SummaryOpenAI has released the GPT-6.2 Astra (frontier) and GPT-6.2 Sol (efficient) models. Astra's price was reduced by 20%, while Sol's price remained unchanged. The models are positioned against Claude Fable 5.2 and Gemini 4 Carbon. Simon Willison's early tests yielded mixed results, with both models lagging behind competitors in a creative 3D animation task. Users have also complained about confusing questions appearing mid-task. |
Why it mattersThe aggressive price cut, particularly the 60% reduction in cached token costs ($0.40/MTok), targets the economics of long-running agentic workflows, where caching is the primary cost driver. However, this launch occurred during a competitive period where Anthropic and Google had already set a high standard. This suggests OpenAI is maintaining its frontier narrative from a defensive position rather than as a leader. |
Highlights- Astra's cached price dropped from $1 to $0.40/MTok, offering a significant cost advantage for long-running agent sessions.
- Willison's Yonkers-mech comparison shows that both 6.2 models lag behind Claude Fable 5.2 and Gemini 4 Carbon in creative coding.
- Theo Browne's pre-launch feedback highlights a user experience issue: Astra asks users ambiguous internal implementation questions mid-task, disrupting the agentic loop.
|
Critical takeThe article's own evidence undermines the 'frontier model' framing. In the only creative coding test shown, both new models lost to competitors. The most detailed user complaint involves the model getting stuck on a simple file naming question instead of solving the task. A 20% price cut for a model that does not clearly lead in benchmarks resembles a defensive repricing rather than a breakthrough. |
This critique was written by the local AI model
Qwen3.8-27B. I chose the story; I do not review each critique before it is published.