GPT-6.1 Sol: Nearly Astra-level performance at one-fifth the price

What's new

OpenAI unveiled GPT-6.1 Sol at its DevDay 2026 developer conference on September 29, 2026. It is an update to GPT-6 Sol and is meant as a middle ground: nearly as capable as OpenAI's flagship GPT-6 Astra, but significantly cheaper.

The model has become the default coding model in ChatGPT and in Codex, OpenAI's coding tool. According to the release notes, it is also the default in the model catalog in Codex CLI 0.159.1 and in catalogs for Amazon Bedrock. It can be used through OpenAI's API, in ChatGPT Work and in Codex.

Codex has also gained several new features: It now runs in cloud environments, so you can start a task on one device and pick it up again on another. Voice control has been added, along with an /agents view for assigning and tracking tasks, and security features for reviewing pull requests (proposed code changes) on GitHub and GitLab.

What's smart

The most practical feature is the cloud capability. A task can keep running even if you close your laptop, and you can pick it up from another device. With the /agents view, you can hand off several tasks and keep an eye on them in one place.

OpenAI also says the share of answers containing factual errors fell from 11.4 to 7.7 percent. That is OpenAI's own measurement, and it deliberately uses difficult conversations, so it does not reflect typical use.

An Ultrafast tier is on the way. It can generate up to 8 times faster, up to 300 tokens (small chunks of text) per second. According to DevX, it already exists for Astra, while the version for Sol is coming soon.

Cheaper or better?

API pricing is $2 per million input tokens and $10 per million output tokens. Cached input, meaning text the model has seen before and remembers, costs $0.10 per million tokens, which is 95 percent below the standard input price.

It is not cheaper than its predecessor: According to DevX, the prices are the same as for GPT-6 Sol. The savings come compared with Astra, where OpenAI says Sol costs one-fifth of the standard price for input and output, or roughly 80 percent less. OpenAI also says the model approaches Astra in several tests, but that Astra still leads on the hardest reasoning and computer-use tasks. The test results come from OpenAI itself.

A developer quoted by InfoQ summed it up as "Luna pricing, Astra-level quality." Monthly prices for regular users are not given in the sources, and it is also unclear whether it is available in Denmark or the EU.

What it's good for

Sol is aimed at coding and at developers and companies that have AI work through large volumes of tasks where price matters. Cached input makes it especially attractive when the same background text is used over and over.

The flexible cloud setup and code review on GitHub and GitLab suit teams that want help reviewing other people's code. For the very hardest problems, OpenAI itself still points to Astra.