If you use GPT-6 Sol in Codex, trying GPT-6.1 Sol at Medium effort is a reasonable next step. OpenAI added 6.1 Sol to Codex and ChatGPT Work on September 29, 2026, and recommends it for complex coding and sustained work. However, the official sources we checked do not give a direct percentage improvement in accuracy or speed over 6 Sol. Keep the reasons to try the newer model separate from what has yet to be measured.
What this article covers
What changed between 6 Sol and 6.1 Sol?
OpenAI's Codex changelog dates the addition of 6.1 Sol to September 29, 2026. GPT-6 Sol was released about a week earlier, on September 22 (US time). The 6 Sol model page currently describes it as designed for complex coding and agentic tasks, while the 6.1 Sol page describes performance close to Astra on complex coding, computer use and professional work. “Close to Astra” is OpenAI's positioning, not a measurement of the difference from 6 Sol.
| Item | GPT-6 Sol | GPT-6.1 Sol |
|---|---|---|
| Model ID | gpt-6-sol | gpt-6.1-sol |
| Official positioning | Complex coding and agentic tasks | Complex coding, computer use and professional work |
| Knowledge cutoff | April 20, 2026 | April 30, 2026 |
| Context window / maximum output | 1,050,000 / 128,000 tokens | Unchanged |
| API reasoning effort | none, low through max | low through max (none unsupported) |
| Input / output | Text and images / text | Unchanged |
Figures and supported features: OpenAI API specifications for 6 Sol and 6.1 Sol. The context window limits input and output combined; it does not mean the model can write 1,050,000 characters in a single response.
Neither model knows September's news from its training alone. When writing about recent developments, finding the original source and checking its publication date and contents matters more than switching models.
Prices are unchanged except for cached input
At standard API speed, with no more than 272,000 input tokens per request, both models charge $2 per million standard input tokens and $10 per million output tokens. Only cached input falls from $0.20 with 6 Sol to $0.10 with 6.1 Sol. Cache writes cost $2.50 for both. Repeatedly supplying the same material can save money when the cache actually hits, but the price table alone offers no savings when all input is new (see OpenAI API pricing).
Per million tokens. New input costs $2.00 for both models.
If you sign in to Codex with a ChatGPT account, do not convert API prices directly into your bill or available task count. Codex's official pricing documentation lists the same standard-speed additional-credit rates for both models: 50 credits per million input tokens and 250 per million output tokens. Cached input costs 5 credits with 6 Sol and 2.5 with 6.1 Sol. However, subscription usage varies with task length, tool calls, reasoning, caching and other factors. Halving one credit rate does not guarantee twice as much use within your plan.
Half-price caching does not halve the cost of the whole task
Example: 1 million new input tokens, 10 million cached input tokens and 1 million output tokens.
- New input
- $2
- Cached input
- $2
- Output
- $10
- New input
- $2
- Cached input
- $1
- Output
- $10
Consider an example using API token prices alone: 1 million new input tokens, 10 million cached input tokens and 1 million output tokens, excluding cache writes and tool fees. In the breakdown above, the $10 output cost stays the same, and only the cached-input difference changes the total. Cached input costs half as much, but the overall saving in this example is only about 7%. Without caching, the difference under this price table is zero. Repeatedly reading long, identical documents can make the saving larger. In practice, though, a new model may use different amounts of output, reasoning and retries. Measuring costs task by task is necessary to establish the total saving.
For API requests exceeding 272,000 input tokens, the entire request is charged at twice the input and cache rates and 1.5 times the output rate. The same rates do not necessarily apply all the way to the 1,050,000-token context limit.
The same pricing page estimates about 15–150 local messages per five hours for 6 Sol and about 15–160 for 6.1 Sol on Plus and Standard Business. These are not fixed counts. Pro currently has no five-hour allowance limit, although weekly and other limits may apply. Check your usage screen for your remaining allowance and reset times. If you connect Codex with an API key, API charges are separate from your subscription.
A newer version does not establish a percentage gain
In its model selection guide, OpenAI suggests 6.1 Sol at Medium as a starting point for complex technical work and deliverables that need revision, and Extra High for more polished deliverables. Its Codex Models page also recommends 6.1 Sol, when available, for complex coding and agentic tasks. This official recommendation is a good reason to try it for everyday work.
However, public scores or runtimes comparing 6 Sol and 6.1 Sol on identical tasks, reasoning effort and tool settings were absent from the OpenAI model specifications, changelog and selection guide we checked. “Close to Astra” positions the model relative to Astra; it is not evidence that it is faster than 6 Sol. Different output token counts can change total costs even when unit prices match. Tool wait times and rework also affect results, particularly in long repository tasks.
Claims such as “the design looks better” or “it follows instructions more closely” are especially hard to judge from one attempt. For a web-page fix, give both models the same completion criteria and compare desktop and mobile screenshots at matching widths. Check readability, spacing, component consistency, horizontal overflow and whether implementation tests pass. A fluent answer is not a success if the required changes are missing. For article proofreading, score factual accuracy against official sources separately from whether the reader's next step is clear. By defining completion criteria before starting, you make the value of switching models easier to assess.
How to switch in Codex
In the desktop app, use the model selector near the input box to choose 6.1 Sol and set reasoning effort to Medium. Some screens require opening advanced settings. According to the official model guide, availability is rolling out to Plus, Pro and Business, while Enterprise and Edu require administrator enablement. Free and Go are excluded at launch. The release is for Work and Codex rather than regular ChatGPT Chat, so do not look only in regular Chat's model list. If the model is missing, check app updates and availability for your account.
In the Codex CLI, use the following command to select the model and reasoning effort for one launch. This lets you try 6.1 without changing your existing work settings.
codex --model gpt-6.1-sol -c model_reasoning_effort="medium"To change defaults for newly created local chats, place these values in your user config.toml without duplicating existing settings. OpenAI documents that the desktop app, CLI and IDE extension share this file.
model = "gpt-6.1-sol"
model_reasoning_effort = "medium"Changing this default alone does not change the model in every saved chat. Check the settings near the input box in each existing chat. Our guide to bulk changes of Codex model and reasoning settings covers the technical tests and limitations. OpenAI does not document a one-click switch for all existing chats, and bulk changes are not a reason to interrupt someone's active chat without asking.
Developers migrating API calls should do more than replace the model ID. Because 6.1 Sol does not support none reasoning, requests that used none with 6 Sol should switch to low and be evaluated again. Use the Responses API for tool calling. Requests with reasoning enabled also impose restrictions on parameters such as temperature and top_p, so check the migration guide before reusing older parameters. The official GPT-6 migration guide recommends preserving your effective reasoning effort and validating behavior on representative tasks.
For example, if a 6 Sol workflow uses medium to read a repository, identify a cause, implement a fix and run tests, initially keep medium with 6.1 Sol. Changing prompts, tools and testing methods at the same time as the model makes it hard to isolate the cause of a different result. If you used 6 Sol at none only for classification, try low with 6.1 Sol, but also consider comparing GPT-6 Luna for frequent tasks with clear requirements; it may fit the use case better.
How to choose for your own work
For everyday Codex use, start with 6.1 Sol at Medium. Official recommendations cover multistep coding, research and work spanning apps or documents. For short, clearly scoped changes, test whether Light is enough. Bring Astra back into consideration for the hardest design problems or reviews where a wrong judgment has significant consequences. For repetitive classification or extraction, compare Luna as well. These are starting points based on task difficulty and how well you can verify the result, rather than a claim that the model name alone determines quality (see OpenAI's use-case guidance).
Specify the deliverable in your comparison prompt
For coding, you might ask: “Fix this function's edge cases without changing the existing public API. Show one failing input, then run the relevant tests after the fix and report the results.” Start both models from the same repository state and check the correctness of the fix, test results and unnecessary changes. Simply asking for “better code” makes it harder to decide which output to adopt.
For writing, fix the requirements: “Do not add numbers absent from this material. Put the conclusion first and give readers a three-step procedure they can actually try.” Mark down unsupported facts added merely because the model is newer. Separately assess research speed, accuracy against original sources, and whether figures and tables help explain the topic. In practice, choose a model by whether it reduces rework in the tasks you repeat, rather than who wins one attempt.
For how the earlier GPT-6 Sol and Luna compare with Astra, including launch-time assessments and small hands-on tests, see our GPT-6 Sol and Luna article. Do not reuse those figures as results for 6.1 Sol. We will update this assessment if primary sources or reproducible tests compare both Sol models under the same conditions.
Conclusion: try the upgrade at Medium effort
GPT-6.1 Sol is OpenAI's new recommendation for people doing sustained, complex work in Codex. Standard API input and output prices are unchanged from 6 Sol, while cached input costs half as much. The context limit is unchanged, and none reasoning is no longer supported. A sensible everyday starting point is 6.1 Sol at Medium. Published sources do not quantify the gain in speed or success rate. For important deliverables before publication, verify original sources, run tests and inspect the actual interface whichever model you choose. Compare the same work before deciding which model to adopt.
Frequently asked questions
Is 6.1 Sol always faster than 6 Sol?
The official OpenAI sources we checked do not report runtimes measured under identical conditions for both models. Comparing speed requires matching the task, reasoning effort and tool wait conditions. A version number alone does not establish which is faster.
Does switching to 6.1 Sol incur an extra charge on Pro?
When you use Codex with a ChatGPT account, selecting the model alone does not create a separate API charge. You use your plan's allowance; if you buy and use additional credits, that balance also applies. API prices apply when you use an API-key configuration. First check how you are signed in.
Will existing chats switch too?
config.toml sets the default model for newly started local chats. Check each saved chat's settings separately. Changing a model does not automatically regenerate earlier replies.
Is there a reason to keep using 6 Sol?
Yes, when comparing with earlier measurements or reproducing an existing API workflow under the same conditions. The 6 Sol API also supports none reasoning, whereas 6.1 Sol does not. Check your settings and representative tasks before migrating.
