“GPT-6 Astra introduced a new generation of intelligence—these models help distribute the benefits of that intelligence by advancing the frontier on cost efficiency,” OpenAI said.
Pricing is one of the biggest changes. GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 for GPT-5.6 Sol. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, compared with $0.20 and $1.20 previously.
OpenAI attributes those reductions to improvements in inference and caching rather than a simple reduction in model capability. The company is also expanding prompt caching across GPT-6, including discounts of 90% on cached input reads.
GitHub told OpenAI that caching improvements over recent months cut the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models.
The company is also emphasizing factual reliability. OpenAI says GPT-6 Sol produces about half as many mistakes as GPT-5.6 Sol on an internal evaluation built from de-identified ChatGPT conversations where users had flagged factual errors.
“On our internal factuality evaluation, which is based on de-identified real-world conversations where users flagged mistakes by our models, GPT-6 Sol makes about half as many mistakes as its predecessor, approaching Astra-level reliability at much lower cost,” OpenAI said.
OpenAI notes that the test focuses on conversations already associated with mistakes and is not representative of typical usage, where factual errors are less common.
The performance gains extend into coding. On FrontierCode, GPT-6 Sol improved substantially over GPT-5.6 Sol and matched Claude Fable 5.1 at a lower reported cost. On DeepSWE 1.1, Sol scored 68.8% at maximum effort, compared with Claude Fable 5’s 69.9% at xhigh effort, while OpenAI says Sol cost about 80% less per task.
GPT-6 Luna also posted stronger coding results than its predecessor. At maximum effort, Luna scored 66.6% on DeepSWE, which OpenAI says was comparable to Claude Opus 5 and Fable 5 at medium effort while costing significantly less per task.
For broader professional workflows, GPT-6 Sol scored 33.2% on AutomationBench at xhigh effort with a reported cost of $0.27 per task. OpenAI says that result exceeded Claude Opus 5 at max effort and Claude Fable 5.1 with fallback, while also beating GPT-6 Astra at low effort.
On Agents’ Last Exam, Sol reached 56.4% at max effort. OpenAI said that result came at 60% lower cost per task than Claude Opus 5’s highest score in the evaluation.
Computer-use performance also improved. On the offline portion of OSWorld 2.0, GPT-6 Sol scored 60.5% at xhigh effort, compared with 60.3% for Claude Opus 5 at medium effort. OpenAI says Sol achieved that result at about 80% lower cost per task.
The company is also bringing Astra’s communication changes to the two models. OpenAI says Sol and Luna should produce clearer, slightly shorter responses with less jargon, fewer unusual turns of phrase and less unnecessary detail.
Those changes are particularly targeted at technical and coding conversations, where OpenAI says the models should be more explicit about what they checked and less likely to jump to conclusions before completing a task.
Alignment is another area where OpenAI says the new models improve on their GPT-5.6 counterparts. Sol and Luna build on the alignment work introduced with Astra and showed lower rates of misleading claims about coding work in the company’s evaluations.
The release also arrives the same day Anthropic introduced Claude Opus 5.5, roughly 90 minutes before OpenAI’s announcement. That puts both companies’ latest models into direct comparison on cost, coding performance, factuality and professional work.
OpenAI makes several comparisons with Anthropic models in its own benchmark results, but those figures come from OpenAI’s chosen evaluations and configurations. Independent testing will be needed to determine how the models compare across broader real-world workloads.
GPT-6 Sol and GPT-6 Luna are available starting today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access GPT-6 Luna in the desktop app.
The models are also available through the OpenAI API as gpt-6-sol and gpt-6-luna. OpenAI said availability in ChatGPT will expand gradually throughout the day.
This analysis is based on reporting from OpenAI.
Image courtesy of OpenAI.
This article was generated with AI assistance and reviewed for accuracy and quality.
