GPT-6 Sol and Luna: Half the Price of GPT-5.6, With Benchmarks and Cost at Every Effort Level
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, at $2 and $10 and $0.10 and $0.50 per million tokens, half the price of GPT-5.6. Sol makes about half as many factual errors as GPT-5.6 Sol. Luna scores 66.6% on DeepSWE for about $0.22 a task. GPT-6 Astra still leads every OpenAI benchmark.
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. They are the mid-tier and budget models under GPT-6 Astra, and both cost half what their GPT-5.6 versions did: $2 and $10 per million tokens for Sol, and $0.10 and $0.50 for Luna. They are available now in the API as gpt-6-sol and gpt-6-luna, and rolling out in ChatGPT Work, Codex, and GitHub Copilot (OpenAI).
The verdict: Sol is the best-value model for high-volume agent work, and Luna is the cheapest model that can still do real coding. Sol at high effort matches Claude Opus 5.5 at high effort on AutomationBench for about a third of the cost per task. Luna at max effort scores 66.6% on DeepSWE for $0.22, the same score Sol gets at xhigh for $1.00. But Sol is not a straight upgrade on hard coding. At max effort, GPT-5.6 Sol still scores higher on DeepSWE and OSWorld offline, and Astra and Opus 5.5 lead every benchmark on which they overlap with Sol.
Key Takeaways
- Price: Sol is $2 input, $0.20 cached, and $10 output per million tokens. Luna is $0.10, $0.01, and $0.50. OpenAI says both are 50% cheaper than GPT-5.6 and that these are "permanent prices, not promotional or introductory pricing."
- Factuality: Sol makes about half as many factual errors as GPT-5.6 Sol, with a 4.5% error rate at its best setting against 3.9% for Astra.
- Workflows: Sol at xhigh scores 33.2% on AutomationBench for $0.27 a task, beating Astra at low effort (30.3% for $1.08).
- Coding: Luna scores 66.6% on DeepSWE for $0.22 a task. Sol reaches 68.8% at max effort, below GPT-5.6 Sol's 72.7%.
- Effort: Use xhigh for Sol. Use max for Luna, whose lower settings fall apart on agentic tasks.
- Specs: Both have a 1.05M-token context window and 128K max output, with six effort levels from
nonetomax. - Caching: On GPT-6 models, changing reasoning effort mid-conversation no longer breaks the cache, and new explicit breakpoints and diagnostics help keep cache hit rates high.
Pricing and Specs
| Item | GPT-6 Luna | GPT-6 Sol | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|
| Input (per million tokens) | $0.10 | $2 | $10 | $4 |
| Cached input | $0.01 | $0.20 | separate rate | not listed here |
| Output | $0.50 | $10 | $50 | $20 |
| Cache write | $0.125 | $2.50 | not listed here | not listed here |
| Context window | 1.05M | 1.05M | not listed here | not listed here |
| Max output | 128K | 128K | not listed here | not listed here |
| Knowledge cutoff | May 18, 2026 | April 20, 2026 | not listed here | not listed here |
Sources: the GPT-6 Sol and GPT-6 Luna model pages, OpenAI's launch post, and our Astra coverage. Batch processing is 50% off for both models. On Sol, requests over 272K tokens cost twice as much for input and 1.5 times as much for output, so very long single requests cost more than the headline rate. Both models support reasoning effort none, low, medium, high, xhigh, and max; Luna defaults to medium. In Chat Completions, Sol only allows function calling when reasoning effort is none, so agent loops belong on the Responses API.
Benchmarks at Every Effort Level
OpenAI published score-versus-cost charts for six evaluations, with one point per effort level. We pulled every value from the launch page's interactive charts. Costs are OpenAI's estimates per task, and every score is vendor-reported.
GPT-6 Sol:
| Benchmark | Low | Medium | High | Xhigh | Max |
|---|---|---|---|---|---|
| AutomationBench 1.0.6 | 21.2% / $0.19 | 26.9% / $0.21 | 31.2% / $0.24 | 33.2% / $0.27 | 32.0% / $0.34 |
| Agents' Last Exam V1 | 48.7% / $0.86 | 53.1% / $1.27 | 52.6% / $1.53 | 55.4% / $1.67 | 56.4% / $2.93 |
| FrontierCode 1.1 Main | 37.3% / $0.45 | 45.9% / $0.80 | 47.7% / $1.08 | 48.4% / $1.37 | 49.3% / $2.14 |
| DeepSWE 1.1 | 37.2% / $0.16 | 56.6% / $0.38 | 65.3% / $0.64 | 66.6% / $1.00 | 68.8% / $2.74 |
| OSWorld 2.0 offline | 43.9% / $0.97 | 54.0% / $1.32 | 58.3% / $1.64 | 60.5% / $2.21 | 64.4% / $3.25 |
| Factual error rate (lower is better) | 11.4% | 6.9% | 5.1% | 4.5% | 4.6% |
GPT-6 Luna:
| Benchmark | Low | Medium | High | Xhigh | Max |
|---|---|---|---|---|---|
| AutomationBench 1.0.6 | 1.2% / $0.006 | 9.4% / $0.016 | 14.5% / $0.021 | 12.6% / $0.025 | 20.7% / $0.037 |
| Agents' Last Exam V1 | 36.3% / $0.025 | 46.8% / $0.11 | 43.6% / $0.11 | 47.9% / $0.11 | 50.9% / $0.15 |
| FrontierCode 1.1 Main | 25.7% / $0.021 | 35.5% / $0.053 | 37.3% / $0.067 | 37.1% / $0.073 | 42.4% / $0.11 |
| DeepSWE 1.1 | 2.4% / $0.006 | 44.5% / $0.052 | 59.3% / $0.084 | 61.3% / $0.11 | 66.6% / $0.22 |
| OSWorld 2.0 offline | 8.3% / $0.030 | 31.5% / $0.062 | 41.4% / $0.12 | 46.7% / $0.16 | 52.7% / $0.27 |
| Factual error rate (lower is better) | 27.7% | 17.5% | 12.5% | 10.2% | 7.6% |
AutomationBench tests business workflows across 47 tools. Agents' Last Exam covers long professional tasks across 55 sub-industries. FrontierCode grades whether a code change is ready to merge. DeepSWE tests long software-engineering tasks in real codebases. The factuality evaluation uses real ChatGPT conversations where users had flagged an earlier model's mistake, so OpenAI notes the error rates are far higher than in typical use.
Which Effort Level to Use
Run Sol at xhigh and Luna at max. The tables above make the case:
- Sol at xhigh gets its best AutomationBench score and its lowest factual-error rate. Max costs more and scores lower on both.
- Sol at high is the budget setting for coding. It reaches 65.3% on DeepSWE for $0.64. Max adds 3.5 points for 4.3 times the cost.
- Use Sol at max only for computer use and long professional tasks, where OSWorld (+3.9 points) and Agents' Last Exam (+1 point) still improve.
- Never run Luna at low for agent work. It scores 2.4% on DeepSWE and 1.2% on AutomationBench. Luna at max costs under $0.30 a task on every chart, so there is little reason to save.
A useful cross-check: Luna at max and Sol at xhigh both score 66.6% on DeepSWE, at $0.22 and $1.00 a task. For coding subagents, try Luna first.
GPT-6 Sol vs GPT-5.6 Sol: Is It an Upgrade?
For reliability, business workflows, and cost, yes. For raw coding ceiling, not quite. Sol's best scores are higher than GPT-5.6 Sol's on AutomationBench (33.2% vs 28.8%), Agents' Last Exam (56.4% vs 53.6%), and FrontierCode (49.3% vs 47.5%). Its best factual-error rate is about half (4.5% vs 8.4%). Each of those results also costs much less per task.
OpenAI's post leaves out one thing. At max effort, GPT-6 Sol scores below GPT-5.6 Sol on DeepSWE (68.8% vs 72.7%) and on OSWorld offline (64.4% vs 66.2%). Below the top, the new model gives the same score for much less: on OSWorld offline, Sol at xhigh scores 60.5% for $2.21 a task, where GPT-5.6 Sol needed $5.93 for 60.9%. If you pinned GPT-5.6 Sol at max for hard coding, re-run your own evals before you switch.
Luna improves more clearly. Its best DeepSWE score rises from 62.2% to 66.6% while the cost falls from $0.53 to $0.22 a task. Its best factual-error rate drops from 12.0% to 7.6%. On Agents' Last Exam, it gains half a point at max effort while costing about 94% less per task ($0.15 vs $2.57).
GPT-6 Sol and Luna vs GPT-6 Astra
Astra is still OpenAI's best model on every chart. Sol gets you most of the way for a fifth of Astra's token price. Luna gets you surprisingly far for a hundredth of it.
| Benchmark (best score) | GPT-6 Astra | GPT-6 Sol | GPT-6 Luna |
|---|---|---|---|
| AutomationBench | 41.4% ($1.73) | 33.2% ($0.27) | 20.7% ($0.04) |
| Agents' Last Exam | 59.3% ($6.23) | 56.4% ($2.93) | 50.9% ($0.15) |
| FrontierCode | 53.3% ($4.59) | 49.3% ($2.14) | 42.4% ($0.11) |
| DeepSWE | 74.1% ($4.43) | 68.8% ($2.74) | 66.6% ($0.22) |
| OSWorld offline | 73.5% ($9.07) | 64.4% ($3.25) | 52.7% ($0.27) |
| Factual error rate | 3.9% | 4.5% | 7.6% |
OpenAI still calls Astra "the world's best model for computer use." OSWorld is where Sol is furthest behind, at 9 points, so keep Astra for computer-use agents. Everywhere else, Sol's gap is 3 to 8 points. OpenAI also points out that Sol at xhigh beats Astra at low effort on AutomationBench, at about a quarter of the cost.
GPT-6 Sol vs Claude Opus 5.5 and Fable 5.1
Claude Opus 5.5 scores higher. Sol is cheaper per task. Anthropic released Opus 5.5 the same day at $4 and $20 per million tokens, twice Sol's price. Both charge $0.20 per million for cached input. The two labs report only two benchmarks in common:
| Benchmark | GPT-6 Sol (best) | Claude Opus 5.5 (best) | Closest-cost comparison |
|---|---|---|---|
| FrontierCode | 49.3% ($2.14) | 54.4% | Opus 5.5 at medium: 54.6% for $0.80 |
| AutomationBench | 33.2% ($0.27) | 40.0% ($1.37) | Sol at high 31.2% for $0.24 vs Opus 5.5 at high 32.0% for $0.70 |
On FrontierCode, Anthropic's own figures put medium-effort Opus 5.5 above Sol's best score for less money per task. On AutomationBench, Sol matches Opus 5.5 at the same effort level for about a third of the cost. That makes Sol the better buy for routine, high-volume automation and Opus 5.5 the better buy for code quality (Anthropic).
Treat cross-vendor cost claims as ranges. OpenAI says Sol at xhigh beats Claude Opus 5 at max on AutomationBench "at just 9% of Opus 5's cost per task." OpenAI's chart prices that Opus 5 run at $3.05 a task. Anthropic's chart, from Zapier's leaderboard, prices the same run at $1.27, which would make Sol about 21% of Opus 5's cost. The two vendors do agree on Astra's cost within 5%.
Against Claude Fable 5.1, OpenAI says Sol at xhigh matches Fable 5.1 at xhigh on FrontierCode (48.4% vs 48.7%) for $1.37 a task against $9.27. OpenAI notes its AutomationBench cost for Fable 5.1 leaves out Opus 5 fallbacks, which ran on about 40% of tasks.
Prompt Caching Changes
Alongside the models, OpenAI shipped a new prompt-caching system for GPT-6. It gets higher cache hit rates by default and gives cache discounts on shared prefixes reused within a 30-minute window (OpenAI). The changes that matter for agent builders:
- Change effort without breaking the cache. Add a
configuration_updatemid-conversation to raise or lower reasoning effort, leaving the request-level setting unchanged. - Keep tools stable. Limit what the model can call with
allowed_tools, or settool_choicetonone, instead of removing tool definitions, which breaks the cache. - Explicit breakpoints let you choose where a cached prefix ends, so fast-changing content can sit after it.
- Prewarming processes known context before the first request arrives.
- A Prompt Caching Dashboard and diagnostics tool report cache misses with a reason, such as
tools_changed, and the number of tokens affected.
GitHub's Mario Rodriguez says Copilot has cut the share of prompt tokens needing fresh processing by more than 50%. Strawberry Browser's CTO says the diagnostics and dashboard cut its costs by 20%.
Availability
- OpenAI API:
gpt-6-solandgpt-6-luna, on the Responses and Chat Completions APIs and Batch. - ChatGPT Work and Codex: Rolling out through launch day for Plus, Pro, Business, Enterprise, and Edu. Free and Go users get Luna in the desktop app. Neither model is in Chat yet (OpenAI). OpenAI's developer announcement also mentions a banked usage reset for Plus, Pro, and Business accounts (OpenAI Developer Community).
- GitHub Copilot: Sol on Pro+, Max, Business, and Enterprise, and Luna on those plus Pro. Both are available in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, and the cloud agent. The rollout is gradual (GitHub Changelog).
Alignment
OpenAI says Sol and Luna carry over Astra's alignment work. Its coding-deception evaluation uses tasks deliberately chosen to push models toward dishonesty. On it, Sol's deception rate falls from 10.4% to 1.3% and Luna's from 9.5% to 2.8%, with Astra at 0.5%. OpenAI stresses that these tests do not measure typical use, and the full results are in the system card. OpenAI also says both models write "slightly shorter answers overall" and use less jargon, a style carried over from Astra.
When to Choose GPT-6 Sol or Luna
Choose Sol at xhigh for CRM updates, ticket triage, research agents, and other multi-app workflows. Its per-task cost is the lowest among models near its score.
Choose Luna at max for subagents, bulk classification, first-pass code changes, and any job you run thousands of times. A failed Luna run costs cents.
Choose Claude Opus 5.5 when code quality matters more than token price. It leads Sol on FrontierCode even at medium effort.
Keep Astra for computer-use agents and the hardest long tasks, and keep GPT-5.6 Sol at max only if your own coding evals favor it.
Before you switch, move agent loops to the Responses API, set effort explicitly, and use the new caching controls so effort changes do not throw away your cache.
Sources
- OpenAI, "Introducing GPT-6 Sol and Luna" (pricing, benchmarks, effort-level charts, availability, alignment; September 22, 2026): https://openai.com/index/introducing-gpt-6-sol-and-luna/
- OpenAI, "Better prompt caching for GPT-6" (caching system, breakpoints, diagnostics, partner quotes; September 22, 2026): https://openai.com/index/better-prompt-caching-for-gpt-6/
- OpenAI, GPT-6 Sol model page (context, output, cutoff, pricing, long-context surcharge, effort levels; checked September 22, 2026): https://developers.openai.com/api/docs/models/gpt-6-sol
- OpenAI, GPT-6 Luna model page (context, output, cutoff, pricing, default effort; checked September 22, 2026): https://developers.openai.com/api/docs/models/gpt-6-luna
- OpenAI Developer Community, "Announcing GPT-6 Sol and GPT-6 Luna" (availability, banked reset; September 22, 2026): https://community.openai.com/t/announcing-gpt-6-sol-and-gpt-6-luna/1399925
- GitHub Changelog, "OpenAI's GPT-6 Sol and GPT-6 Luna now available" (Copilot plans and surfaces; September 22, 2026): https://github.blog/changelog/2026-09-22-openais-gpt-6-sol-and-gpt-6-luna-now-available/
- Anthropic, "Introducing Claude Opus 5.5" (Opus 5.5 scores and cost per task; September 22, 2026): https://www.anthropic.com/claude-opus-5-5
Read next
Keep building the workspace playbookClaude Opus 5.5: Fable 5.1-Level Scores at $4 and $20, With Benchmarks, Effort Costs, and Migration Guide
Claude Opus 5.5, released September 22, 2026, beats Claude Fable 5.1 on every benchmark Anthropic published and costs $4 and $20 per million tokens, 60% less than Fable 5.1. It scores 66.4% on Terminal-Bench 4.0 against 57.9% for GPT-6 Astra. Thinking is always on, the default effort is medium, and four API changes can break Opus 5 code.
OpenAI Launches GPT-6 Astra for Computer Use and Complex Work
OpenAI is rolling out GPT-6 Astra, a model focused on computer use, browsing, software engineering, cybersecurity, science, and complex professional workflows. Access starts with a limited set of organizations before expanding to ChatGPT plans and developer platforms.
OpenAI Previews GPT-5.6: Sol, Terra, and Luna in Limited Preview
OpenAI announced a limited preview of the GPT-5.6 family: Sol, a next-generation frontier flagship OpenAI calls a step function better than GPT-5.5; Terra, a balanced model competitive with GPT-5.5 at 2x lower cost; and Luna, its most cost-efficient model. Access starts with trusted partners in Codex and the API.
Frequently Asked Questions
What are GPT-6 Sol and GPT-6 Luna?
They are OpenAI's mid-tier and budget GPT-6 models, released September 22, 2026, alongside the flagship GPT-6 Astra. OpenAI says they were trained with methods similar to Astra's and are built for cost-efficient coding, agents, and everyday work.
How much do GPT-6 Sol and Luna cost?
GPT-6 Sol costs $2 per million input tokens, $0.20 cached, and $10 output. GPT-6 Luna costs $0.10 input, $0.01 cached, and $0.50 output. Both are 50% cheaper than GPT-5.6, and OpenAI says the prices are permanent.
Is GPT-6 Sol better than GPT-5.6 Sol?
Mostly. Sol makes about half as many factual errors and scores higher on AutomationBench, Agents' Last Exam, and FrontierCode for much less per task. At max effort, though, GPT-5.6 Sol still scores higher on DeepSWE and OSWorld offline.
How does GPT-6 Sol compare with Claude Opus 5.5?
Opus 5.5 scores higher on the two benchmarks both labs report: 54.4% vs 49.3% on FrontierCode and 40.0% vs 33.2% on AutomationBench. Sol costs half as much per token and reaches Opus 5.5's high-effort AutomationBench score for about a third of the cost per task.
Which reasoning effort should I use?
Use xhigh for Sol, where AutomationBench and factuality peak and max adds cost for small gains. Use max for Luna. Its low and medium settings collapse on agentic benchmarks, and even max costs under $0.30 a task on every OpenAI chart.
Where can I use GPT-6 Sol and Luna?
In the OpenAI API as gpt-6-sol and gpt-6-luna, in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu, and in GitHub Copilot. Free and Go users get Luna in the desktop app. Neither is in Chat yet.