OpenAI launches GPT-6.2 Astra and Sol, touting record benchmarks at sharply lower costs
by Paul

OpenAI launches GPT-6.2 Astra and Sol, touting record benchmarks at sharply lower costs

OpenAI has announced the release of two new models: GPT-6.2 Astra, a frontier model pitched at the hardest professional work, and GPT-6.2 Sol, an efficient everyday companion. The company says both models achieve better results on demanding tasks with fewer output tokens, framing the launch as frontier performance “at costs that make sense at scale”. Astra set new highs on multistep agent benchmarks — scoring 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost on Agents' Last Exam — while OpenAI called it “the best model in the world for scientists”, citing new records on science-focused evaluations.

The announcement leans heavily on professional-work benchmarks: Astra more than doubled Claude Fable 5.2's pass rate on a legal-drafting test from Harvey at about one-sixth the cost, led OpenAI's internal GDPval v2 on long-horizon workflows, and topped the AA-3D design leaderboard with 1,318 Elo. It also demonstrated finer-grained computer use, from modeling mechanical assemblies in CAD to building a Python solver simulating car chassis dynamics in Blender. Alongside these gains, OpenAI flagged a significant jump in cybersecurity capabilities: both models now meet the critical threshold under the company's Preparedness Framework, prompting tighter safeguards on higher-risk assistance, and Astra became OpenAI's first externally released model to hit the high threshold for AI self-improvement. To manage that risk, the company built Control Room, a system that tracks research workloads and can halt model output when intervention is needed, along with expanded monitoring that can automatically flag unauthorized behavior.

On pricing, Astra's API rates drop 20% versus its predecessor to $8 per million input tokens, $40 for output, and $0.40 for cached input, a 60% discount, while Sol carries forward GPT-6.1 Sol's rates of $2/$10/$0.10. Both models are available today to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, with Ultrafast mode extended for the first time to Pro 200 subscribers. OpenAI also disclosed a candid alignment note: the new models were slightly more likely to omit mentioning obstacles that blocked their work, such as a missing file, an issue the company says it has substantially reduced through instructions in Codex and is continuing to investigate.

Paul
Written by Paul
Follow
Add as a preferred source on Google
ChatGPT iconChatGPT
  795
  • ...

ChatGPT, developed by OpenAI and launched in November 2022, is a generative AI chatbot powered by large language models (LLMs). It excels in understanding and generating human-like text, images, and speech. With a user rating of 4.1, it features Dark Mode, a Chat Bot interface, and is web-based. It is often compared with alternatives like Bard, ChatSonic, and Claude.

Other apps mentioned in this article

Comments

0

OpenAI will use token based subscriptions next month, so a $20 subscription will give $20 of tokens for the month (maybe up to $50 if OpenAI is generous) so such a model won't be able to complete any task. Not that OpenAI cares about theses plans anyway, it is focusing now on business plans, where the money is, instead of million low usage users that consume Altman's precious RAM.