OpenAI launches GPT-6 Sol and Luna with sharply lower API prices

Why it matters

OpenAI is pushing GPT-6 into high-volume agent workloads where token costs, caching and model routing decide whether products can operate profitably.

OpenAI launched GPT-6 Sol and GPT-6 Luna on September 22nd, extending its newest model family into the lower-cost coding and agent workloads that account for sustained, high-volume usage.

https://x.com/ChatGPT/status/2102462020339613966

The models are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go subscribers can access Luna through the ChatGPT desktop app. OpenAI said neither model is available in the regular Chat interface at launch.

The release gives OpenAI a broader price ladder beneath GPT-6 Astra, its highest-capability model. Sol is positioned for complex coding and agentic workflows, while Luna is designed for focused tasks that need to run cheaply at high volume. That division matters for developers building agents: the strongest available model can be excessive when a workflow involves hundreds of routine tool calls, tests or document operations.

OpenAI cuts the cost of its workhorse models

In the API, GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. Those prices are half the promotional rates for GPT-5.6 Sol.

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. Its predecessor was priced at $0.20 and $1.20, respectively, meaning the output-token reduction is greater than the 50% headline discount OpenAI used to describe the release.

Both models support reasoning settings from none through max, a 1.05 million-token context window and up to 128,000 output tokens. OpenAI's documentation lists support for web search, file search, image generation, code execution, computer use, MCP connections and other tools through the Responses API. The callable model IDs are gpt-6-sol and gpt-6-luna.

The pricing makes Luna the more consequential product for applications where inference volume determines whether a feature can be offered profitably. At $0.10 per million input tokens, developers can route classification, extraction and tightly specified agent steps to Luna while reserving Sol or Astra for decisions that require deeper reasoning. OpenAI is effectively selling GPT-6 as a model portfolio rather than a single default intelligence tier.

The benchmark pitch is capability per dollar

In its launch announcement, OpenAI emphasized cost per completed task rather than raw benchmark leadership.

OpenAI said GPT-6 Sol at its xhigh reasoning setting scored 33.2% on AutomationBench, compared with 26.9% for Claude Opus 5 at max effort. OpenAI calculated Sol's cost at $0.27 per task, or about 9% of the Claude model's cost under the tested configurations. The comparison depends on OpenAI's chosen reasoning settings, API prices and evaluation harness, so it should be read as a vendor benchmark rather than an independent cost study.

On DeepSWE v1.1, OpenAI reported a 68.8% score for Sol at max effort, 1.1 percentage points behind the result it cited for Claude Fable 5 at xhigh effort. OpenAI said Sol cost roughly 80% less per task in that comparison. Luna scored 66.6% at max effort, according to OpenAI, with materially lower estimated costs than the Claude models included in the test.

OpenAI also claims Sol produces about half as many factual errors as GPT-5.6 Sol on an internal evaluation built from conversations where users flagged mistakes. The test uses error-inducing conversations rather than a representative sample of ordinary ChatGPT traffic, a limitation OpenAI acknowledges in the announcement.

Cheaper context is part of the release

OpenAI paired the model launch with changes intended to improve prompt caching for agents and long-running conversations. Cached input reads receive a 90% discount, while developers can alter reasoning effort or tool availability without automatically invalidating earlier cached context.

That change attacks another large cost inside agent products. Long-running systems repeatedly send instructions, files, conversation history and tool definitions back to the model. Better cache reuse can reduce the share of those tokens that require fresh processing. OpenAI cited GitHub data showing that caching improvements cut freshly processed prompt tokens by more than 50% across billions of requests, though the announcement does not break out how much of that gain comes specifically from the new models.

The rollout is aimed first at work surfaces where usage and cost are easiest to measure: coding, document production, computer use and multi-step agents. OpenAI is keeping Astra at the top of the family while making Sol and Luna cheap enough to serve as routine infrastructure. For developers, the launch turns model selection into an economic decision made step by step inside a workflow, rather than a single choice made for an entire application.