If you use GPT-6 Sol in Codex, switching to GPT-6.1 Sol is fine. Start by testing it at Medium (standard), the same setting you use now. On September 29, 2026, OpenAI added 6.1 Sol to Codex and ChatGPT Work and recommends it for complex coding and long-running work. API input and output prices are the same, and cached input is half price. This article compares the two only to help you decide whether to switch.

Summary · Everyday use in Codex
Tasks to start with6.1 Sol · MediumA starting point for work that spans research, code, and documents. Compare it with 6 Sol at the same Medium setting.
Why keep your current settingTasks that need comparisonIf you want to reproduce past results, keep 6 Sol too and compare both on the same task at the same effort level.
Pricing essentials: API standard input and output prices are the same. Only cached input is half price for 6.1 Sol. It does not mean you can get twice as many uses from your Codex plan.
The basis for these choices is OpenAI's Models page and model selection guide. Checked on September 30, 2026.

6.1 Sol and 6 Sol: What Changed

In OpenAI's Codex changelog, 6.1 Sol was added on September 29, 2026. GPT-6 Sol was released about a week earlier, on September 22 (US time). The GPT-6 Sol model page describes it as "for complex coding and agentic work," and the GPT-6.1 Sol page describes it as "performance close to Astra on complex work at lower cost." This is a comparison with Astra, not a measured difference from 6 Sol. For an overview of the differences among 6 Sol, Luna, and Astra, see our GPT-6 Sol and Luna explainer.

ItemGPT-6 SolGPT-6.1 Sol
Model IDgpt-6-solgpt-6.1-sol
Official positioningComplex code and agentic workComplex code, computer use, and specialized work
Knowledge cutoffApril 20, 2026April 30, 2026
Context / max output1.05 million / 128,000 tokensSame
API reasoning effortnone, low to maxlow to max (none not available)
Input / outputText, images / textSame

Figures and supported ranges come from the OpenAI API model specifications for 6 Sol and 6.1 Sol. The context window is the combined limit for input and output. It does not mean you can write 1.05 million characters at once.

The knowledge date is newer, but only by 10 days
6 SolApril 20
6.1 SolApril 30

Neither model has learned about September news. If you need current information, it matters more to search the original source and check the publication date and content than to switch models.

For the difference between a "knowledge cutoff" and "information you can check on the web," see our guide to knowledge cutoffs for major AI models.

Price Unchanged, Only Cached Input Gets Cheaper

If you use the API at standard speed with an input of 272,000 tokens or fewer, both models have the same standard input price of $2 and output price of $10 per 1 million tokens. Only cached input is half price: it drops from $0.20 for 6 Sol to $0.10 for 6.1 Sol. Cache writes cost $2.50 for both. You'll see a difference in work where you send the same material repeatedly and the cache is actually hit. If all input is new, there is no difference in unit price (OpenAI API pricing).

API prices for cached input

Per 1 million tokens. Standard new input is $2.00 for both models.

6 Sol$0.20
6.1 Sol$0.10
The bars show the ratio between the two cached input prices. Sources: 6 Sol and 6.1 Sol.

If you are signed in to Codex with a ChatGPT account, do not convert API prices directly into your bill or usage limits. In Codex's official pricing explanation, the standard-speed rate for additional credits is the same for both models: 50 credits per 1 million input tokens and 250 credits per 1 million output tokens. Cached input costs 5 credits for 6 Sol and 2.5 credits for 6.1 Sol. Usage on a subscription plan varies with task length, tool calls, reasoning, and caching, so even though one rate is half price, you do not get twice as much work within your plan.

Half-Price Cache Doesn't Mean Half-Price Work

Same token volume, total goes from $14 to $13

Example: 1 million standard input tokens, 10 million cached input tokens, and 1 million output tokens.

6 Sol
Standard input
$2
Cached input
$2
Output
$10
Total $14
6.1 Sol
Standard input
$2
Cached input
$1
Output
$10
Total $13
This is a calculation example that applies the pricing table to the same volume. It does not include cache write costs or tool fees.

In the example above (1 million standard input, 10 million cached input, and 1 million output tokens), the $10 output cost stays the same, and only the cached input difference affects the total. The total difference is about 7%. Without caching, there is no difference, and the more often you read the same long material, the bigger the gap becomes. In practice, output, reasoning, and retry volume also vary by model, so measure total cost for each task.

Assumptions: This example keeps each request's input at 272,000 tokens or fewer and applies standard API prices to the total token volume across multiple calls. It does not include cache write costs, tool fees for search or computer use, surcharges for long inputs, or taxes. It is also not an estimate of usage limits on a Codex subscription plan.

In the API, if a single request's input exceeds 272,000 tokens, input and cached input prices are doubled and output prices are multiplied by 1.5 for the entire request. The price is not the same all the way up to the 1.05 million token limit.

The guide for Plus and Standard Business on the same pricing page gives about 15 to 150 local messages per 5 hours for 6 Sol and about 15 to 160 for 6.1 Sol. These are not fixed numbers. Pro currently has no five-hour limit, but weekly or other limits may apply, so check your remaining usage and reset time on your usage dashboard. If you connect to Codex with an API key, usage is billed separately from your subscription plan.

Compare Quality and Speed on Your Own Work

In its model selection guide, OpenAI suggests 6.1 Sol at Medium as a starting point for complex technical work and deliverables that involve revision, and Extra High for deliverables that need more polishing. The Models page for Codex also recommends 6.1 Sol for complex coding and agentic work, if it is available to you. This official recommendation is a good enough reason to switch your everyday use.

On the other hand, OpenAI's model specifications, changelog, and selection guide do not include public scores or completion times that measure 6 Sol and 6.1 Sol under the same tasks and reasoning effort. The most reliable way to see how much better it is, is to compare on the work you often ask for. If output token counts differ, total cost changes even when unit prices are the same.

A pitfall when comparingDon't call it a "version difference" if you compare 6 Sol at Medium with 6.1 Sol at Light. Use the same reasoning effort, and look at answer accuracy, rework, completion time, and usage limits together. OpenAI also cautions that reasoning effort levels do not map perfectly across generations.

Judgments like "it looks better" or "it follows instructions better" are hard to make from a single impression. For example, for a web page fix, give both models the same completion criteria, compare PC and smartphone screenshots at the same width, and check whether the tests pass after implementation. For article proofreading, score factual accuracy separately from whether the reader can clearly see what to do next. Defining the completion criteria for the work first makes it easier to measure whether switching models was worth it.

How to Switch in Codex

In the desktop app, choose 6.1 Sol from the model selector near the input field, and set reasoning effort to Medium (standard). Depending on your screen, you may need to open "Advanced settings." According to the official model guide, the rollout is staged for Plus, Pro, and Business, while Enterprise and Edu require an administrator to enable it. Free and Go are not eligible at the time of publication. This is for Work and Codex, not regular ChatGPT Chat, so don't look only in the regular Chat model list. If you can't find the model, check that your app is up to date and that your account has access.

To specify the model and reasoning effort for a single Codex CLI session, start it like this. This is useful when you want to try 6.1 without changing your existing work settings.

codex --model gpt-6.1-sol -c model_reasoning_effort="medium"

To change the default for new local chats, put the following values in your user config.toml file, without duplicating existing settings. According to official documentation, the desktop app, CLI, and IDE extension all use the same file.

model = "gpt-6.1-sol"
model_reasoning_effort = "medium"

Changing this default alone does not change the model for saved chats. Check the setting in each existing chat's input field. For how to change them in bulk and the limits of doing so, see Can you change Codex model and reasoning settings in bulk?. You don't need to interrupt your active chats to change them all at once.

Developers moving from the API should not stop after changing only the model ID. 6.1 Sol does not support none reasoning, so requests that used none with 6 Sol should be changed to low and evaluated again. Use the Responses API for tool calls. Requests with reasoning enabled also have restrictions on temperature, top_p, and similar parameters, so check the migration guide to see whether your existing parameters can be passed as they are. OpenAI's official GPT-6 migration guide recommends keeping your current effective reasoning level and checking behavior on representative tasks.

For example, if you have built a workflow with 6 Sol at medium that "reads a repository, identifies the cause, and carries out the fix and tests," start by running it with 6.1 Sol at medium. If you change the prompt, tools, and test method along with the model, you won't know what caused any change in results. If you used 6 Sol at none only for classification, compare 6.1 Sol at low together with GPT-6 Luna, since one of them may fit your use case better.

How to Choose for Your Own Work

For everyday Codex use, start with 6.1 Sol at Medium. OpenAI recommends it for multi-step coding, research, and work that spans apps and documents. For short, clearly scoped changes, check whether Light is enough. For the most difficult design or high-impact reviews, consider Astra, and for repetitive classification or extraction, consider Luna (OpenAI's guide by use case). For more on choosing among Astra and Luna, see our GPT-6 Sol and Luna article. The figures there are for 6 Sol, not 6.1 Sol.

A short comparison procedure to decide the switch yourself
01 · Pick the same taskFor example, one small bug fix and one article proofread that requires checking sources. Write the criteria for a correct result first.
02 · Match the conditionsUse Medium for both models, the same materials, the same tool permissions, and the same stopping conditions.
03 · Decide by resultsRecord accuracy, number of corrections, waiting time, and usage limits. Don't decide which is better from a single impression.
OpenAI's model selection guide also recommends testing with the same inputs.

For comparisons, specify the deliverable in your prompt

For coding work, for example, you might ask: "Fix the boundary condition in this function. Don't change the existing public API, show one input that fails, and after the fix, run the related tests and report the results." Give the same repository state to both models, and check the correctness of the fix, the test results, and whether there are unnecessary changes. A vague request like "make the code better" makes it hard to judge which output to use.

For writing, fix the conditions with something like: "Don't add figures that aren't in the source, put the conclusion first, and give readers the steps to try in three stages." Deduct points if facts not in the source are added. The practical test is whether it reduces rework in tasks you repeat.

On October 26, 2026, OpenAI also released GPT-6.1 Terra at half the price of 6.1 Sol ($1 input / $5 output). In OpenAI's measurements, Terra beats Sol on the professional-work benchmark, while Sol is ahead on the coding and computer-use benchmarks. We cover how to choose between them in "What Is GPT-6.1 Terra? Pricing, Performance, and How to Choose Between It, Sol, Astra and Claude".

Summary: Try Switching from 6 Sol at Standard Settings

GPT-6.1 Sol is OpenAI's new recommended model for people who keep working on complex tasks in Codex. API input and output prices stay the same as 6 Sol, and cached input is half price. The context limit does not change, and none reasoning is no longer available. For everyday use, 6.1 Sol at Medium is a good starting point. OpenAI does not publish official figures on how much speed or success rate improves, so for important deliverables, compare the same task on both models, and decide only after checking tests and the actual screen output.

Frequently Asked Questions

Is 6.1 Sol always faster than 6 Sol?

OpenAI's official materials do not include completion times measured under the same conditions for both models. Compare them yourself with the same task, reasoning effort, and tool wait times.

Does switching to 6.1 Sol cost extra on Pro?

If you use Codex with a ChatGPT account, choosing a model does not add a separate API charge. Your plan's usage limits apply, and if you buy and use additional credits, that balance is involved. If you use an API key, API pricing applies. Check how you are signed in first.

Do previous chats switch too?

Changing the default model in config.toml applies to new local chats. Check the settings for saved chats individually. Changing the model does not automatically regenerate past responses.

Are there reasons to keep using 6 Sol?

Yes, if you want to compare against past results or reproduce existing API workflows under the same conditions. Also, the 6 Sol API supports none reasoning, but 6.1 Sol does not. Check your settings and representative tasks before migrating.