OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September, expanding its newest model generation downward into the cheaper tiers and cutting API prices by half or more in the process. Coming just 19 days after GPT-6 Astra, the release says less about raw capability ceilings than about where frontier competition has actually moved: cost per completed task, not peak benchmark score.
The headline numbers are straightforward. For standard short-context workloads, GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 for GPT-5.6 Sol. GPT-6 Luna runs at $0.10 and $0.50, against $0.20 and $1.20 for its predecessor. Long-context workflows are priced separately, at $4/$15 for Sol and $0.20/$0.75 for Luna.
Two Models, Two Jobs
The split between the models is a positioning decision rather than an architectural one. Sol is aimed at complex coding and agentic workflows — long chains of tool calls where reasoning quality compounds across steps. Luna targets focused, high-volume tasks that need to run cheaply: classification, extraction, routing and the unglamorous middleware of production AI systems.
Both share the same envelope. Each supports reasoning settings ranging from none through max, a 1.05 million-token context window and up to 128,000 output tokens. Tool support covers web search, file search, image generation, code execution, computer use and MCP connections via the Responses API — a notable signal in itself, given how quickly the Model Context Protocol has become the default plumbing for agent integrations. The model identifiers are `gpt-6-sol` and `gpt-6-luna`.
Availability is deliberately staged. The models are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, while Free and Go subscribers can reach Luna through the ChatGPT desktop app. Neither model appears in the regular chat interface at launch — a clear indication that OpenAI sees this pair as infrastructure for builders first and a consumer product second.
The Benchmarks Tell a Nuanced Story
On AutomationBench, GPT-6 Sol at extra-high effort scored 33.2%, ahead of Claude Opus 5's 26.9%, at roughly $0.27 per task — about 9% of the cost per task of its rival. On DeepSWE 1.1, Sol reached 68.8%, trailing Claude Fable 5's 69.9% by just over a point while costing roughly 80% less per task. Luna scored 66.6% on the same benchmark, which is the more surprising result given its price point. OpenAI also says Sol makes about half as many mistakes as its predecessor on the company's internal factuality evaluation.
The caveat is equally important, and OpenAI's own charts make it visible: Sol scores slightly lower than GPT-5.6 Sol's top setting on two of six headline evaluations, while costing around 60% less per task. Read honestly, this is an efficiency release. The frontier of what these models can do has not moved much in three weeks; the frontier of what they cost to run has moved a great deal.
Read the Pricing Footnotes
OpenAI attributes the reduction to improvements in caching and inference rather than to a smaller model, and a company spokesperson has said that while GPT-5.6's pricing was promotional, the new GPT-6 rates are the default.
Buyers should still check the fine print. The 50% cut is measured against GPT-5.6 Sol's promotional rate of $4/$20, which OpenAI guarantees only at least through 21 November 2026, and the pricing page does not indicate what GPT-5.6 Sol will cost afterwards. Comparisons framed against a promotional baseline deserve the same scepticism in AI infrastructure as anywhere else.
Why It Matters
For anyone running agents in production, cost per task has quietly become the binding constraint. An agentic workflow that makes forty model calls to finish one job multiplies every per-token decision by an order of magnitude, which is why a 50% price cut changes what is economically viable far more than a one-point benchmark gain does.
Three practical implications follow:
- Tiering gets easier. With Luna at $0.10 per million input tokens and near-Sol performance on DeepSWE, routing the majority of steps to the cheap model and reserving Sol for genuinely hard reasoning becomes the obvious default architecture.
- Long context is now priced separately, so teams should measure how often their workloads actually cross the threshold rather than assuming a single blended rate.
- Migration is cheap but not free. Identical context windows and tool support lower the switching cost, but reasoning-effort settings change behaviour enough that evaluation suites should be re-run rather than assumed portable.
The broader read is that the model generation released this week is competing on unit economics. That is what a maturing platform market looks like — and it is generally good news for the companies building on top of it.
