OpenAI Releases GPT-6.1 Sol - Input and Output Prices Unchanged, Only Cached Input Moved

OpenAI released GPT-6.1 Sol on September 29, 2026, timed to its developer event DevDay 2026. The company frames it as an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra’s intelligence on agentic coding, computer use, and professional work, at one-fifth of Astra’s standard input and output token prices¹.

That “one-fifth” is a comparison with Astra, not with the previous generation. Standard, short-context pricing is $2 per million input tokens and $10 per million output tokens¹, exactly the same as GPT-6 Sol released a week earlier³. What actually moved is cached input, cut from $0.20 to $0.10, which OpenAI describes as 50% below GPT-6 Sol’s cached input pricing.

Which Row in the Price Table Changed

Placing the two models’ published prices side by side, exactly one row differs (all figures per million tokens, standard processing, short context)²³.

Item	GPT-6 Sol	GPT-6.1 Sol
Input	$2	$2
Cached input	$0.2	$0.1
Cache writes	$2.5	$2.5
Output	$10	$10

As a ratio, cached input fell from 10% to 5% of the uncached input rate. The benefit of this generation change therefore reaches only workloads that resend the same prefix repeatedly. For one-off calls that do not reuse the front of a prompt, the bill does not change by a single token.

If you are running agents, there is room for this to matter. Setups that stack a long system prompt and tool definitions at the front of every request are usually designed so that most input tokens hit the cache, and the rate on that portion is now halved. How much is actually hitting is visible in the prompt caching dashboard released in August. Migrating while hit rates stay low delivers almost none of the reduction.

Note that prompts above 272K input tokens are billed at 2x for input and cache rates and 1.5x for output, for both models². If long contexts are routine for you, the halved cached input sits inside that multiplier too.

none Is No Longer Available

The migration snag may be the supported reasoning efforts rather than the price. GPT-6 Sol accepted none for reasoning.effort; GPT-6.1 Sol accepts neither none nor minimal. The available settings are low, medium (default), high, xhigh, and max²³.

Tool calling changed as well. GPT-6 Sol supported function calling through Chat Completions as long as reasoning_effort was set to none, while GPT-6.1 Sol requires the Responses API for tool calling and supports Chat Completions only without tools. Swapping gpt-6-sol for gpt-6.1-sol and nothing else will break calls built around none.

The rest of the specifications are largely unchanged: a 1,050,000-token context window, a maximum of 922,000 input tokens, and 128,000 max output tokens are identical across the two. Only the knowledge cutoff advanced, by ten days, from April 20 to April 30, 2026²³. There is no reason to switch for context window size.

The Benchmarks Are In-House, and They Do Not Replace Astra

Every performance claim is a number OpenAI published itself. The company notes that competitor scores were taken from publicly available reports, and that its own GPT evaluations ran in its research environment or via its API, which may produce slightly different output from production ChatGPT because of differences in system prompts, available tools, and efforts¹.

On that footing, here is what the company reports. On DeepSWE v1.1, which evaluates software engineering in real codebases, GPT-6.1 Sol matches Astra at roughly one-fifth of the cost and exceeds GPT-6 Sol’s best score by 6.4 percentage points. On AutomationBench, which measures multi-step business workflows, it scores 2.2 points above Opus 5.5 at medium reasoning effort at roughly a third of the cost. On the offline set of OSWorld 2.0, for computer use, it beats GPT-6 Sol by seven points at maximum reasoning effort and comes within 2.1 points of Astra¹.

The company is explicit, however, that the hardest scientific research tasks should go to Astra. On Terminal-Bench Science 0.1, which covers scientific workflows, Astra holds the highest score among the models tested, and while GPT-6.1 Sol costs substantially less per task, its score does not reach Astra’s¹. Summarizing this as “near parity at one-fifth the price” takes only one side of the company’s own account.

There are figures on factual errors too. The largest improvement comes at low reasoning effort, where the share of responses containing a factual error fell from 11.4% for GPT-6 Sol to 7.7%¹. That evaluation, though, was assembled from de-identified conversations in which users had flagged an earlier model’s error, and OpenAI states that these are deliberately difficult prompts that do not represent typical usage. The number cannot be read as a general error rate.

On alignment evaluations, the company reports improvement over GPT-6 Sol that brings the model closer to Astra, with lower failure rates on disclosing a broken search tool, respecting explicit restrictions, and avoiding unauthorized outcomes during agentic tasks¹. Here too the tasks are built to elicit failures and are not presented as measuring failure rates in normal use.

Where It Is Available, and Where It Turns On by Default

In ChatGPT, it is available to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. It is not yet in Chat¹. Per the documentation, Enterprise and Edu keep it off by default until an administrator enables it, and Free and Go are not included at launch. Standard and Fast modes are available, with Ultrafast support coming later⁵.

In the API, the model ID is gpt-6.1-sol. It supports US and EU data residency, though Fast mode is unavailable with EU data residency².

GitHub Copilot began rolling it out the same day, September 29, to Copilot Pro+, Max, Business, and Enterprise users⁴. Billing is at provider list pricing under usage-based billing. GitHub writes that in early testing it reliably completed tasks using noticeably fewer tokens and steps than earlier models in the GPT-6 and GPT-5.6 families. On enablement, the changelog states that new models turn on automatically unless an administrator has switched off the global default or explicitly disabled the model. That default is worth reading alongside the default policy for new features announced on September 24.

Judging a migration from the price table alone invites reading unchanged numbers as a discount. The bill only moves for workloads with high cache hit rates; everywhere else, what remains is the change in performance and the loss of none. Claude Sonnet 5.5, which Anthropic released the day before, likewise held its per-token prices steady while positioning the model as cheaper per task because it needs fewer tokens. The window in which unit prices alone can decide a model is narrowing. The real difference only shows up once you measure cache behavior and token consumption on your own workload.

Sources

1. Introducing GPT-6.1 Sol - OpenAI official announcement (September 29, 2026)
2. GPT-6.1 Sol - OpenAI documentation (pricing and specifications; accessed September 30, 2026)
3. GPT-6 Sol - OpenAI documentation (pricing and specifications for comparison; accessed September 30, 2026)
4. GPT-6.1 Sol in GitHub Copilot - GitHub changelog (September 29, 2026)
5. Models - OpenAI documentation (availability in ChatGPT and Codex; accessed September 30, 2026)