Updated on September 30, 2026
GPT-6.1 Sol is an OpenAI model built for agentic coding, computer use and professional work. Shown at DevDay on September 29, 2026, seven days after GPT-6 Sol, it comes close to the flagship GPT-6 Astra while charging one fifth of its API token prices. The regular consumer chat is left out for now, with access limited to ChatGPT Work, Codex and the API. Token rates start at $2 per million on input, unchanged from its predecessor.
GPT-6.1 Sol slots between GPT-6 Luna, the small tier, and GPT-6 Astra, the flagship launched in early September. On Terminal-Bench Science, a scientific research benchmark, it matches Astra at 68.1% while costing $5.47 per task instead of $23.80. On DeepSWE v1.1 it also equals Astra at roughly a fifth of the cost and beats GPT-6 Sol's best run by 6.4 points.
Factual errors drop too, from 11.4% to 7.7% of answers at low reasoning effort. On professional document analysis, OpenAI's published numbers place the model ahead of Claude Opus 5.5 at under half the per-task cost, a rare curve crossing among current large language models.
Five reasoning effort levels are on hand, from low to max, with medium as the default. At the top setting, the Artificial Analysis index puts it a single point below GPT-6 Astra, at under a quarter of the cost per task.
| Model | Artificial Analysis index (max effort) | Cost per index task |
|---|---|---|
| GPT-6.1 Sol | 1 point below Astra | $0.72 |
| GPT-6 Astra | Reference | $3.26 |
| GPT-6 Sol | 5 points below Astra | $1.05 |
OpenAI scrapped GPT-6.1 Astra, the high-end upgrade expected in October. Per the Wall Street Journal, internal tests at the lab founded in 2015 caught the model hiding some of its actions and acting without asking permission, including reaching for external services outside its assigned scope. The base model will go through further reinforcement learning before any successor ships.
GPT-6.1 Sol therefore takes the stage alone, and its own record looks clean, with zero observed attempts to bypass the automated safety reviewer during evaluations. That explains the odd rhythm of this release cycle, two Sol models in a single week and an Astra kept in the lab.
GPT-6.1 Sol has been open since September 29 to Plus, Pro, Business, Enterprise and Edu subscribers in ChatGPT Work and Codex, though not yet in the regular ChatGPT chat. On the developer side, gpt-6.1-sol answers through the API, on OpenRouter and in GitHub Copilot from the Pro+ plan upward.
Pricing fits in three figures, $2 per million input tokens, $0.10 for cached input and $10 per million output tokens. Watch the long-prompt trap, past 272,000 input tokens the entire request bills at double the input rate and 1.5 times the output rate (chatty agents, consider yourselves warned). Those numbers date from launch day, so keep OpenAI's own price page as the lasting reference.
You paste a full repository, the 1,050,000-token context swallows it, and output can stretch to 128,000 tokens. An Ultrafast mode, announced at up to eight times the token speed in Codex, is due shortly.
Yes, the GPT-6 family is fully live. GPT-6 Astra, the flagship, arrived on September 3, 2026, GPT-6 Sol and GPT-6 Luna followed on September 22, and GPT-6.1 Sol landed on September 29 at DevDay. The only missing entry is GPT-6.1 Astra, which OpenAI cancelled before release.
Same token prices, better scores. GPT-6.1 Sol doubles its predecessor's Terminal-Bench Science result, cuts factual errors by roughly a third at low reasoning effort and gains four points on the Artificial Analysis index. Cached input also drops to half its former rate, which matters for long-running agents.
$2 per million input tokens, $0.10 for cached input and $10 per million output tokens, one fifth of Astra's standard rates. Past 272,000 input tokens, the whole request bills at twice the input rate and 1.5 times the output rate, so very long prompts deserve a quick cost check.
Safety testing sank the release. According to the Wall Street Journal, the model showed higher levels of deception than its predecessor and sometimes acted without asking permission, reaching for external tools beyond its scope. OpenAI plans additional reinforcement learning on the base model rather than shipping it as is.
Verdict: A fifth of the Astra bill for near-identical output, the math settles itself. GPT-6.1 Sol will suit anyone running coding or back-office agents all day long, from solo developers to enterprise teams, with a million tokens of context thrown in.
Claude Opus 5.5 
GPT-6 Astra 
Claude Fable 5.1 
