OpenAI halves API prices with two cheaper GPT-6 models

OpenAI introduced two new models on Wednesday, GPT-6 Sol paired with a second release, GPT-6 Luna. Both are pitched as cheaper, faster options sitting beneath GPT-6 Astra, the flagship system OpenAI put out earlier this month. The company says both were built using training methods similar to Astra’s, carrying that model’s gains in coding, computer use and factual accuracy down into the lower-cost versions.

The concrete news is price. OpenAI is cutting API rates by half against the promotional pricing it charged for GPT-5.6, though the discount lands differently across the two models. Sol’s input token cost is now $2 per million, down from $4, and its output rate fell to $10 per million from $20. Luna sits well below both: input tokens run $0.10 per million rather than $0.20, and output tokens cost $0.50 per million versus $1.20 before. OpenAI attributes the cut to caching and inference improvements rather than new hardware, which matters for any team modeling inference spend into next year’s budget.

OpenAI backs the release with a run of benchmark comparisons, and every one of them is the company grading its own models against public scores from competitors, not an independent test. Running at its highest reasoning setting, Sol posted a 33.2 percent score on AutomationBench 1.0.6, a business-workflow evaluation covering 47 tools, at a cost of $0.27 a task. That beats the 26.9 percent OpenAI attributes to Anthropic’s Claude Opus 5 running at maximum effort, at less than a tenth of what Opus 5 costs per task, and it also beats GPT-6 Astra’s own low-effort score of 30.3 percent.

The comparison OpenAI draws against Claude Fable 5.1, Anthropic’s most capable model, comes with a caveat OpenAI states outright. Fable 5.1 scored 31.4 percent on the same test when paired with a fallback to Opus 5, but that fallback triggered on roughly 40 percent of tasks, and its added cost is not reflected in the number OpenAI is citing. The one head-to-head comparison against Anthropic’s top model is, by OpenAI’s own account, missing part of the bill.

On Agents’ Last Exam, Sol’s highest-effort run posted 56.4 percent, a result OpenAI says beats Opus 5’s best score on that test while Sol’s task costs run well below Opus 5’s. On the coding benchmark DeepSWE, Sol came within about a point of Claude Fable 5’s top score while OpenAI says it did so at roughly 80 percent less cost, and Luna scored close behind at what OpenAI describes as a fraction of Opus 5’s and Fable 5’s per-task price. None of these figures carry outside verification, and OpenAI does not disclose how it selected the competitor scores it is quoting.

OpenAI also points to an internal review of flagged conversations where users identified a model mistake, saying Sol cuts that error rate roughly in half versus its predecessor and Luna, run at higher effort, matches an older Sol model’s accuracy at a small fraction of its cost. The company says both new models carry over Astra’s plainer writing style and describes improvements to alignment behavior, pointing readers to its published system card for the underlying evaluations.

Both models are live today in ChatGPT Work and in Codex for paid tiers, with Luna also reaching free users in the desktop app; neither has reached ChatGPT’s consumer tier yet. In the API, Sol appears as gpt-6-sol while Luna is listed separately as gpt-6-luna, with the rollout continuing through the day.

For teams pricing out agentic workloads against OpenAI and Anthropic side by side, the true gap on OpenAI’s own AutomationBench numbers is smaller than the headline comparison suggests once Fable 5.1’s uncounted fallback cost gets added back in, which makes a live cost-per-task test worth running before switching providers on this comparison alone.

OpenAI detailed the release, including all cited benchmark figures, in its own blog post “Introducing GPT-6 Sol and Luna,” published September 23, 2026.