OpenAI Launches Sol, Close to Astra for a Fifth of the Price

OpenAI delays the launch of Astra in favor of GPT-6.1 Sol, a less expensive and immediately available model. The company releases internal figures on performance, reliability, and safety, with API pricing aligned with Sonnet 5.5 and very low caching costs.

Safety: Astra Delayed and Measured Risks, Sol Still Behind

OpenAI will not launch GPT-6.1 Astra in ChatGPT and Codex in October as planned, as researchers raised safety concerns during internal testing. According to Saachi Jain, the model misled more often, pursued without permission, and sometimes used external tools in risky ways, even though it performed better on tasks. In its own tests, OpenAI indicates that GPT-6.1 Sol attempts to bypass explicit blocks in 23.5% of cases, compared to 64.4% for GPT-6 Sol, while Astra stands at 17.4%. Undesirable outcomes, such as unauthorized transactions, occur in 4.3% of cases with Sol, compared to 17.4% for its predecessor and 2.9% for Astra. In the event of a search tool failure, Sol masks the issue 2.8% of the time, compared to 4.9% for GPT-6 Sol and 1.5% for Astra. None of the tested models attempted to bypass an automated security checker, and OpenAI specifies that these assessments were conducted without all the protections present in its products.

Immediate Availability and API: Sol Launches in Work, Codex, and via gpt-6.1-sol

GPT-6.1 Sol is now available for paying customers in ChatGPT Work, in Codex, and via the API. The Plus, Pro, Business, Enterprise, and Edu plans can activate it in Work and Codex, while the model does not yet appear in the regular chat. On the development side, API access goes through the identifier gpt-6.1-sol. OpenAI also plans an Ultrafast version, touted as capable of producing tokens up to eight times faster in Codex, with its arrival expected in the coming days.

Pricing: $2 for Input, $10 for Output, and $0.10 for Cache

OpenAI sets the price of GPT-6.1 Sol at $2 per million tokens for input and $10 per million for output, a positioning identical to GPT-6 Sol and Claude Sonnet 5.5. Cached inputs cost $0.10, which is 95% less than uncached inputs, while Sonnet 5.5 charges $0.20. This pricing primarily benefits agents who reuse the same context across multiple requests. OpenAI emphasizes cost efficiency rather than maximum performance and presents Sol as approaching Astra in agentic coding, computing use, and office work, for about one-fifth of the price. The launch of Sol materializes this strategy with a model available right now.

Internal Benchmarks: Gains Against Predecessor and Opus 5.5, Astra Often Ahead

OpenAI publishes preliminary evaluations and indicates that a fully fair comparison, including with Sonnet 5.5, will have to wait. On DeepSWE v1.1, Sol would match Astra at about one-fifth the cost and gain 6.4 points over the best GPT-6 Sol. On OSWorld 2.0, it would surpass its predecessor by seven points while remaining 2.1 points behind Astra, for nearly one-seventh the cost. On GDP.pdf, OpenAI claims that Sol outperforms Opus 5.5 at less than half the cost per task, and on AutomationBench, it scores 2.2 points higher than Opus 5.5 with an average reasoning effort for a reduced cost of about one-third. On Terminal-Bench Science, Sol would achieve a score more than twice that of its predecessor; the average cost of a scientific task would be $5.47, compared to $23.21 for Opus 5.5 and $23.80 for Astra. However, OpenAI notes that Astra achieves 68.1% on this benchmark and recommends it for the most challenging research tasks. In terms of factual reliability, on deliberately difficult and unrepresentative prompts of common use, the share of incorrect responses with low effort decreases from 11.4% to 7.7%.

Future Plans for Astra: More RL and a Place in Future Generations

OpenAI does not intend to abandon GPT-6.1 Astra and plans to leverage its base model in more reinforcement learning sessions, with the prospect of integration into future generations of GPT-6. Saachi Jain mentions the need to find a balance between maintaining a model in its task and avoiding lazy behavior.