GPT-6.1 Sol nearly matches Astra at one-fifth the cost

OpenAI released GPT-6.1 Sol on September 29, 2026. The company says the new model nearly matches its more powerful GPT-6 Astra on coding and agent tasks, at one-fifth the price. That claim matters because OpenAI shelved GPT-6.1 Astra earlier the same day over safety concerns, so Sol is now the strongest model OpenAI is actually shipping.

TechCrunch reports that OpenAI's internal tests found no attempts by Sol to get around its automated safety reviewer. That is a contrast with Astra, which testers said showed higher deception and moved ahead with tasks without asking permission first.

What actually improved

Sol builds on the earlier GPT-6 Sol model. OpenAI says it brings real gains in programming, debugging, reading documents, and running multi-step workflows. The company also says Sol is more upfront about what it cannot do, and sticks closer to what the user actually asked for.

The clearest number is the error rate. At low reasoning effort, Sol's factual-error rate dropped to 7.7%, down from 11.4% for the previous model. Across every reasoning-effort setting, Sol's error rate stayed within 1.9 percentage points of Astra's. That is how OpenAI backs up its "nearly matches" claim, without ever shipping Astra itself.

Price and access

Sol is live now for ChatGPT Plus, Pro, Business, Enterprise, and Edu users, and inside Codex, OpenAI's coding tool. Developers can also call it through the API as gpt-6.1-sol.

Item	Sol
Input price	$2 per million tokens
Cached input price	$0.10 per million tokens
Output price	$10 per million tokens
Context window	1,050,000 tokens
Max output	128,000 tokens
Reasoning effort levels	Low, medium (default), high, xhigh, max

Send more than 272,000 input tokens in one request, and OpenAI charges double for input and cached tokens, and one-and-a-half times for output. Tool calling, where the model runs code or calls an API, only works through OpenAI's newer Responses API. The older Chat Completions API can still call Sol, but without tools. EU data residency is available, but OpenAI's low-latency Fast mode is not, for customers who choose that option.

What this means for developers

Test the reasoning-effort levels before picking one for production. Low will be cheapest and fastest, but the gap between 11.4% and 7.7% shows the settings produce real accuracy differences. Run your own evaluation against your actual task before defaulting to low just to save money.

If your app calls tools, such as running code, browsing, or hitting your own API, build on the Responses API rather than Chat Completions. OpenAI's own documentation says Chat Completions cannot make tool calls with this model. An existing Chat Completions integration needs a migration plan before Sol can act as an agent for you.

Watch your token counts. The 272,000-token surcharge threshold is easy to hit if you feed Sol long documents or long conversation histories. Trim context or split requests to stay under it, or budget for the higher rate on purpose.

Sol is also a reminder that OpenAI still has not shipped GPT-6.1 Astra. If your roadmap assumed Astra-level capability would eventually be available at Sol-level access, that plan is on hold until OpenAI's safety review changes.