September 23, 2026
GPT-6 Sol becomes cheaper for agents: OpenAI cuts API prices by 50%
On September 22, GPT-6 Sol arrived in Codex with API pricing of $2 per 1M input tokens and $10 per 1M output tokens. That is half the $4 and $20 pricing of GPT-5.6 Sol. The model is available to Plus, Pro, Business, Enterprise, and Edu users.
Dan Shipper
@danshipper
Breaking: @OpenAI just released GPT-6 Sol. I made it my main everyday model in Codex: it falls just short of Astra, but is close enough for most daily work, responds faster, and costs 50% less than 5.6 Sol. We ran it on the tasks we actually work on at @every and compared it head-to-head with Opus 5.5. Here are my impressions: • Writing: On a task to write a paragraph from my real work, Sol scored almost as highly as Astra, and Astra is still my favorite model for writing. It writes cleanly and concisely, putting the main point first. Opus 5.5 is pleasant to work with, but the main point still gets buried in its drafts. • Computer use: If you like how Astra controls a computer, you will like Sol too. In an early preview, Sol performed as well as Astra in 17 of 18 attempts across six simpler Hands tasks, while costing noticeably less in tokens. • Code: Sol is better than GPT-5.6, including at writing Ruby in @kieranklaassen's tests. But Opus 5.5 has a higher ceiling in long autonomous builds. One thing frustrates me: Codex's new safety classifier repeatedly stopped work we had already approved and asked us to confirm the action. The classifier creates this friction, not necessarily a limitation of Sol, but long runs have become harder to leave unattended. Overall: Sol feels like an S-release iPhone: it has a lot of Astra in it, while costing roughly one-fifth as much as Astra. Opus 5.5 was more surprising. On some of our tasks, it matches or beats Fable 5.1, while costing less than Opus 5. Several team members who switched to Codex began to question their choice after using Opus 5.5. If you are already in the Claude ecosystem, you will like this model. If you use Codex, consider it for your hardest coding tasks. You will like Sol 6 if you spend all day reading, writing, and getting things done in Codex. It is fast, much cheaper than Astra, and the one I keep coming back to. You will like Opus 5.5 if you want to give an agent a complex coding or visual project and see how far it can go. Its best results in our evaluations went further than Sol's, enough to draw part of the team back toward Claude. You can explore all our evaluations in more detail, including every real-world task and the scoring method, at the links below: Writing: https://checks.every.to/p/dans-editorial-checks Knowledge and real-world tasks: https://checks.every.to/p/mikes-checks Reading: http://checks.every.to/p/dans-reading-checks We will publish our full impressions at @every soon.
Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe. GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale. We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.
· 136K views
Previously, the GPT-5.6 Sol API cost $4 per 1M input tokens and $20 per 1M output tokens. GPT-6 Sol now costs $2 and $10, while cached input tokens cost $0.20. In the Codex CLI, enable the model with `codex --model gpt-6-sol`; in an open session, switch it with `/model`.
The ceiling remains higher. In hands-on evaluations, Every found Sol close to Astra in writing and computer use, but Opus 5.5 went further in long autonomous builds. Codex's new safety classifier repeatedly stopped work that had already been approved, making long runs harder for the team to leave unattended.
Every says it will publish the full results of its evaluations later.
