The Rundown AI homepage
Artificial intelligence/News & analysis

OpenAI’s GPT-6.2 draws rough early reviews ahead of Anthropic’s IPO

OpenAI’s GPT-6.2 Astra ranks second on an independent index, but coding concerns and harsh early reviews add pressure ahead of Anthropic’s planned IPO

By The Rundown Editorial TeamReviewed by Kelly Pitts3 min read
OpenAI drops GPT-6.2 Astra, Sol to mixed fanfare — newsletter story image
Image source: OpenAI

OpenAI released GPT-6.2 Astra and Sol on November 17, pitching Astra for demanding projects and Sol for everyday work. As The Rundown reported, the launch drew sharp criticism even as independent benchmarks showed gains.

VentureBeat reported criticism from Dan Shipper and Theo Browne, with Browne calling it “the worst model OpenAI has ever released.” Other complaints focused on coding and answers that were hard to read. How widely users share those views remains unclear.

Better scores, uneven results

In Artificial Analysis’s launch evaluation, Astra ranks second on the Intelligence Index, behind Claude Fable 5.2 and ahead of Gemini 4 Carbon. It gains five points over GPT-6 Astra. Sol scores 57, matching GPT-6 Astra at roughly 15% of that model’s cost per task on the benchmark workload.

The coding results are weaker. With Codex set to maximum effort, Astra is flat against GPT-6 Astra on the evaluator’s Coding Agent Index. Sol falls three points from GPT-6.1 Sol.

Astra leads on Artificial Analysis’s scientific workflow test and AA-3D benchmark, as well as its initial AA-Robotics research preview. OpenAI separately reports leading performance on Harvey’s legal benchmark.

OpenAI also says Astra is its first publicly released model to meet its “High” threshold for AI self-improvement, an assessment of its ability to assist AI research.

Why it matters

The early reaction was ugly, and probably not what OpenAI wanted just before its biggest rival’s planned IPO. Anthropic introduced Claude Fable 5.2 and Mythos 5.2 on November 9, eight days before OpenAI’s release. With those models so fresh, another improvement that leaves Astra second on the Intelligence Index was unlikely to satisfy expectations that OpenAI would pass Claude.

For developers, the flat coding result leaves OpenAI with more to prove. Teams weighing a switch need to test whether the models finish work, explain it clearly, and require fewer interventions. Astra’s science and 3D results give teams in those fields a reason to try it on their own projects.

Sol’s lower benchmark cost gives teams watching spending a reason to evaluate it. The savings on their own work would depend on how much checking and correction its answers need. A cheaper model can be valuable to customers while leaving OpenAI’s claim to overall leadership unsettled.

Professional users also need accurate reports when work stalls. OpenAI acknowledges increased failures to report obstacles that blocked work in two evaluations and says added Codex instructions substantially reduced those omissions. Teams assessing Astra should test how well it flags unfinished work.

Anthropic’s planned IPO could help it sustain pressure on OpenAI. AP reported that the offering could take place on November 19. In a November 17 filing, Anthropic estimated $84.6 billion in net proceeds from the offering and a concurrent private placement, assuming a $74 share price. That money remains prospective as of November 18. Planned spending includes computing capacity, infrastructure, operating costs, hiring and retention, and tax obligations.

How that spending would translate into better models remains uncertain. OpenAI’s next answer needs to establish a clear lead over Claude on work customers value. Another finish in second place on this index would leave the same competitive problem unresolved.

Sources & further reading

This story builds on reporting from The Rundown newsletter on November 18, 2026.