guide

GPT-6.1 Sol pricing: near Astra, lower task costs

Sol is one Index point behind Astra at a much lower measured task cost. Ordinary API prices stay put, cache reads halve, and migration has a tool-calling catch.

EiliyaSeptember 30, 20267 min read

GPT-6.1 Sol sits one point below GPT-6 Astra on Artificial Analysis's Index, at $0.72 per evaluation task to Astra's $3.26. The September 29 release also keeps GPT-6 Sol's ordinary API prices while halving its cached-input rate.

You get a newer model without paying more for fresh input or output. How much you save in your own agent loop needs testing: token prices, benchmark task costs and ChatGPT subscription allowances measure different things.

GPT-6.1 Sol pricing, including cache writes

The API model ID is gpt-6.1-sol. These are Standard processing prices in US dollars per million tokens, for requests with no more than 272,000 input tokens:

Token category	GPT-6.1 Sol	GPT-6 Sol	GPT-6 Astra
Uncached input	$2	$2	$10
Cached input reads	$0.10	$0.20	$1
Cache writes	$2.50	$2.50	$12.50
Output	$10	$10	$50

The full OpenAI pricing table includes cache writes. As the prompt-caching guide explains, you pay the $2.50 write rate instead of the $2 ordinary input rate for those tokens. If your agent keeps creating new prefixes, include those writes in your estimate.

With caching disabled, a hypothetical request using 20,000 input tokens and 5,000 billable output tokens costs $0.09 on Sol: $0.04 for input plus $0.05 for output. The same counts cost $0.45 on Astra. This calculation uses the rates above and excludes tool charges and regional premiums; it isn't an observed task bill.

Output can include invisible reasoning tokens, as the reasoning documentation explains. You can get a short answer and still pay for a lot of output.

The price rises before you fill the context window

Sol's model specification lists a 1,050,000-token context window, with at most 922,000 input tokens and 128,000 output tokens. Those limits are separate from its pricing threshold.

Once input exceeds 272,000 tokens, the full request moves to higher rates: $4 for input, $0.20 for cached reads, $5 for cache writes and $15 for output, per million tokens. Loading a large repository or document collection can therefore change the bill before you approach the model's capacity.

Check both limits when planning your agent. Fitting the prompt into the window doesn't make it cheap, or prove the model can reliably retrieve material from every part of it.

One Index point behind Astra, at about 22% of its task cost

Artificial Analysis's model pages give the following snapshot, checked on September 30, 2026:

Evaluated configuration	Intelligence Index	Cost per Index task
GPT-6 Sol, max	48	$1.05
GPT-6.1 Sol, max	52	$0.72
GPT-6 Astra, max	53	$3.26
Claude Opus 5.5, max with fallback	58	$5.98

Sources: the evaluator's pages for GPT-6 Sol, GPT-6.1 Sol, Astra and Opus.

Sol's $0.72 is about 22% of Astra's $3.26. It gains four points over the earlier Sol while lowering measured task cost, despite unchanged ordinary token rates. You get better price-performance on this evaluation; equal success rates on production jobs still need testing.

The Index combines scores rather than reporting a percentage of tasks completed. Its task cost is a weighted benchmark average, so don't treat it as a quote for fixing your app. Keep Claude's fallback label attached when sharing its result: that configuration differs from a bare model run.

The live Sol page showed 69.3 output tokens per second when checked. That reading can move, and it leaves out time spent in tools or waiting for the answer to start.

Astra still leads OpenAI's scientific evaluation

OpenAI describes Sol as close to Astra across agentic coding, computer use and professional work. Its launch post reports an average Terminal-Bench Science cost of $5.47 per task, compared with $23.21 for Opus 5.5 and $23.80 for Astra, at maximum effort.

The same post says Astra has the highest score among the tested models on that scientific evaluation, at 68.1%. OpenAI still recommends it for the most difficult scientific research.

These are OpenAI's results, measured in its research environment or API, with competitor results drawn from public reports. Keep that label separate from Artificial Analysis's measurements.

A lower task bill and a higher success rate are separate outcomes. Sol can suit your budget while Astra completes difficult work Sol misses. Decide how much those misses cost you before choosing a default.

A week between releases, and a tool-calling catch

The API changelog dates GPT-6 Sol to September 22 and GPT-6.1 Sol to September 29. The cache-read rate dropped by half, and the independent Index score rose. The ordinary $2 input and $10 output rates stayed put.

Check your integration before switching. Earlier GPT-6 Sol supports none reasoning effort and Chat Completions function calling when reasoning is off. GPT-6.1 Sol supports low, medium, high, xhigh and max, with medium the default. It supports neither none nor minimal.

The new model requires the Responses API for tools. Chat Completions works without tool calling, so changing only the model string in an existing function-calling integration can break your request.

The quick release cadence has developers questioning the earlier launch. In the OpenCode Reddit discussion, WiggyWongo asks why GPT-6 Sol arrived only a week before its successor. Fair question. For your migration, compatibility and accepted results still decide the answer.

Before migrating, inventory your endpoint, reasoning setting and tool usage. Replay representative requests, including tool calls, and keep your current model available while investigating failures. A small check now beats debugging a live integration after switching everything.

Work, Codex and API access; regular Chat has to wait

OpenAI's launch post lists ChatGPT Work and Codex access for Plus, Pro, Business, Enterprise and Edu users. It explicitly says Sol is not yet available in regular Chat. Developers can access it through the API.

Work and Codex use subscription allowances and credits, with accounting separate from API billing. To learn how many jobs your subscription will finish, measure its consumption: model choice, context, reasoning and tools all affect it.

The model accepts text and image inputs and produces text. The tools listed in its specification don't give it native image, speech or video output. Budget separately if your workflow calls an image-generation tool.

Sol Ultrafast is announced for the coming days. On September 30, it's still pending, so you'll need an available speed tier for a production rollout.

When to make Sol your default

Test Sol on repeated coding, document and browser tasks where inference cost limits how much work you can run. Stable prompt prefixes put the cheaper cache rate to use. Constantly changing instructions or a need for your strongest scientific reasoning model weaken the case.

Give each task an acceptance check: a test suite that catches the intended bug, a document with verifiable figures, or a browser workflow with a clear final state. Record accepted results, retries, total billed tokens and elapsed time. Count reviewer time when an answer needs a lot of repair.

Then compare a modest reasoning setting with a higher one. The published max-effort benchmark tells you about that configuration. It doesn't tell you to run every request at maximum effort.

I'd test Sol as the economical default and keep Astra for tasks where your results justify the extra cost. The one-point Index gap is a good reason to run that experiment.

FAQ

How much does GPT-6.1 Sol cost?

Standard API rates are $2 input, $0.10 cached input, $2.50 cache writes and $10 output per million tokens at up to 272,000 input tokens. Longer prompts cost more.

What is the GPT-6.1 Sol API model ID?

Use gpt-6.1-sol. Tool calling requires the Responses API. Chat Completions is supported without tools.

Is GPT-6.1 Sol available in ChatGPT chat?

OpenAI says it is available in ChatGPT Work and Codex on eligible paid plans, but not yet in regular Chat.

Is Sol as good as GPT-6 Astra?

Artificial Analysis scores max-effort Sol at 52 and Astra at 53. That close composite result does not mean equal performance on every task.