OpenAI Releases GPT-6.1 Sol at One-Fifth Its Flagship’s Token Prices

OpenAI’s new pitch is flagship-like work without flagship token prices. The company released GPT-6.1 Sol at DevDay on September 29, 2026, saying it nearly matches GPT-6 Astra on coding, computer use and professional tasks. The price difference is concrete: standard input and output tokens cost one-fifth as much. The performance case, however, comes from OpenAI’s own evaluations—not independent tests of production workloads.

The upgrade arrives about a week after GPT-6 Sol. It keeps that model’s basic API rates while cutting the price of cached input, the reusable prompt material that can recur across requests. That makes this release two different propositions: an unchanged standard price for claimed better results, and a direct discount on repeated context.

The discount depends on the comparison

GPT-6.1 Sol’s API pricing is $2 per million input tokens, $10 per million output tokens and $0.10 per million cached input tokens. Tokens are the units used for this pricing; input covers material sent to the model, while output covers what it generates. Astra’s corresponding standard input and output rates are $10 and $50.

Against the previous Sol, the change is narrower. Uncached input and generated output cost the same. Cached input falls from $0.20 to $0.10 per million tokens, a 50% reduction. Developers switching between those two models therefore do not get a blanket cut to every part of the token bill.

The cached rate is relevant to agents—systems that use tools across multiple steps. A coding agent or document workflow can reuse instructions, repository context, policies or reference files between calls. Cheaper reads of that repeated material can reduce a recurring expense, while newly supplied input and generated output remain separately priced.

Three tests, different gaps and costs

OpenAI’s benchmark results give the near-Astra claim more detail than a single intelligence label. They cover software engineering, business workflows and computer use. They also use different reasoning settings—the amount of work a model puts into an answer—so the results should be read as separate comparisons, not one universal score.

• Software engineering: On DeepSWE v1.1, which tests long-running work in real codebases, OpenAI says Sol matches Astra at roughly one-fifth the cost. It also beats GPT-6 Sol’s best result by 6.4 percentage points.
• Business workflows: AutomationBench tests end-to-end tasks across 47 tools in areas including sales, support, finance and HR. OpenAI says Sol beats Claude Opus 5.5 by 2.2 percentage points at medium reasoning effort, at roughly one-third the cost.
• Computer use: On the OSWorld 2.0 offline set, OpenAI says Sol finishes within 2.1 percentage points of Astra at maximum reasoning effort. Its reported cost per task is roughly one-seventh Astra’s.

Those task-cost comparisons vary even though the standard token-price ratio stays fixed. The distinction is important: a rate card prices units of input and output, while a benchmark’s cost figure describes running the tested task. One-fifth the token price is not a promise that every completed job will cost one-fifth as much.

OpenAI says its evaluations run in its research environment or through the API and can differ from production ChatGPT. That caveat limits what the scores establish. They support comparisons under the tested conditions, rather than a guarantee that Sol and Astra will be interchangeable inside a customer’s workflow.

Accuracy and permission are separate tests

OpenAI also reports fewer factual errors on difficult prompts. At low reasoning effort, the share of responses containing an error falls from 11.4% for GPT-6 Sol to 7.7% for the new version. That is a result on the company’s challenging evaluation, not an overall error rate for ordinary use.

A different set of tests addresses how the model acts. OpenAI says Sol fails less often than its predecessor at flagging broken search tools, following explicit restrictions and avoiding unauthorized outcomes. It also says it observed no attempts to bypass an automated safety reviewer, matching GPT-6 Sol and Astra. These are company-reported test findings, not a claim that failures are impossible.

Work and Codex first, not regular Chat

GPT-6.1 Sol is available through the API as gpt-6.1-sol. OpenAI also lists access for Plus, Pro, Business, Enterprise and Edu customers in ChatGPT Work and Codex. It is not yet available in regular Chat, so availability in those work-focused products should not be mistaken for a release across every ChatGPT surface.

A Sol version of Ultrafast is still ahead. OpenAI says support is coming in the coming days, with the tier promising up to eight times faster token generation in Codex. That is a planned addition, not a capability already available with today’s standard Sol release.