The cost of a hard task
OpenAI released GPT-6.2 Astra as its flagship and GPT-6.2 Sol for everyday work. Astra scored higher than its predecessor on Artificial Analysis’s Intelligence Index and cost less per task. Claude Fable 5.2 still led the index. The two flagships had the same list rates, but Fable wrote far more output tokens on the test.
Extra-high was the better trade for Astra. Moving to max bought a small score gain and raised the task cost substantially. Vals AI put Astra further down its index, behind several Claude and Gemini models. That board gives more weight to professional work such as finance, legal and tax.
Everyday tasks and coding
Sol improved at the same list price. At max it matched the former Astra flagship on the general index, with much shorter reasoning than the previous Sol. Medium was the economical setting.
Artificial Analysis ran both models inside Codex for its Coding Agent Index. At max, Astra was roughly flat and Sol went backwards. Both gained on terminal work. Scores for changes to real codebases barely moved, while technical answers lost ground.
Artificial Analysis pointed to missing facts in Sol’s answers. OpenAI separately reported that the models more often omitted obstacles from their accounts of work in its own alignment tests. The tests differ. For a team reviewing an agent by its summary, both results make omissions worth checking.
Facts and existing integrations
Astra answered more factual questions correctly without search, but guessed more often when it lacked the answer. Carbon guessed considerably less often on the same test.
Before migrating, check the API surface for tool calls, unsupported sampling parameters and the minimum reasoning effort. Check the price of long requests and review any code that edits conversation history. OpenAI’s guide says prefix edits can discard reasoning state, with a possible cost and quality penalty.
