OpenAI Introduces GPT-6 Sol and Luna With Lower Costs and Frontier-Level Performance

OpenAI introduced GPT-6 Sol and Luna on September 22, expanding the company’s latest model generation beyond the earlier Astra release. Sol targets demanding reasoning, coding, and computer-use tasks, while Luna is designed for faster, lower-cost inference across everyday workloads. OpenAI described both models as sharing the GPT-6 generation’s core advances while offering different balances of capability, speed, and price.

The release marks a shift from presenting a single frontier model toward a portfolio built around workload economics. GPT-6 Astra remains the high-end model introduced earlier in September, while Sol and Luna bring parts of that capability to a broader set of users and developers. TechCrunch described the pair as updated smaller models built from the same technical foundation as Astra.

GPT-6 Sol is positioned as the stronger option for complex professional work, including software development, long-running computer tasks, and advanced reasoning. OpenAI says the model makes fewer mistakes while also reducing the number of tokens needed to complete coding tasks. Dataconomy reported OpenAI CEO Sam Altman’s claim that Sol is 54% more token-efficient for AI coding than earlier models.

That efficiency matters because agentic systems pay for every step of a task. A model that reaches the same result with fewer generated tokens lowers both direct API costs and the time spent waiting for tool calls. OpenAI’s framing places Sol between Astra’s maximum capability and the cheaper Luna tier, giving teams a model for difficult work without assigning every request to the most expensive system.

GPT-6 Luna is the speed and price-focused member of the launch, aimed at high-volume requests, routine coding, and tasks where response time matters more than maximum reasoning depth. The published pricing is $1 per 1 million input tokens and $6 per 1 million output tokens, according to Simon Willison’s analysis of OpenAI’s announcement.

Those prices put Luna directly into the cost-performance contest with Anthropic and other model providers. A comparison published by Trilogy AI says Luna delivered roughly 24 DeepSWE benchmark points per estimated API dollar, compared with 4.5 for Claude Opus 4.8 and 3.2 for Claude Fable 5. The comparison depends on reported benchmark peaks and estimated usage, so it measures a specific coding scenario rather than overall model quality.

OpenAI’s release makes model choice a budgeting decision as much as a capability decision. Sol is priced at $5 per 1 million input tokens and $30 per 1 million output tokens, while Luna costs $1 and $6 respectively. The gap gives developers a reason to route simple requests to Luna and reserve Sol for tasks that benefit from deeper reasoning.

That structure reflects a broader industry move toward tiered frontier models. Ars Technica characterized the current competition as a comparison-shopping phase, with OpenAI and Anthropic both promising more capability for less money. Price per token remains an imperfect measure because reasoning effort, output length, caching, tool use, and task success all affect the final bill. Still, the lower entry price expands the range of applications that can run continuously rather than only in limited trials.

Developers choosing between GPT-6 Sol and Luna will need to test complete workflows rather than rely on headline benchmarks. Luna’s advantage is strongest when requests are numerous, latency is important, and failures are cheap to retry. Sol is better suited to coding agents, multi-step research, and computer-use tasks where one incorrect decision can erase the savings from a lower per-token rate.

The release also raises operational questions around routing, monitoring, and fallback behavior. Teams can use Luna for classification, drafting, and routine automation, then escalate difficult cases to Sol or Astra. OpenAI’s claims about reduced mistakes and token efficiency make that strategy attractive, but production evaluations still need to measure accuracy, tool-call reliability, latency, and total cost for each organization’s workload.

GPT-6 Sol and Luna extend OpenAI’s effort to make frontier intelligence available across more price points. Sol brings the GPT-6 generation’s higher-end reasoning to professional workflows, while Luna attacks the cost and speed barriers that limit large-scale deployment. Together, they give OpenAI a broader answer to competitors offering their own cheaper high-capability models.

The immediate test will be adoption, not launch-day benchmarks. Enterprises will compare completed task costs, developers will measure agent reliability, and consumers will notice only if the models improve everyday work without adding friction. The lineup also gives OpenAI room to tune products around different budgets as model inference becomes a central competitive lever.