GPT-6.1 Sol: Cheaper Cached Input, What Changed From GPT-6 Sol and How It Fits Codex

GPT-6.1 Sol pricing: what changed from GPT-6 Sol

The clearest pricing change is a 50% reduction in cached-input cost: GPT-6.1 Sol charges $0.10 per million cached input tokens, compared with GPT-6 Sol’s $0.20. Standard input remains $2 per million tokens, and output remains $10.

The OpenAI API changelog records GPT-6 Sol and GPT-6 Luna’s September 22 launch, followed by GPT-6.1 Sol’s September 29 release. The following prices apply to the Standard tier for prompts with up to 272,000 input tokens.

GPT-6.1 Sol also has a cache-write price of $2.50 per million tokens. Keep that separate from the cached-input rate when reviewing the published prices: $0.10 is not a universal rate for every token involved in caching.

How much does the cached-input reduction save?

For 10 million tokens billed as cached input, GPT-6 Sol’s listed rate produces a $2 charge, while GPT-6.1 Sol’s produces a $1 charge. That calculation covers only cached input; it excludes ordinary input, output, and cache-write charges.

The distinction matters when interpreting “cheaper.” The cached-input component is half the price, but the total request cost does not automatically fall by half. Requests billed entirely at the ordinary input and output rates retain the same listed token prices.

GPT-6.1 Sol’s cached-input rate is also 95% below its own uncached-input rate: $0.10 versus $2 per million tokens. Longer prompts and other processing tiers have separate pricing, so these figures should not be extended beyond their stated scope. For background on the predecessor, see our GPT-6 Sol overview.

What improved in coding, automation, and factuality?

The GPT-6 Sol vs GPT-6.1 Sol comparison includes measured capability changes as well as cheaper cache reads. OpenAI’s published evaluations report improvements in coding, automation, computer use, and difficult factuality prompts.

These results need their evaluation settings attached. A score at one reasoning effort is not interchangeable with a score at another, and a task-cost result is not the same thing as a token-price reduction.

Coding: higher DeepSWE score at lower effort

The September 30 developer announcement reports a DeepSWE v1.1 score of 75.2% for GPT-6.1 Sol at high reasoning effort. GPT-6 Sol’s best reported result was 68.8% at maximum effort.

That is a 6.4-percentage-point improvement, with approximately 76% lower cost per task in the reported comparison. Both details are useful: the newer model scored higher while using a lower named reasoning setting.

The 76% figure describes that evaluation’s task cost. It does not mean ordinary input or output became 76% cheaper, and it does not establish the same savings for every coding workload.

Automation, computer use, and factuality

The GPT-6.1 Sol launch announcement reports three additional comparisons with GPT-6 Sol:

The factuality qualification is particularly important. Those percentages describe OpenAI’s difficult prompt set, so they should not be presented as the error rate developers will see in ordinary use.

Where Astra fits

OpenAI describes GPT-6.1 Sol as offering “Near-Astra intelligence for a fifth of the price.” That intelligence comparison is OpenAI’s evaluation-based claim.

The Standard input/output price comparison is concrete: GPT-6 Astra costs $10/$50 per million tokens, while GPT-6.1 Sol costs $2/$10. Each Sol rate is one fifth of the corresponding Astra rate. Our GPT-6 Astra overview provides a separate reference for that model.

The caching improvements started before Sol 6.1

Several caching changes associated with this release sequence were already announced with GPT-6 Sol and GPT-6 Luna on September 22. Those included higher default cache-hit rates, a Prompt Caching Dashboard and diagnostics, and explicit prefix breakpoints.

The earlier announcement also described preservation of previously cached context when developers changed reasoning effort or tool availability. GPT-6.1 Sol’s September 29 cached-input price reduction builds on that earlier caching announcement.

Separate caching behavior from caching prices

There are two changes to track: the September 22 caching features and the September 29 reduction in the cached-input rate. Treating all of them as newly introduced by GPT-6.1 Sol would blur the release history.

For a developer evaluating the upgrade, the useful cost questions are how much input receives cached billing, what ordinary input and output contribute, and whether cache-write charges apply. The published price table answers the rate question; it does not establish a universal cache-hit percentage or total savings figure for a particular application.

This also explains why the benchmark cost reductions and the cached-input discount should be assessed separately. The brief’s reported task-cost improvements belong to specific evaluations, while the $0.10 rate applies to tokens billed as cached input within the stated pricing scope.

How GPT-6.1 Sol fits Codex

GPT-6.1 Sol became the default model in the bundled catalog with Codex CLI 0.159.1, released September 29. That release also added it as the default in the Amazon Bedrock Mantle and Runtime catalogs.

The Codex CLI 0.159.1 release establishes the dated default change. By September 30, the documented installation command used version 0.159.3:

npm install -g @openai/codex@0.159.3

Version 0.159.3 added optional account-security setup reminders. The model-default change had already arrived in 0.159.1.

Model identifier, reasoning settings, and context

The versioned 0.159.1 model catalog identifies the model as gpt-6.1-sol and describes it as “Latest workhorse model for coding and everyday work.”

These are Codex catalog settings. In particular, the context fields should not be used as a statement of the API’s maximum context window.

The low default also puts the coding benchmark in perspective: the 75.2% DeepSWE result used high effort. Selecting the model with its default reasoning level does not reproduce that evaluation setting.

Availability depends on more than the catalog

At launch, GPT-6.1 Sol was available in ChatGPT Work and Codex to Plus, Pro, Business, Enterprise, and Edu users, as well as through the API. It was not yet available in Chat.

Codex availability depends on plan, client, and workspace settings. The catalog default therefore needs to be read alongside those access conditions. Developers comparing their coding-tool options can use our Claude Code vs Codex guide as a separate reference.

API support, delegation, and Ultrafast status

GPT-6.1 Sol launched with support for both /v1/responses and /v1/chat/completions. The September 29 API entry directs developers to use the Responses API for tool calling.

That entry also announces Multi-agent support in beta, allowing delegation to subagents within a Responses request. The beta status is part of the release detail and should remain attached when describing the feature.

Choose the endpoint around the documented workflow

For a tool-calling integration, the launch guidance points to Responses. Support for Chat Completions does not change that specific instruction.

The model identifier is gpt-6.1-sol, but endpoint support, delegation support, and Codex defaults describe different parts of the release. An API integration should be evaluated against the API announcement; a Codex setup should also account for its client catalog and access settings.

Ultrafast was still “coming soon”

GPT-6.1 Sol Ultrafast was announced as “coming soon,” with up to 8× faster token generation in Codex promised. Its availability by September 30 was unconfirmed in the dated material covered here.

The speed figure is therefore an announced promise, not evidence of an already available mode at this publication cutoff. It should not be folded into claims about the released model’s current Codex performance.

Key takeaways

• GPT-6.1 Sol launched September 29, 2026, with the model identifier gpt-6.1-sol.
• Standard cached input fell from $0.20 to $0.10 per million tokens; ordinary input and output remained $2 and $10 for prompts up to 272,000 input tokens.
• OpenAI reported a 75.2% DeepSWE v1.1 score at high effort, versus GPT-6 Sol’s best 68.8% at maximum effort, with approximately 76% lower cost per task.
• Codex CLI 0.159.1 made GPT-6.1 Sol the bundled catalog default, with low as its default reasoning level.
• GPT-6.1 Sol Ultrafast availability remained unconfirmed as of September 30; its promised speed increase was not a released-performance result.

FAQ

What is GPT-6.1 Sol, and when did it launch?

GPT-6.1 Sol is the model released September 29, 2026, under the identifier gpt-6.1-sol. It supports the Responses and Chat Completions API endpoints and was available at launch through the API, ChatGPT Work, and Codex, subject to access conditions.

How much cheaper is GPT-6.1 Sol than GPT-6 Sol?

Its Standard cached-input rate is 50% lower: $0.10 instead of $0.20 per million tokens. Ordinary input remains $2 and output remains $10 per million tokens for prompts up to 272,000 input tokens, so total savings depend on the billed token mix.

Is GPT-6.1 Sol the default in Codex?

Codex CLI 0.159.1 made GPT-6.1 Sol the default in its bundled model catalog. The catalog sets default reasoning to low, while availability depends on plan, client, and workspace settings.

Was GPT-6.1 Sol Ultrafast available on September 30?

Availability was unconfirmed in the dated material reviewed for the September 30 cutoff. Ultrafast had been announced as “coming soon,” with up to 8× faster token generation in Codex promised.