GPT-6.1 Sol Is One-Fifth of Astra's Price, but Only the Cached Rate Is New
A.I. / news
GPT-6.1 Sol Is One-Fifth of Astra's Price, but Only the Cached Rate Is New
OpenAI's model page lists the same $2 and $10 per million tokens Sol has carried since July. The change is a cached-input rate of $0.10.

OpenAI released GPT-6.1 Sol on Tuesday, September 29, pricing it at $2 per million input tokens and $10 per million output tokens, a fifth of what GPT-6 Astra costs. The one rate that actually moved is cached input, which fell to $0.10 per million tokens.
The figures come from OpenAI's API model page and from RuntimeWire's launch summary. OpenAI's launch post returned an HTTP 403 error to an automated request, so the benchmark figures below are second-hand and are OpenAI's own.
What the model page lists
The API page describes GPT-6.1 Sol as offering "near-Astra performance at a lower cost for complex coding, computer use, and professional work." It takes text and image input and returns text only.
The context window is 1.05 million tokens, with 922,000 of those available for input and 128,000 for output. Knowledge runs through April 30, 2026. Reasoning effort can be set from low to max, and minimal effort is not offered.
Above 272,000 input tokens the price doubles for input and rises 50 percent for output, according to the same page. That puts a long prompt at $4 in and $15 out per million tokens.
| Rate, per million tokens | GPT-6.1 Sol | GPT-6 Astra |
|---|---|---|
| Input | $2 | $10 |
| Output | $10 | $50 |
| Cached input | $0.10 | Not listed |
- GPT-6.1 Sol10 USD
- GPT-6 Astra50 USD
Source: RuntimeWire launch summary of OpenAI pricing, accessed 2026-09-29
Why the headline overstates the cut
FourWeekMBA argued that the "one-fifth of Astra" framing is accurate but misleading. Its analysis said the $2 and $10 standard rates are identical across GPT-5.6 Sol, GPT-6 Sol and GPT-6.1 Sol since July, and that the cached read was cut 50 percent to $0.10.
That is a single-outlet reading, and OpenAI has not published a price history on the model page. The comparison with Astra is real. The comparison with the previous Sol is not a cut, on FourWeekMBA's numbers.
The cached rate matters for agents, which resend the same long prefix on every step. OpenAI's page lists cache writes at 1.25 times the standard input rate, so a workload that rarely rereads its prefix pays more, not less.

OpenAI's benchmark claims
RuntimeWire listed OpenAI-reported results, all from internal evaluations rather than independent testing. On DeepSWE v1.1 the model matches Astra at roughly a fifth of the cost. On OSWorld 2.0 in its offline setting it lands within 2.1 points of Astra at about a seventh of the cost.
OpenAI also reported that the factuality error rate at low reasoning effort fell from 11.4 percent on GPT-6 Sol to 7.7 percent. RuntimeWire said Sam Altman announced the model on X, aimed at developers building agents.
FourWeekMBA pointed to a result OpenAI did not lead with. On a reliability test that flags broken tools, Astra failed 1.5 percent of the time against 2.1 percent for Sol.
Access and limits
The model is in the API as gpt-6.1-sol on the Chat Completions and Responses endpoints, plus Batch. RuntimeWire said it is also in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users.
Rate limits on the model page run from 500 requests and 500,000 tokens a minute at the entry tier to 15,000 requests and 40 million tokens a minute at the top. Fine-tuning and predicted outputs are not supported.
OpenAI has not said whether Sol's cached rate will be matched on Astra. The page is live now, so developers can check their own bills against it.
Astra is the model the UK AI Security Institute ran through 500 simulated cyber runs with its classifiers off, as reported on September 28. OpenAI is also one of two labs named in Cal Newport's call for a congressional investigation.
Sources
More in A.I.
- 01Runta's 27,280-Run Harness Eval Finds the Model Explains 92.8% of Agent CostFrontierHarness v1.1 ran 22 coding-agent harnesses against five open models, and Codex CLI went from first place to sixteenth when the model changed.
- 02Qwen3.8-27B Has 6.9 Million Downloads on Hugging Face. Artificial Analysis Rates It Slow and VerboseAlibaba's Apache 2.0 dense model beats its size class on an independent index of 20, and takes 38.8 minutes a task against a 12.6-minute median.
- 03Claude Code Team Publishes Its Verification Rules: 39 Checks, a Three-Round Limit and Observe ModeA Nov. 16 post says about 90% of the team's code is written and reviewed by Claude, and lays out the gates that keep it shippable.
- 04Inkling 2 Small Weights Land on Hugging Face Under Apache 2.0, and the Card Asks for 500GB of VRAMThinking Machines opened the weights a week after launching on Tinker. The safety section still says release is pending.