OpenAI cuts Astra’s API prices by 20%
OpenAI says the upgrade can stretch subscriber allowances, while its fastest mode comes with a usage tradeoff.
By Saeed Ezzati8 min read

The audio edition
Listen to this newsletter
0:005:02
Read transcript
OpenAI released GPT-6.2 Astra and Sol on November 17, and cut Astra’s standard input and output prices by 20 percent. Both models are available through the developer API, ChatGPT Work and Codex. Sol’s price is unchanged from GPT-6.1 Sol, and OpenAI continues to recommend Astra for the hardest projects. The price cut is only part of the pitch. OpenAI says both models use fewer output tokens on demanding tasks, which could let subscribers complete more work before hitting usage limits. That is a separate potential saving from the lower price per token. In Codex, OpenAI says Astra is better at modifying and testing existing software; early testers reported cleaner code, more focused tests and fewer revision rounds. Those are user reports, not a measured estimate of time saved. Sol remains the cheaper choice for everyday coding and writing. OpenAI also points to a result on Cognition’s FrontierCode 2.0: Sol finished within half a point of Astra at about one-fifth the cost, with both models using maximum reasoning effort. The test evaluates changes to real codebases. It does not show that Sol and Astra are interchangeable across other kinds of work. There’s a separate speed-versus-usage tradeoff in Ultrafast mode. OpenAI says it can generate tokens at up to 4.5 times the standard rate, while consuming more of a user’s allowance. That is a token-generation rate, not a promise that an entire project finishes 4.5 times faster. The $200-a-month Pro plan gets Ultrafast for Sol; the $500 plan gets it for both models. Eligible Enterprise and Edu workspaces can enable it for both through an administrator. So the practical choice is not simply cheaper or faster: it’s which model and speed fit the task, and how much usage you’re willing to spend. That same question of control shows up in Anthropic’s new linked-billing beta for supported Claude tools. Customers can authorize a third-party platform to charge their existing Anthropic organization without giving it an API key. They can use an existing spending commitment or pay as they go, set a monthly cap for each linked platform, review its spending in Console, and disconnect it at any time. The important caveat: the requests run in the tool provider’s organization, not the customer’s. That means the customer’s IP restrictions, encryption-key settings and data-retention rules do not carry over. A billing link avoids handing over a credential; it does not extend the customer’s security controls into the outside tool. An Owner, Admin or Billing user approves the connection for the organization, and launch supporters include Linear, Ramp and Warp. And with Gemini, Google is moving from answering questions to taking action. It began rolling out free tools for appointments, bill analysis and shopping across the U.S., Japan, South Korea, Canada, Australia and India, with advanced features starting in the U.S. Google says Gemini connects with more than 60 popular apps. The examples include calling dealerships to schedule a service appointment, reviewing utility bills, and comparing products before checkout. Google executive Josh Woodward described these as his own experiences, not independent accuracy tests. Some shopping tasks still hand the final step back to the user when a website requires it. Users can set spending caps and require confirmation for critical actions, such as clicking Buy; routine actions can be pre-approved. Google says each task runs in a hardware-isolated virtual machine, with defenses against unsafe links and prompt injection. Users can interrupt tasks, disconnect apps and delete chat history. The rollout’s reach is broad, but availability and how much a task can actually complete still depend on location, account access and the website. The same gap between doing a task and understanding its limits is central to a new account from theoretical physics. Suvrat Raju, co-creator of TheoPhysEval, told Understanding AI he has stopped taking new PhD students while he assesses their prospects. The benchmark’s authors report GPT-6.2 Astra, running in Codex with literature-search and calculation tools, solved 49 of 50 unpublished problems at an average cost below three dollars each. Astra also identified subtle errors in human solutions to apparent misses, prompting a grading correction. That is a striking result on checkable research calculations, not proof that an agent can independently make a major discovery. The paper says removing web search and background papers changed accuracy only slightly on average, suggesting the model drew substantially on learned literature. Physicists interviewed for the report also raise an apprenticeship concern: doing calculations can help students build the intuition needed to choose worthwhile questions. Across these stories, the useful thing to watch is where the systems’ boundaries actually lie—usage costs, permissions, website handoffs, and the human judgment that still has to guide the work.



