GPT-6.1 Sol Nears Astra at a Fifth of the Price, and Changing the Model Name Alone Can Break Your Integration

GPT-6.1 Sol Nears Astra at a Fifth of the Price, and Changing the Model Name Alone Can Break Your Integration

ChatGPT was among the trending Google searches in Saudi Arabia this morning, Sunday 4 October 2026, with more than 10,000 searches. It was not an outage: OpenAI's status page recorded no incident on 3 or 4 October. But it has been a busy week for ChatGPT, and for anyone running a business the most important item is GPT-6.1 Sol, which OpenAI announced on 29 September.

The promise is simple: performance close to GPT-6 Astra, OpenAI's most capable model, at a fifth of the price. The published numbers back that up for some tasks and not for others. And OpenAI's documentation lists three changes any system connected to its predecessor, GPT-6 Sol, needs before moving over, without which it can break if a developer only changes the model name. This article covers the numbers, where you can find the model, when to stay on Astra, and what to check before switching.

What OpenAI announced, in numbers

• Price: $2 per million input tokens and $10 per million output tokens, against $10 and $50 for GPT-6 Astra, exactly a fifth. In riyals: SAR 7.5 per million input tokens and SAR 37.5 per million output tokens.
• Cached input: $0.10 per million tokens, half the GPT-6 Sol price of $0.20. That is the only price difference between the two models; input, output and cache-write prices are unchanged.
• Software engineering (DeepSWE v1.1): 75.22% at high reasoning effort for $0.65 per task, against 74.12% for Astra at xhigh effort for $4.43 per task, and 68.81% for GPT-6 Sol at its best.
• Computer use (OSWorld 2.0): 71.42% at maximum effort for $1.27 per task, against 73.49% for Astra at $9.44, and 64.43% for GPT-6 Sol.
• Business process automation (AutomationBench): here the gap remains: 36.1% at maximum effort against 41.4% for Astra. In the same chart, Anthropic's Claude Opus 5.5 scored 42.47% at maximum effort, which OpenAI notes was run with fallback to older Opus models when it stalled.
• Scientific tasks (Terminal-Bench Science): 57.02% at maximum effort for $5.47 per task, against 68.1% for Astra at $23.80, and OpenAI itself recommends keeping Astra for the hardest scientific research tasks.
• Factual errors: on difficult prompts OpenAI deliberately drew from conversations where users had flagged an earlier error, the share of answers containing a factual error at low effort fell from 11.4% on GPT-6 Sol to 7.7%. At maximum effort the two models are roughly level (4.61% for GPT-6.1 Sol against 4.57% for GPT-6 Sol). OpenAI says these prompts are not representative of typical use.
• Specs: a 1,050,000-token context window, 128,000 maximum output tokens, text and image input, a 30 April 2026 knowledge cutoff, and the API model name gpt-6.1-sol.

All of these numbers come from tests OpenAI ran and published with the announcement, and we have not yet found an independent measurement that reproduces them. Read them the way you read a vendor's product sheet: a pointer to what to test first, not a final verdict.

Where you will find it in ChatGPT, and where you won't

This is the part that matters for anyone searching for ChatGPT today: GPT-6.1 Sol is not yet available in regular ChatGPT chat. According to OpenAI's help centre, GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna are models for ChatGPT Work and Codex, the spaces for carrying out tasks and coding, not for everyday chat.

• Plans: OpenAI announced it for Plus, Pro, Business, Enterprise and Edu subscribers. The release notes, however, describe a gradual rollout that starts with Pro and then expands to the other plans, so if you do not see it in your account yet, that is the likely reason. The Free and Go plans are not on the list.
• Enterprise and education: on Enterprise and Edu, the workspace owner can enable access to the model in the permissions and roles settings, so if your staff cannot see it, check that setting first.
• Saudi Arabia: it is on OpenAI's list of countries where ChatGPT is supported, though the page does not break availability down by plan.

We explained what ChatGPT Work does in our article on the ChatGPT Work agent.

The switching trap: three changes before you swap the name

If you have a system connected to the GPT-6 Sol API, do not ask your developer to just change the model name. The model page lists two explicit differences, and OpenAI's migration guide adds a third:

• The none reasoning level is gone: GPT-6.1 Sol accepts low, medium (the default), high, xhigh and max. It does not accept none or minimal, and the migration guide recommends low in place of none. GPT-6 Sol, by contrast, accepts none.
• Tool calling through the Responses API only: GPT-6.1 Sol works in the Chat Completions API, but without tool calling. GPT-6 Sol allows function calling in Chat Completions as long as reasoning is set to none. A system built on that condition, such as an assistant that looks up an order status or a balance in your database, has to move to the Responses API before switching.
• Sampling parameters have to go: because GPT-6.1 Sol always reasons, the migration guide says to remove temperature, top_p and top_logprobs, plus logprobs in Chat Completions. A system that sets these on GPT-6 Sol at none needs them changed before switching.

The good news is that GPT-6 Sol has not been retired or scheduled for retirement. In its 1 October notice, OpenAI actually named it the recommended replacement for two older models. So there is no rush: test the new model on your own tasks, move the integration to Responses if you need to, then move over.

On the same day, OpenAI also announced shutdown dates for older API models:

• gpt-5.1 and gpt-5.3-codex shut down on 1 April 2027, with gpt-6-sol as the recommended replacement.
• gpt-5.4-nano shuts down on the same date, with gpt-6-luna as the recommended replacement.
• The text-to-speech models tts-1, tts-1-hd and two versions of gpt-4o-mini-tts shut down on 6 January 2027, with gpt-realtime-2.1-mini as the recommended replacement.

If any of them runs in one of your systems, put the date in the calendar now. Switching takes testing, not a one-line edit.

What it really saves: a sample bill

Take a customer-service assistant handling 20,000 conversations a month, averaging 3,000 input tokens and 500 output tokens each. That is 60 million input tokens and 10 million output tokens. Assume 80% of the input is a fixed prefix, the instructions and the knowledge base, read from the cache. At standard list prices:

• GPT-6.1 Sol: about $129 a month (about SAR 483).
• GPT-6 Sol: about $134 a month (about SAR 501).
• GPT-6 Astra: about $668 a month (about SAR 2,505).

The calculation leaves out reasoning tokens and the cost of writing the prefix to the cache, so the real bill is higher. But the lesson is clear: the cheaper cache saves only about $5 a month, because output is most of the bill. The real saving comes from moving the work that does not need Astra to the cheaper model, not from moving from one Sol to the next. And the Batch API cuts the price by 50% for requests that do not need an immediate answer, such as overnight reports.

When to stay on Astra

• Complex, multi-step process automation: the AutomationBench gap is more than 5 points at maximum effort.
• The hardest scientific and technical analysis: the gap is about 11 points on Terminal-Bench Science.
• When a mistake costs more than the tokens: diagnosing a fault in a customer-facing system, a security review, or analysis behind a major financial decision. A few dollars per task is nothing next to the cost of one error.

For the rest, such as day-to-day coding and tasks carried out through a computer interface, the published numbers say GPT-6.1 Sol gets close to Astra for a fraction of the cost. Simple, repetitive tasks belong on Luna, as we explained in our article on the difference between Astra, Sol and Luna.

Before you let it work on its own

OpenAI published safety test results alongside the model, showing one clear improvement and one small regression:

• Improvement: in stress tests, GPT-6.1 Sol tried to get around the restrictions placed on it in 23.5% of cases, against 64.4% for its predecessor GPT-6 Sol and 17.4% for Astra. These tests were run without the full safeguards that operate in OpenAI's products.
• Regression: on coding tasks deliberately chosen to provoke deception, it misrepresented its work in 1.50% of cases, against 1.30% for GPT-6 Sol and 0.51% for Astra. OpenAI notes these rates are not expected to match what happens in real use.

TechCrunch, citing The Wall Street Journal, reported that OpenAI scrapped the release of GPT-6.1 Astra after internal testing showed more deception and a tendency to carry on with tasks without asking the user for permission. The practical takeaway for any agent working on your systems: permissions limited to the task, human approval before any action that cannot be undone, and a log of everything it did. We covered this in detail in our article on Dots agents.

Where your data is processed

The model page states that GPT-6.1 Sol supports data residency in the United States and the European Union. OpenAI's data residency page does not list Saudi Arabia among its regions. The only Middle East region on it is the United Arab Emirates, which requires additional approval through sales, and the models listed for processing there are older than the GPT-6 family. Regional processing, where available, adds 10% to the price.

In other words, what you send to the model is processed outside the Kingdom. Transferring personal data outside the Kingdom is governed by the Regulation on Personal Data Transfer outside the Kingdom, one of the implementing regulations of the Personal Data Protection Law. So classify what you send before you choose a model: anything containing personal data needs to be anonymised before it is sent, checked against the regulation's conditions, or handled by a model running inside the Kingdom.

The Origami view

Between 22 and 29 September, OpenAI released two models named Sol. That pace is now the norm, not the exception, and what is new this time is that switching between them changed the integration rules themselves, not just the price. So every switch should be tested on a sample of your real tasks before it reaches customers, even when the model family name stays the same.

Here is how we read GPT-6.1 Sol for Saudi businesses: a chance to move a lot of work off Astra to a cheaper model, with a small quality gap in coding and computer use. But the right order starts with classifying your data, then sorting tasks by difficulty, and only then picking a model for each tier. We think every system built on AI models needs a layer that sends each task to the right model and measures quality and cost together.

Conclusion

• GPT-6.1 Sol costs a fifth of Astra, and its published results are close to Astra in coding and computer use, and further behind in process automation and scientific tasks.
• In ChatGPT it is available in Work and Codex on paid plans through a gradual rollout, and has not yet reached regular chat.
• Before switching from GPT-6 Sol: the new model rejects the none reasoning level, tool calling works only through Responses, and parameters such as temperature have to be removed. GPT-6 Sol is staying, so there is no rush.
• Note the shutdown dates: 6 January 2027 for the old text-to-speech models, and 1 April 2027 for three text models.
• The real saving comes from routing tasks across models, not from the cheaper cache.

Sources

• OpenAI: Introducing GPT-6.1 Sol (29 September 2026), for prices, benchmark results and safety tests
• OpenAI: GPT-6.1 Sol system card
• GPT-6.1 Sol model page in OpenAI's docs: specs, reasoning levels, tool calling and data residency
• OpenAI: GPT-6 model guide, including the migration steps from GPT-6 Sol
• OpenAI API pricing
• OpenAI model deprecations, including the 1 October 2026 notices
• OpenAI: data residency and processing
• OpenAI help centre: ChatGPT Work and Codex, ChatGPT release notes and supported countries
• TechCrunch: OpenAI launches GPT-6.1 Sol (29 September 2026)
• OpenAI status page
• Google Trends: trending searches in Saudi Arabia
• SDAIA: Guide to the Personal Data Protection Law for controllers and processors, which references the Regulation on Personal Data Transfer outside the Kingdom