Frontier AI models have usually come with frontier prices. OpenAI’s latest release tries to break that pattern: GPT-6.1 Sol, launched on September 29, 2026 at DevDay, is a mid-tier model that the company says gets close to its flagship on the tasks developers care about most, at a fraction of the cost.
What OpenAI released
GPT-6.1 Sol is an upgrade to GPT-6 Sol, the middle model in OpenAI’s GPT-6 family. According to OpenAI’s announcement, it targets agentic coding, computer use and professional work. It is available in the OpenAI API as gpt-6.1-sol, and in ChatGPT Work and Codex.
The GPT-6 lineup now has three tiers, and the price gap between them is wide. Prices below are per million tokens, listed as input, output and cached input.
GPT-6 Astra, the flagship, costs $10 for input, $50 for output and $1 for cached input. GPT-6.1 Sol costs $2 for input, $10 for output and $0.10 for cached input. GPT-6 Luna, the smallest tier, costs $0.10 for input, $0.50 for output and $0.01 for cached input.
Where the “one-fifth” comes from
The headline claim is that Sol costs one-fifth of Astra’s standard token prices, and the numbers above show it: $2 against $10 for input, $10 against $50 for output.
It is worth being precise about what that means. Sol’s standard input and output prices are the same as those of GPT-6 Sol, the model it replaces. The saving is relative to the flagship, not to the previous Sol. The one price that did fall is cached input, which dropped from $0.20 to $0.10 per million tokens, so cached reads now cost 5% of the normal input rate instead of 10%.
That change matters most for AI agents. They resend the same system prompt, tool definitions and conversation history at every step, so a cheaper cache lowers the cost of long-running tasks more than the headline rate suggests.
What the benchmarks say
All numbers below are reported by OpenAI in its launch post, and OpenAI says competitor figures come from public reports. They have not been independently reproduced.
- Coding: On DeepSWE v1.1, Sol matches Astra at roughly a fifth of the cost. It also beats GPT-6 Sol’s best score by 6.4 percentage points while using a lower reasoning effort.
- Computer use: On the offline set of OSWorld 2.0, Sol gains 7 points over GPT-6 Sol at maximum effort and lands within 2.1 points of Astra, at around one-seventh of the cost per task.
- Professional work: OpenAI reports that Sol beats Anthropic’s Claude Opus 5.5 on a benchmark of answers over complex professional PDFs, at less than half the cost per task, and edges ahead on AutomationBench at about a third of the cost.
- Science: On Terminal-Bench Science 0.1, Sol more than doubles GPT-6 Sol’s score at maximum effort, with an average cost of $5.47 per task against $23.80 for Astra. Astra still leads at 68.1%, and OpenAI itself recommends it for the hardest research.
- Factuality: At low effort, the share of responses containing a factual error falls from 11.4% to 7.7%, on conversations built from cases where users had flagged earlier mistakes.
Technical details for developers
According to the model page, the main specifications are:
- A context window of about 1.05 million tokens and up to 128,000 output tokens.
- A knowledge cutoff of April 30, 2026.
- Text and image input, with text output.
- Five reasoning-effort levels, from low to max, with medium as the default.
- Tool calling through the Responses API.
- Batch and Flex processing at half the standard price, and a fast mode at double.
- US and EU data residency, though fast mode is not available with EU residency.
- No fine-tuning.
One pricing detail deserves attention: prompts above 272,000 input tokens are billed at twice the input and cache rates and 1.5 times the output rate, for the whole request. Teams that rely on very long contexts will pay noticeably more than the headline numbers suggest.
OpenAI also announced a GPT-6.1 Sol Ultrafast option for Codex, promising up to eight times faster token generation. The company said it would arrive within days of launch and has not published its price.
How to read the claim
A few caveats apply before treating Sol as a drop-in replacement for the flagship:
- The benchmarks come from OpenAI. Independent testing will show whether the results hold across real workloads.
- List price is not total cost. What a task costs depends on how many tokens a model uses to complete it, and different vendors count tokens differently. The per-task figures OpenAI cites are more informative than per-token prices, but they are still the company’s own numbers.
- Astra keeps the top end. For the hardest scientific and research tasks, the flagship still scores higher.
- Sol is closed and hosted only. It can’t be self-hosted and has no fine-tuning, which rules it out for some use cases.
Daily AI roundups also report that a separate GPT-6.1 Astra release was cancelled. We could not find an explanation from OpenAI in the sources we reviewed, so we treat it as an unconfirmed detail.
Why it matters
The pattern this release points to is a narrowing gap between a company’s mid-tier and its top model. If Sol holds up in practice, many teams building agents, coding tools and document workflows will have little reason to pay Astra prices for everyday work, and may reserve the flagship for a small set of difficult cases.
It also puts pressure on rivals. At $2 for input and $10 for output, Sol sits at the same list price as comparable mid-tier models from other labs, such as Anthropic’s Claude Sonnet 5.5, but with half the cached-input rate. For developers, the practical advice is to test it on their own tasks, compare cost per completed task rather than per token, and then decide where each tier belongs.
