Glonce

OpenAI launches GPT-6 Sol and Luna, cuts API prices 50%

modelsThe Glonce editorsHacker News🔥84Glonce rating84Impact75Novelty65Relevance100Surprise50Players95Credibility88
OpenAI launches GPT-6 Sol and Luna, cuts API prices 50%

OpenAI has expanded its GPT-6 lineup with two new models, GPT-6 Sol and GPT-6 Luna, joining GPT-6 Astra, which OpenAI introduced earlier the same month as its most intelligent and aligned model. OpenAI says Sol and Luna were trained with methods similar to Astra's, carrying over its advances in professional work, factuality, coding, computer use and alignment into faster, cheaper models, and it is cutting API prices for Sol and Luna by 50% compared with their GPT-5.6 promotional pricing.

On AutomationBench, a test of business workflows across apps, GPT-6 Sol at xhigh effort outperforms Claude Opus 5 at max effort while costing just 9% of Opus 5's price per task; GPT-6 Luna at high effort improves on its own predecessor by 5.4 percentage points at 58% lower cost per task. OpenAI also says Sol beats Claude Fable 5.1 at far lower cost and even outperforms Astra itself running at low effort. On Agents' Last Exam, a test of complex professional workflows, GPT-6 Sol at max effort scores 56.4%, above Claude Opus 5's highest score in the evaluation, at 60% lower cost per task.

On OpenAI's internal factuality evaluation, based on de-identified real conversations where users flagged model mistakes, GPT-6 Sol makes about half as many mistakes as its predecessor, approaching Astra-level reliability at much lower cost; GPT-6 Luna also improves substantially, matching GPT-5.6 Sol's factuality at higher effort levels for roughly a hundredth of the cost. OpenAI frames part of the release around its own coding-agent usage: valued at API prices, daily token usage inside the company has exceeded $600 for the median researcher and $7,000 for researchers at the 90th percentile, the cost pressure Sol and Luna's pricing is meant to ease.

On FrontierCode, which checks whether an agent's code changes are ready to merge into real codebases, GPT-6 Sol improves substantially over GPT-5.6 Sol and matches Claude Fable 5.1 at xhigh effort while costing much less. On DeepSWE v1.1, a complex software-engineering benchmark, GPT-6 Sol at max effort scores 68.8%, within 1.1 percentage points of Claude Fable 5's highest score in the evaluation (69.9% at xhigh effort) at about 80% lower cost per task; GPT-6 Luna at max effort scores 66.6%, comparable to Claude Opus 5 and Fable 5 at medium effort, while costing 93% less per task than Opus 5 and 96% less than Fable 5. On computer use, OpenAI says Astra remains the world's best model, but on OSWorld 2.0 offline GPT-6 Sol at xhigh effort scores 60.5% versus Claude Opus 5's 60.3% at medium effort, at about 80% lower cost per task, and GPT-6 Luna at max effort exceeds GPT-5.6 Sol at medium effort while costing one tenth as much.

OpenAI also says it improved prompt caching for GPT-6, raising cache hit rates by default and giving a 90% discount on cached input-token reads, alongside a caching dashboard, a diagnostics tool for missed cache opportunities and explicit breakpoints for where cached prefixes end; reasoning effort and enabled tools can now be changed without breaking the cache. GitHub is cited as reporting that these changes have cut the share of prompt tokens needing fresh processing by more than 50% across billions of Copilot requests to OpenAI models. Sol and Luna also carry over alignment improvements from Astra, including lower rates of misleading claims about their own coding work, and Astra's more concise communication style. Both models are live today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with Free and Go users limited to Luna in the desktop app; neither is yet available in plain ChatGPT Chat, and both are addressable in the API as gpt-6-sol and gpt-6-luna. OpenAI notes that its evaluations deliberately test challenging situations rather than typical-use failure rates, that competitor scores were taken from public reports rather than run by OpenAI itself, and that it substituted Claude Fable 5 scores wherever Fable 5.1 scores were not available.

Key facts

  • API prices for GPT-6 Sol and Luna are cut 50% compared with their GPT-5.6 promotional pricing.
  • On AutomationBench, GPT-6 Sol at xhigh effort beats Claude Opus 5 at max effort while costing just 9% of Opus 5's price per task; GPT-6 Luna at high effort gains 5.4 percentage points over its predecessor at 58% lower cost.
  • On DeepSWE v1.1, GPT-6 Sol at max effort scores 68.8%, within 1.1 percentage points of Claude Fable 5's best score (69.9% at xhigh effort) at about 80% lower cost per task; GPT-6 Luna scores 66.6%, at 93% less cost than Opus 5 and 96% less than Fable 5.
  • On OpenAI's internal factuality evaluation, GPT-6 Sol makes about half as many mistakes as its predecessor, and GPT-6 Luna at higher effort matches GPT-5.6 Sol's factuality at roughly a hundredth of the cost.
  • GPT-6 Sol and Luna go live today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users (Free and Go get Luna in the desktop app only), with API access under the model ids gpt-6-sol and gpt-6-luna.

Why it matters

OpenAI is splitting GPT-6 into a cost tier rather than shipping one flagship. Earlier this month it released GPT-6 Astra as its most capable and aligned model, reserved for the most demanding work. Sol and Luna carry over Astra's training methods and its gains in professional work, factuality, coding, computer use and alignment, but aimed at cheaper, faster everyday use, with API prices for both cut 50% against their GPT-5.6 promotional rates. For computer use specifically, OpenAI says Astra stays the world's best model; Sol and Luna instead compete on cost, matching or beating older-generation rivals at a fraction of the price.

Who it affects

Developers on the OpenAI API get two new model ids, gpt-6-sol and gpt-6-luna, at half the promotional price of their GPT-5.6 predecessors, plus better prompt-caching discounts. ChatGPT Work and Codex subscribers on Plus, Pro, Business, Enterprise and Edu plans get both models starting today; Free and Go users get GPT-6 Luna only, through the desktop app; neither model is yet in plain ChatGPT Chat. Coding-agent-heavy teams are a direct target: OpenAI says its own median researcher now runs about $600 a day in coding-agent token usage at API prices, and its 90th-percentile researcher about $7,000 a day. Every comparison in the announcement is against Anthropic's Claude Opus 5 and Claude Fable 5 or 5.1, positioning the release directly against Anthropic's current lineup.

How to use it

GPT-6 Sol and GPT-6 Luna are live today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu accounts; Free and Go users can reach Luna only through the desktop app. Neither model is yet available in ChatGPT Chat. In the API they are addressable as gpt-6-sol and gpt-6-luna, at 50% lower per-token prices than Sol and Luna's GPT-5.6 promotional pricing, and both carry Astra's more concise communication style, with less jargon and shorter answers. OpenAI has also improved prompt caching for GPT-6: a 90% discount on cached input-token reads, a caching dashboard, a diagnostics tool for missed cache opportunities, and explicit breakpoints for where cached prefixes end, with reasoning effort and enabled tools now adjustable without breaking the cache.

How solid is it

Every figure comes from OpenAI's own announcement: the AutomationBench, Agents' Last Exam, FrontierCode, DeepSWE v1.1 and OSWorld 2.0 offline scores, and the internal factuality and alignment evaluations, are all run and reported by OpenAI itself. The Claude Opus 5 and Fable 5 or 5.1 numbers used as comparators, by OpenAI's own account, were taken from publicly available reports rather than tested by OpenAI directly, and Claude Fable 5 scores stand in wherever Fable 5.1 scores were not available. No independent benchmark and no comment from Anthropic is included.

Risks and caveats

OpenAI states directly that its evaluations deliberately test challenging situations and do not measure failure rates in typical use, so the improvement figures should not be read as everyday error rates. The competitor numbers are secondhand, drawn from public reports rather than OpenAI's own testing of Claude models, and substituting Fable 5 for missing Fable 5.1 scores means some comparisons use an older Anthropic model than the one actually named. The announcement gives no exact calendar date for the release or for Astra's earlier launch, no per-token dollar prices for Sol or Luna, and no parameter counts or architectural detail for any of the three models; it also does not say whether Astra's own pricing changed alongside Sol and Luna's cut.