OpenAI Launches GPT-6 Sol and GPT-6 Luna, Cutting API Prices 50% Versus GPT-5.6

OpenAI released two new models, GPT-6 Sol and GPT-6 Luna, on September 22, 2026, expanding the GPT-6 lineup it began earlier this month with GPT-6 Astra. Both new models are priced 50% lower than their GPT-5.6 predecessors, according to OpenAI, and are rolling out today to ChatGPT Work and Codex.

What's new

GPT-6 Sol targets complex coding and agentic workflows, positioned as the large model in the lineup. GPT-6 Luna is built for focused, high-volume tasks and sits at the smaller end of the tier structure. Notably absent is a mid-sized "GPT-6 Terra" model — OpenAI appears to be consolidating the GPT-6 family into three tiers: Astra (extra large), Sol (large), and Luna (small), dropping the medium tier that existed in the GPT-5.6 generation.

OpenAI says both models were trained using methods similar to GPT-6 Astra, carrying over what the company describes as Astra's state-of-the-art performance in professional work, factuality, coding, computer use, and alignment — at lower cost and with fewer tokens consumed per task, according to OpenAI's release notes in the ChatGPT desktop app.

Benchmark claims

On OSWorld 2.0 offline, a computer-use benchmark, OpenAI claims GPT-6 Sol running at "xhigh" effort scores 60.5%, edging out Claude Opus 5 at medium effort (60.3%) — at approximately 80% lower cost per task, according to OpenAI. The company also claims GPT-6 Luna at maximum effort exceeds GPT-5.6 Sol at medium effort while costing one-tenth as much. These figures come directly from OpenAI and have not been independently verified.

GPT-6 Astra remains OpenAI's top-performing model for computer use, the company says; Sol and Luna are positioned as more cost-efficient alternatives rather than outright replacements at the high end.

Pricing and availability

OpenAI has not disclosed specific per-token dollar pricing for Sol or Luna, but confirmed both are priced 50% below the GPT-5.6 promotional rates they replace. OpenAI attributes the cuts to improvements in caching and inference infrastructure that lower serving costs.

Both models are available today in ChatGPT Work and Codex — but not in standard ChatGPT Chat — for Plus, Pro, Business, Enterprise, and Edu users. Free and Go tier users can access Luna only through the desktop app. Enterprise administrators must manually enable the new models for their workspaces, and rollout timing varies by account.

OpenAI's announcement also cited internal token usage data: median researchers at the company now consume a daily median of $600 in API-equivalent tokens on coding agents, with the 90th percentile exceeding $7,000 per day, underscoring the cost pressure driving this release.

What this means

This release is about economics, not a capability leap. OpenAI is acknowledging that as coding agents and computer-use tasks run longer and consume more tokens, sustained cost becomes the binding constraint on adoption — not raw intelligence. Cutting prices 50% while claiming near-parity with Astra-tier performance is a direct answer to that pressure, and a competitive response to Anthropic's Claude Opus 5, which OpenAI explicitly benchmarks against on OSWorld 2.0. Dropping the mid-tier Terra model in favor of a three-tier structure (Astra/Sol/Luna) also simplifies OpenAI's product lineup, mirroring how GPT-5.6 was structured. Whether the claimed 80% cost reduction per task and 10x cost efficiency gains hold up under independent testing remains to be seen — these are OpenAI's own figures, not third-party validated results.