OpenAI launched GPT-6.1 Sol on Sep 29, 2026, as an upgrade to GPT-6 Sol, with access starting in ChatGPT Work, Codex and the API. It brings stronger results in OpenAI’s coding and agent evaluations at Sol’s existing standard input and output token prices. Cached input costs less than it did with GPT-6 Sol.
Where GPT-6.1 Sol runs
Plus, Pro, Business, Enterprise and Edu users can use GPT-6.1 Sol in ChatGPT Work and Codex. Developers can select gpt-6.1-sol in the API. It is not yet in Chat. Tool-using integrations need the Responses API; Chat Completions supports this model without tools.
OpenAI plans an Ultrafast Codex mode in the coming days, promising up to eight times standard mode’s token-generation speed.
What the four models cost
At standard API rates per million tokens, GPT-6.1 Sol keeps GPT-6 Sol’s fresh-input and output prices and halves its cached-read price. Astra and Opus 5.5 cost more than 6.1 Sol on all three measures.
| Model | Fresh input | Cached read | Output |
|---|---|---|---|
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 |
Cheaper cached reads matter when an agent reuses context. The 1.05 million-token context window has a cost catch: requests above 272,000 input tokens carry higher rates for the full request.
The workflow tradeoff
On OpenAI’s AutomationBench 1.0.6, GPT-6.1 Sol completes more workflows than earlier Sol at lower average model cost per task. At medium effort, Astra scores slightly higher but costs more; Opus 5.5 with fallbacks scores lower and also costs more. Zapier counts a workflow as correct only when every assertion passes. The graph shows both medium and maximum effort.
At medium effort, GPT-6.1 Sol produces about 30% more correct benchmark workflows per model-API dollar than GPT-6 Sol, 3.7 times as many as Opus 5.5 with fallbacks, and 6.2 times as many as Astra. The comparison uses reported average model costs and excludes tool fees and human review.
At maximum effort, Opus 5.5 with fallbacks leads GPT-6.1 Sol by 6.4 completion-rate points, but costs 4.8 times as much per task. Astra also scores higher and costs more. Peak success and correct workflows per API dollar are different priorities.
Where premium models still lead
The pattern is not universal across tasks. On OpenAI’s Terminal-Bench Science 0.1 result at maximum effort, Astra resolves 68.1% at $23.80 per task and Opus 5.5 with fallbacks resolves 63.3% at $23.21. GPT-6.1 Sol resolves 57.0% at $5.47.
Premium models buy a higher success ceiling on this science workload, while Sol resolves more tasks per reported model-API dollar.
For coding agents, OpenAI reports 75.2% on DeepSWE v1.1 for GPT-6.1 Sol at high effort and $0.65 per task, above GPT-6 Sol’s best shown 68.8% at $2.74 and Astra’s 74.1% at $4.43. Its published chart does not include Opus, so it cannot settle a four-way coding ranking. These evaluations were run in research or API environments; the results need not reproduce in production ChatGPT.
What do I think of this?
I see GPT-6.1 Sol as a useful upgrade for teams building coding and workflow agents on a Sol budget. OpenAI has improved the model’s benchmark results without raising standard input or output prices, and repeat use of cached context is cheaper. In the AutomationBench comparison, 6.1 Sol turns a fixed API budget into more correct completions than the other three models.
Opus 5.5 and Astra still earn their higher prices when the priority is the best result from a difficult attempt. Their maximum-effort workflow and science scores remain higher. My choice for repeated work on a budget would be 6.1 Sol; I would pay the premium when the higher success ceiling matters more than volume.