterprise and Edu workspaces is not used to improve models by default, and personal plans can choose. A dot’s notes to itself and its proactive research are not trained on directly, but may inform an eligible conversation or task depending on settings. Sensitive actions such as changing a password always stay with the person.

Where OpenAI’s dots are one persistent agent per person, this is "one agent shared by a team". The reason it is here is that in the same week vendors put always-on, shared agents at the centre of their products. The figures are the company’s own case studies, with no evidence of reproducibility in other environments.

Speed is an add-on, its pricing is in the "Ultrafast guide", and this issue did not confirm it. The monthly price of Pro 500 is also absent from the text I retrieved. OpenAI and NVIDIA say speed matters most when an agent loops through "write, call a tool, check" many times. Before you consider it, check the price in the primary sources (the Ultrafast guide and the plan pricing page).

Organisations that run on AWS now have a way to use OpenAI agents without the data leaving AWS. But the AWS side is in preview, and this issue did not confirm production terms such as pricing or availability guarantees. The Decisions API is only said to be "broadly released in the coming days", so I will check whether it shipped in the next issue.

The change is that it keeps running with your laptop closed. On the other hand, a security scan run in the cloud means handing the reading of whole repositories and the preparing of fixes to OpenAI’s cloud, so decide the scope of repositories and permissions before you turn it on. Security Cloud is for Pro and above; Plus is not included.

If you restrict permissions with rules, it is worth moving to 2.1.289, after Mods arrived. The two fixes landed in 2.1.289, so earlier versions could evidently be affected (under the conditions in the fix notes: managed machines and organization-managed MCP servers). `allowedProviders` is a way for administrators to narrow where models are reached from.

The money in an OpenAI contract can now move as a budget for neighbouring tools. For the software companies, it adds a path for OpenAI’s customers to buy their products. For buyers, the conditions for applying a commitment (which contracts qualify, any cap, how it is settled) are not stated in the announcement, so do not read it as freely usable until you confirm with your account team.

This puts "the number of people who can deploy it" at the centre of competition, not model performance. It targets customers’ and partners’ engineers and goes through the Anthropic account team, so it is not a scheme an individual can apply to. How the credential is valued in the market will not be known until the first graduates in early 2027.

Systems that make reasoning portable or replayable become an attack surface (OpenAI itself writes that such systems may face related risks). The same week, Anthropic shipped Sonnet 5.5 with classifiers against reasoning extraction. Using a vendor model’s outputs or thinking to train or reproduce another model is the kind of use now being shut down (OpenAI says it violates its terms). But the attribution to Moonshot AI is OpenAI’s assessment, and this issue could not confirm that company’s response from a primary source.

This is a proposal from OpenAI, not a standard. It has no legal force and no guarantee that others will adopt it. Even so, alongside the misalignment reports in issue 5 and Anthropic’s embedded outside evaluator, the direction of making training-time safety verifiable from outside is lining up across vendors’ official documents, which is worth noting.