OpenAI is positioning GPT-6.1 Sol as a cheaper model for demanding AI work. The company says it approaches GPT-6 Astra on agentic coding, computer use and professional tasks while charging one-fifth of Astra’s standard input and output token prices. The prices are concrete; the capability comparison is harder to verify.
Through OpenAI’s API, GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. The performance case rests on OpenAI’s evaluations, which do not establish that Sol will match Astra across independently tested, real-world workflows.
Developers can use the gpt-6.1-sol API model, while eligible subscribers can access it in ChatGPT Work and Codex. OpenAI says it is not yet available in Chat. For anyone choosing a model to deploy, the question is how often Astra’s additional capability, if any, justifies its higher standard token rates.
OpenAI Says Sol Improved on GPT-6 Sol
In its GPT-6.1 Sol announcement, OpenAI describes a model intended to handle complex tasks closer to Astra’s level at a lower price. It highlights agentic coding, computer use and professional work, tasks that can involve more than producing an answer. A coding agent might inspect files, make changes and run tests; a computer-use agent may need to carry out a sequence of actions while respecting instructions and permissions.
OpenAI also claims improvements over the earlier GPT-6 Sol in programming and debugging, document understanding and multistep workflows, according to TechCrunch’s account of the launch. Those categories are useful, though they are not measurements by themselves. A developer needs to know whether the model completes their tasks reliably, how much intervention it needs and what those completed tasks cost.
Agents make that assessment especially demanding. A model that solves an isolated programming problem can still struggle to navigate a repository, recover from a failed tool call or recognize that an action requires permission. OpenAI’s comparison with Astra is more consequential than a claim about writing quality, and harder to settle with a single result.
The Reported Accuracy Gain Has Limits
One of the clearest numerical comparisons in the launch reporting concerns factual errors. TechCrunch reports that, on OpenAI’s difficult-prompt evaluation at low reasoning effort, the share of GPT-6 Sol responses containing a factual error was 11.4%, compared with 7.7% for GPT-6.1 Sol. That is a 3.7-percentage-point improvement on that measure.
The result describes a particular test setting, not a general-purpose error rate. The prompts were selected for difficulty, the figure concerns responses containing factual errors, and the cited result uses low reasoning effort. Reading 7.7% as the chance that Sol will make a mistake in an ordinary conversation or coding session would be misleading.
As reported by TechCrunch, OpenAI also says GPT-6.1 Sol is better than GPT-6 Sol at recognizing broken search tools, following explicit restrictions and avoiding unauthorized outcomes during challenging tasks. For computer use, finishing the task is only part of the job if an agent ignores a boundary along the way. The reported evaluations do not guarantee that an agent will always detect tool failures or obtain permission appropriately in production.
The Astra comparison needs care, too. OpenAI says Sol approaches Astra, not that it equals Astra on every task. The reported evaluations use differing environments and effort or cost settings. A controlled test on the same tasks under the same conditions would be more informative than a headline-level comparison, and the supplied launch evidence does not include independent real-world validation of parity.
A team already using Astra can test a representative sample of its own work, tracking completion rate, correctness, unauthorized actions, retries and human review. The difficulty of failures matters as well. A small gap on routine tasks may be acceptable; a gap on the hardest cases could be the reason the team chose Astra in the first place.
The API Price Is Precise; the Total Bill Is Not
OpenAI lists these GPT-6.1 Sol API rates:
- Input: $2 per million tokens
- Cached input: $0.10 per million tokens
- Output: $10 per million tokens
The model ID is gpt-6.1-sol. OpenAI’s one-fifth-of-Astra claim applies to standard input and output token prices. The supplied pricing evidence gives Sol’s cached-input rate but does not establish an equivalent one-fifth comparison for Astra’s cache pricing.
For example, one million input tokens and one million output tokens would cost $12 on Sol at the listed rates. Applying OpenAI’s stated one-fifth ratio to those same token quantities implies $60 at Astra’s standard rates. This holds token counts fixed; it does not predict what either model would consume to complete a job.
An agentic workflow may require more turns, longer tool interactions or retries on the cheaper per-token model. It may also finish with no meaningful extra work. Caching can change the bill further when input qualifies for the discounted rate. Sol’s published price advantage remains, but a purchasing decision should compare cost per successfully completed task, not just cost per million tokens.
At $10 per million output tokens, generated text costs five times as much per token as standard input on Sol. Workflows that produce long responses or repeated intermediate outputs will have a different cost profile from those that mainly read large prompts and return short answers.
Where GPT-6.1 Sol Is Available
OpenAI says developers can access GPT-6.1 Sol through its API using gpt-6.1-sol. It also lists availability in for users. OpenAI explicitly says the model is .
Frequently Asked Questions
4 questions
1How Much Does GPT-6.1 Sol Cost Through the API?
GPT-6.1 Sol costs $2 per million standard input tokens, $0.10 per million cached input tokens and $10 per million output tokens. OpenAI says its standard input and output rates are one-fifth of GPT-6 Astra’s. That ratio does not establish a comparable discount for cached input, and an application’s total bill will depend on its token use and eligible caching.
2
Sources
- GPT-6.1 Sol announcementopenai.com
- TechCrunch’s account of the launchtechcrunch.com
