OpenAI launched GPT-6 Sol and GPT-6 Luna on September 22, 2026, expanding the GPT-6 family beyond the earlier GPT-6 Astra release.
The pitch is clear: Astra remains OpenAI's highest-capability model, but Sol and Luna bring some of the same GPT-6 improvements into cheaper, higher-volume tiers. Sol is aimed at complex coding and agentic workflows. Luna is the low-cost model for focused, repeatable tasks at scale.
Both models are available in the OpenAI API as gpt-6-sol and gpt-6-luna. OpenAI says they are also rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can try GPT-6 Luna in the desktop app, but OpenAI says the models are not yet available in regular Chat.
The price cut is the headline
OpenAI says GPT-6 Sol and Luna are 50% cheaper than the GPT-5.6 promotional prices they replace.
| Model move | Input price | Output price | OpenAI framing |
|---|---|---|---|
| GPT-5.6 Sol to GPT-6 Sol | $4 to $2 / 1M tokens | $20 to $10 / 1M tokens | 50% cheaper |
| GPT-5.6 Luna to GPT-6 Luna | $0.20 to $0.10 / 1M tokens | $1.20 to $0.50 / 1M tokens | 50% cheaper |
The API pricing page adds more detail. In standard short-context use, Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 per million input tokens and $0.50 per million output tokens. Cached inputs are 10% of the uncached input price, while cache writes are 1.25x the uncached input price.
For very large prompts, OpenAI applies long-context pricing above 272K input tokens. Under that tier, Sol rises to $4 input and $15 output per million tokens, while Luna rises to $0.20 input and $0.75 output.
That matters for coding agents because long sessions often reuse repo context, prior tool outputs, and large instruction prefixes. A model that is slightly weaker on a benchmark can still win in production if it lets teams run more attempts, more subagents, or broader validation for the same budget.
Sol versus Luna
OpenAI positions Sol as the serious work model and Luna as the volume model:
| Model | Official positioning | Standard price | Context | Max output |
|---|---|---|---|---|
| GPT-6 Sol | Complex coding and agentic workflows | $2 input / $10 output | 1.05M tokens | 128K tokens |
| GPT-6 Luna | Focused, high-volume tasks | $0.10 input / $0.50 output | 1.05M tokens | 128K tokens |
Both support reasoning effort from none through max, image input, Responses API tools, web search, file search, code interpreter, hosted shell, computer use, MCP, skills, and structured outputs. The catch: OpenAI's model guidance says Chat Completions function calling with Sol and Luna only works when reasoning_effort is set to none; developers should use the Responses API for reasoning with tools.
OpenAI's benchmark claims
OpenAI's launch post argues that Sol and Luna are strongest where cost per task matters:
| Area | OpenAI-reported signal |
|---|---|
| Business workflows | GPT-6 Sol at xhigh effort scores 33.2% on AutomationBench at $0.27 per task, above Claude Opus 5 max at 26.9% and cheaper by OpenAI's estimate. |
| Professional agents | GPT-6 Sol at max effort scores 56.4% on Agents' Last Exam, above Claude Opus 5's highest score in that evaluation at lower reported cost per task. |
| Coding | On FrontierCode, OpenAI says GPT-6 Sol improves substantially over GPT-5.6 Sol and matches Claude Fable 5.1 xhigh at much lower cost. |
| DeepSWE v1.1 | GPT-6 Sol at max scores 68.8%; GPT-6 Luna at max scores 66.6%. |
| Computer use | GPT-6 Sol xhigh reaches 60.5% partial reward on OSWorld 2.0 offline, close to Claude Opus 5 medium at 60.3% but with lower reported cost. |
The most interesting Luna claim is economic: OpenAI says Luna at max is comparable to stronger Claude models at medium effort on DeepSWE while costing far less per task. That is the shape of the release. Luna is not supposed to beat frontier models everywhere; it is supposed to make repeated agentic work cheap enough to run by default.
Independent benchmark read
Vals AI published launch-day evaluations for both models. Its numbers are less flattering than OpenAI's own framing, but still useful:
| Vals metric | GPT-6 Sol | GPT-6 Luna |
|---|---|---|
| Vals Index | 62.57%, #8 of 65 | 58.45%, #20 of 65 |
| Cost per Vals test | $7.56 | $0.42 |
| Vibe Code Bench | 87.82%, #6 of 103 | 81.65%, #15 of 103 |
| Code Migration | 57.20%, #4 of 68 | 42.55%, #21 of 68 |
| Terminal-Bench 2.1 | 83.15%, #6 of 73 | 73.03%, #22 of 73 |
Vals says Sol lands just behind GPT-5.6 Sol on its overall index and about four points behind GPT-6 Astra, but at less than half of Astra's cost per test. Luna is the more unusual story: it ranked only #20 overall, but Vals says it was the lowest-cost model in the top 20 and roughly 18x cheaper per test than Sol.
That makes the buyer's decision more nuanced. Sol looks like the safer pick for complex coding, migrations, and harder terminal tasks. Luna looks attractive for cheap sweeps, first-pass subagents, data cleanup, simple code edits, and cases where running several attempts beats paying for one expensive attempt.
Safety and alignment notes
OpenAI added a GPT-6 Sol and Luna appendix to the GPT-6 Astra system card. It says both models are treated as High capability in the Cybersecurity and Biological/Chemical domains under OpenAI's Preparedness Framework, but neither reaches the High threshold for AI self-improvement.
The appendix reports that GPT-6 Sol and Luna improve over GPT-5.6 models on static and multiturn jailbreak evaluations. It also says GPT-6 Sol produced about 36% fewer severity-3-or-higher misalignment flags than GPT-5.6 Sol in a simulated deployment over internal Codex traffic. OpenAI cautions that those results are a signal about deployment risk, not a guarantee for every external use case.
There are also trade-offs. On HealthBench, OpenAI says Sol and Luna maintain or improve performance on HealthBench Professional and HealthBench Consensus, but regress on HealthBench and HealthBench Hard. The company ties some of that to substantially shorter default answers.
What builders should do now
The practical test is not "is GPT-6 Sol smarter than every rival?" It is "does this model lower the cost of accepted work?"
For Codex and API teams, a good eval should compare Sol, Luna, GPT-6 Astra, GPT-5.6 Sol, and a non-OpenAI baseline on the same internal tasks. Track accepted diffs, failed tests, review comments, retries, wall-clock time, tool calls, cached tokens, and total cost. For agent systems, test mixed routing too: Luna for first-pass exploration, Sol for implementation, and Astra only when the task is high-stakes or repeatedly failing.
The short version: GPT-6 Sol and Luna are a cost-curve release. Astra remains the premium model, but Sol and Luna make GPT-6 more usable for the repetitive, iterative, agent-heavy work where token bills decide what teams are willing to automate.