GPT-6.1 Sol: OpenAI brings Astra to one-fifth of the price

On September 29, OpenAI introduced GPT-6.1 Sol, just one week after launching GPT-6 Sol. Having another version come out so quickly is noteworthy, but the change that really matters is somewhere else: performance and cost.

OpenAI is positioning Sol 6.1 as a model for agentic coding, computer use, and complex professional work that gets quite close to GPT-6 Astra, but with a considerable difference in price.

That changes the conversation for those of us who use these models as a daily tool.

Five times cheaper than Astra

The published prices for GPT-6.1 Sol are:

Additionally, cached input for GPT-6.1 Sol drops to US$0.10 per million tokens. This is especially interesting for applications that reuse large instructions, documents, or frequent context.

The figures were reported alongside the launch and place Sol in a rather particular category: we're no longer talking about a budget model that sacrifices much of the capability of the higher-tier model. OpenAI is trying to bring it much closer to Astra's level.

Sources: OpenAI — ChatGPT Work and Codex and TechCrunch — GPT-6.1 Sol launch. OpenAI Help Center

The big shift is in coding

This is where it gets really interesting for me.

DeepSWE 1.1 measures software engineering tasks on real repositories. Depending on the level of reasoning used, GPT-6.1 Sol comes very close to Astra and even surpasses it in some published results.

For example, with one of the reported High configurations, Sol 6.1 scored 75.22%, while Astra came in at 73.23%.

The average cost of that evaluation was even more striking: approximately US$0.65 per task with Sol versus US$3.92 with Astra. HeyMark

This should be interpreted carefully. A specific benchmark doesn't prove that Sol is better than Astra overall. Changing tools, the amount of reasoning, or the type of problem can completely change the result.

In fact, in more demanding scientific tests, Astra maintains a significant advantage.

Useful reading is another matter: there are already development tasks where paying five times more for the superior model can contribute very little.

It also improves when working with the computer

OpenAI points to three main areas for GPT-6.1 Sol:

• agentic coding;
• computer use;
• multi-step professional work.

In OSWorld 2.0, an evaluation related to computer use, Sol 6.1 also gets quite close to Astra depending on the configuration.

That matters because current models are no longer limited to responding with text.

When an agent needs to open tools, interpret documents, execute several actions, and verify the result, each step consumes time and tokens. A small price difference multiplied by hundreds or thousands of executions ends up being a lot of money.

That's where I find Sol's real value.

Sol 6.1 doesn't replace Astra in everything either

It would be easy to stick with the headline of "Astra performance for one-fifth the price," but the results themselves show a more nuanced story.

In Terminal-Bench Science 0.1, for example, using the maximum reasoning level, Sol 6.1 appears at 57.02% while Astra reaches 68.10%.

The difference is considerable.

The cost also changes: approximately US$5.47 per task for Sol versus US$23.80 with Astra within that evaluation. HeyMark

So I still see Astra for problems where I need to squeeze out the maximum available performance. For a large amount of everyday work, Sol is starting to look like the much more logical option.

Where can it be used?

There's a detail that can cause confusion.

GPT-6.1 Sol is not currently available as a model in the conventional ChatGPT chat.

OpenAI is rolling it out in ChatGPT Work and Codex, starting with Pro users and expanding access to Plus, Business, Enterprise, and Edu. In Enterprise and Edu organizations, administrators must enable it. OpenAI Help Center

GitHub also announced that GPT-6.1 Sol began rolling out generally in GitHub Copilot, especially for agentic coding and jobs involving several consecutive operations. The GitHub Blog

For developers, it's probably where it will first make sense to try it.

Ultrafast: paying more to wait less

Along with Sol 6.1, another interesting piece appears: Ultrafast.

It's a premium inference mode designed to significantly increase generation speed. Launch reports mention up to eight times more speed within Codex and up to six times in the API, depending on the model and configuration. Venturebeat

It may seem like a luxury until we think about agents.

Waiting an extra ten seconds for an isolated response almost never matters. Waiting that long in a flow that performs fifty consecutive operations does.

In that scenario, latency starts to become part of the cost.

The context around Astra also matters

The launch also comes at a strange time for OpenAI.

The company halted the release of GPT-6.1 Astra after internal evaluations found issues related to safety, authorization tracking, and autonomous behavior. Reuters and AP reported the decision on September 28 and 29. Reuters

That temporarily leaves Sol 6.1 in an even more interesting position.

OpenAI can continue advancing in capability and efficiency with a version it has decided to deploy, while keeping the update to its most powerful model on hold.

For those of us building products with AI, this also serves as an important reminder: the technically most capable model won't always be the first one worth putting into production.

My takeaways

The improvement I care most about with GPT-6.1 Sol isn't that it managed to score a few extra points on a benchmark.

It's that the gap between mid-tier models and flagships is shrinking much faster than the price gap.

Not long ago, it made sense to reserve budget models for simple processes and move up to the strongest one when a task started getting complicated.

With results like these, I'm going to try just the opposite: start with Sol and scale up only when I actually find a case where Astra makes a measurable difference.

That also forces a change in how we evaluate models.

Rather than asking which one gets the best absolute score, I'm interested in measuring how much it costs to correctly complete a real task from my project.

Conclusion

GPT-6.1 Sol seems to represent quite well where this generation of AI is heading: models capable of coding, handling tools, and completing long processes, but with a cost low enough to use them continuously.

I still want to see independent testing after several days of real-world use. A controlled benchmark and a repository with years of technical debt usually tell different stories.

But if the difference compared to Astra turns out to be as small as these initial evaluations show, US$2 for input and US$10 for output per million tokens make GPT-6.1 Sol one of the OpenAI releases most worth trying in development.

And that does have a concrete effect: more tasks where I can use a strong model without first having to justify the cost of the flagship model.