OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, extending the GPT-6 family beyond the flagship Astra model. API prices for both models are 50 percent lower than GPT-5.6 promotional pricing, and both are live in the API, Codex and ChatGPT Work today.
Sol is positioned for complex coding and agentic workflows, Luna for focused, high-volume tasks. Both build on the training advances behind GPT-6 Astra, which OpenAI released earlier this month as its frontier model. The company describes the new pair as bringing most of Astra capability to faster, cheaper models rather than introducing new frontier capability of their own.
The release comes two weeks after Astra, an unusually short gap for OpenAI. The company has been under visible pricing pressure from Google, Anthropic and cheaper open-weight competitors, and the 50 percent cut reads as a direct answer to that pressure rather than a routine product refresh.
What the benchmarks claim
OpenAI put forward two cost-adjusted comparisons. On OSWorld 2.0, a computer-use benchmark, GPT-6 Sol at xhigh reasoning effort scored 60.5 percent, roughly matching Claude Opus 5 at medium effort at about 80 percent lower cost per task. GPT-6 Luna at maximum effort exceeded GPT-5.6 Sol at medium effort while costing about a tenth as much.
Both numbers follow a familiar pattern: capability parity with a rival flagship, framed through cost rather than raw performance. Anthropic has made the same argument in reverse with its own pricing updates. The contest between the labs has shifted from who tops the leaderboard to who delivers acceptable capability per dollar, and this release is OpenAI playing that game deliberately rather than defensively.
Specification details matter for developers planning migrations. Sol carries a 1.05 million token context window with a 922,000 token input maximum and a 128,000 token output allowance. Both models default to medium reasoning effort, with settings ranging from none to max. The knowledge cutoff is April 2026, which puts both models reasonably current for fast-moving technical work.
Availability and the missing Terra
Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users starting today. Free and Go users get Luna in the desktop app. Neither model is selectable in standard ChatGPT conversations, where Astra remains available as GPT-6 Pro for paid plans. The split is deliberate: OpenAI wants consumer users on the hosted experience and developers on the models they can tune and scale.
One absence is notable. There is no GPT-6 Terra. The GPT-5.6 generation had Sol, Terra and Luna as its three-tier lineup, with Terra serving as the balanced middle option for everyday work. The public GPT-6 catalog currently lists only Astra, Sol and Luna. Developers who built on Terra as a cost-performance compromise have no direct replacement, and OpenAI has not said whether one is coming. Some will read the gap as deliberate upselling toward Sol; others will see a lineup that simply is not finished.
Enterprise administrators must enable the new models in workspace settings, which means adoption inside larger organizations will lag the general rollout by days or weeks depending on internal review processes and data governance checks.
The pricing war in context
The 50 percent cut follows a string of price moves across the industry this quarter. Google cut Gemini pricing at its developer conference. Anthropic dropped cached token prices by 75 percent on its Claude 5.1 line. StepFun, a Chinese lab, released a 600B-parameter model at $1 per million input tokens, roughly a seventh the price of GPT-5.6 Sol at the time, with open weights promised for October 15.
Inference cost is becoming the main competitive axis for API business. Model quality differences at the top end have narrowed to the point where most workloads run acceptably on several models, so the buying decision increasingly comes down to price, speed and context length. OpenAI cutting its own prices 50 percent within a single model generation is the clearest sign yet that the company would rather trade margin for volume than defend premium pricing against a widening field of competitors.
OpenAI also said it has made caching and inference more efficient, which funds part of the cut. The rest comes out of margin. With Anthropic reportedly pacing toward $100 billion in annualized revenue and a $2 trillion IPO expected in November, and Google pressing on distribution through Workspace and Android, OpenAI is choosing growth over near-term profitability in its developer business. Investors in the upcoming round will note the tradeoff.
What developers should watch
For teams running agent workloads, the practical question is token efficiency. OpenAI claims Sol and Luna use fewer tokens than their GPT-5.6 counterparts on the same tasks, which compounds with the lower per-token price. If the claim holds in production rather than in demos, effective cost falls by more than half for many workflows, and agent-heavy applications that were marginal at GPT-5.6 pricing become viable.
The second question is Codex adoption. Sol becomes the default model for complex coding in Codex, and early developer sentiment on the release thread has been positive on writing quality in particular. Independent benchmarks will take weeks to accumulate, but the early signal suggests the GPT-6 generation is competitive with Claude on coding at a lower price point, which is exactly the comparison enterprise buyers are making right now.
For now the message to the market is straightforward: the frontier stays at Astra, the volume business moves to Sol and Luna, and anyone waiting for AI API prices to stabilize should not hold their breath. The next data point arrives when Anthropic and Google respond, and neither is likely to wait long.