OpenAI releases GPT-6.2 Astra and Sol; several early testers criticize Astra

For a technology chief running model procurement at a bank

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, and several power users who tested them before launch, including Matt Shumer and Theo Browne, criticized the models, VentureBeat reported.1

Shumer said he gave Astra work as a sub-agent and the Claude models supervising it redid the work, according to VentureBeat's account of his posts.1 Browne said Astra spent seven days on his TypeScript compiler rewrite in Rust without landing an improvement and produced a version that ran 56% slower after focusing on reducing its size.1

Astra now costs $8 per million uncached input tokens and $40 per million output tokens, while GPT-6.2 Sol keeps its predecessor 6.1 Sol's $2 input and $10 output rates, VentureBeat reported; it also reported that Astra came with a 20% price cut over the API, but did not say what the earlier price was.1

Both are proprietary, hosted models, available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users and through the OpenAI API as gpt-6.2-astra and gpt-6.2-sol.1

VentureBeat reported that Anthropic launched Claude Fable 5.2 and Mythos 5.2 eight days before OpenAI's release, with Fable priced at $8 and $40 per million standard input and output tokens, the same as Astra, though it said Anthropic charges half as much as OpenAI for cache reads on those flagships.1

The feedback is mixed: Browne said Astra caught code-review issues other models missed and still preferred Claude Opus 5.5 for development, Ethan Mollick called it "very uneven & weird," and Ben Hylak and Ashe Magalhaes were more positive about its visual and video-editing work.1

Why it matters

Token prices do not show the cost of a finished job. If a model stops early, needs rework or is redone by another model, the effective cost rises, so banks and fintechs should test these models on their own workloads before switching. The reports come from a few testers on their own tasks and are not a general verdict.

From the brief

Our take. Astra now costs the same as Anthropic's Fable after a 20% cut, while one tester watched it spend seven days producing code that ran 56% slower; a cheaper model you cannot steer saves nothing until your own tests agree.

Action. Model risk: Pin production API calls to current model versions and rerun bank workload evaluations on gpt-6.2-astra before any switch.

What our records say

ChatGPT Enterprise is OpenAI's workplace version of ChatGPT, launched in 2023, and it sells banks seats and API access.1 In 2026 OpenAI added on-site engineering deals, a financial-services edition built with investment-bank design partners, and a separately funded deployment company.1 Strategy and product: OpenAI sells banks ChatGPT seats for employees and API access to its models.1

  1. ChatGPT Enterprise: AI programme summary (part 1)

From our records

Sources

  1. ReportOpenAI launches GPT-6.2 Astra and Sol to negative reception from power users Venturebeat