OpenAI releases GPT-6.2 Astra and Sol with lower pricing and mixed benchmarks
OpenAI launched GPT-6.2 Astra and Sol on November 17, 2026, offering reduced API costs but showing uneven performance in complex coding tasks compared to rival models.
OpenAI introduced two new frontier models, GPT-6.2 Astra and GPT-6.2 Sol, on November 17, 2026. These updates bring significant price reductions for developers while aiming to balance high-end capability with everyday efficiency. Early tests reveal a complex trade-off between cost savings and raw performance against competing systems.
What happened
The release splits the model lineup into two distinct tiers. GPT-6.2 Astra targets difficult, resource-intensive projects, while GPT-6.2 Sol serves as a versatile option for routine development work. Both are immediately available via the API under the identifiers gpt-6.2-astra and gpt-6.2-sol. Simon Willison updated his LLM tooling to version 0.38 to support these new endpoints, signaling rapid adoption in the developer community.
Pricing represents a major shift in the market dynamics. GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens. This is an 80% price point relative to its predecessor, GPT-6 Astra, which charged $10/$50. The cached input price drops even further to $0.40 per million tokens, down from $1 per million. This structure matches the pricing of Claude Fable 5.2, creating direct competition at this specific performance tier. GPT-6.2 Sol maintains the existing $2/$10 price point, which remains half the cost of the older GPT-5.6 Sol.
Performance evaluations show mixed results. In a comparative test involving a complex Three.js animation request, neither GPT-6.2 Astra nor Sol performed as well as Claude Fable 5.2 or Gemini 4 Carbon. The task required morphing a map of Yonkers into a mech character using OpenStreetMap data. While the models generated functional code, the visual execution lacked the fidelity of their competitors. Additionally, user feedback highlights usability issues with Astra’s agentic behavior during long-running tasks.
How it works
These models operate within the standard transformer-based architecture but utilize optimized inference pathways to achieve the new price points. The significant discount on cached tokens suggests improved state management for long-context sessions. This allows agents to retain information across extended interactions without incurring the full cost of re-processing previous inputs. This mechanism is critical for agentic workflows where the model must maintain context over hours of operation.
The distinction between Astra and Sol lies in their compute allocation and reasoning depth. Astra employs higher effort levels, such as the "Ultra" setting used in testing, to tackle multi-step problems. It can spawn and manage subagents to handle parallel tasks. However, this complexity introduces new failure modes. The model may surface internal implementation details to the user when blocked, asking questions about file paths or system states that are irrelevant to the human operator. This behavior indicates a gap in the model’s ability to filter its internal reasoning process before presenting queries to the user.
Key details
- GPT-6.2 Astra is priced at $8/MTok input and $40/MTok output, matching Claude Fable 5.2.
- Cached input tokens for Astra cost $0.40 per million, a significant reduction from the previous $1 rate.
- GPT-6.2 Sol retains the $2/MTok input and $10/MTok output pricing of its predecessors.
- In a Three.js animation benchmark, both new models underperformed compared to Claude Fable 5.2 and Gemini 4 Carbon.
- Astra can manage multiple subagents but may confuse users by asking obscure technical questions when blocked.
- Support for the new models was added to LLM version 0.38 shortly after release.
Why it matters
For software engineers and technical founders, the pricing adjustment makes high-capability models more accessible for production use. The reduction in cached token costs specifically benefits applications that rely on long-running agentic sessions. Developers can now afford to keep context windows open for extended periods, enabling more complex autonomous workflows without prohibitive expenses. This aligns the cost structure of OpenAI’s top-tier model with its primary competitor, forcing a focus on performance and reliability rather than just price.
However, the performance gaps highlight that cost efficiency does not yet equal superiority in all domains. The inability to match the visual and logical coherence of Claude Fable 5.2 or Gemini 4 Carbon in creative coding tasks suggests that teams should maintain a multi-model strategy. Relying solely on the newest OpenAI models may lead to inferior outputs in specialized areas like generative UI or complex spatial reasoning. Furthermore, the reported confusion in Astra’s communication style requires careful prompt engineering and oversight to prevent workflow stalls.
What you can do
- Update your API clients to use
gpt-6.2-astraandgpt-6.2-solendpoints immediately to access lower rates. - Implement caching strategies for input tokens to leverage the $0.40 per million token price for Astra.
- Benchmark your specific use cases against Claude Fable 5.2 and Gemini 4 Carbon before fully migrating workloads.
- Monitor agentic logs for obscure internal questions from Astra and add filtering logic to your application layer.
- Use GPT-6.2 Sol for routine coding tasks where the $2/$10 price point offers sufficient performance.
- Test the "Ultra" effort level only for tasks requiring deep reasoning, as it incurs higher latency and potential confusion.



