OpenAI shipped GPT-6 Sol and GPT-6 Luna on 22 September 2026, hours after Anthropic released Claude Opus 5.5. Both new models are priced at roughly half their GPT-5.6 equivalents. Flagship GPT-6 Astra is unchanged at $10 in and $50 out.
Pricing
| Model | Input / 1M | Output / 1M | Cached input / 1M |
|---|
| GPT-6 Astra | $10.00 | $50.00 | $1.00 |
| GPT-6 Sol | $2.00 | $10.00 | $0.20 |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 |
Cached input reads carry a 90% discount across the line. Both Sol and Luna are priced at 50% of the GPT-5.6 models they replace.
Specs
Sol and Luna both run a 1.05M token context window with 128K max output. Same envelope as Astra. The difference between the three tiers is capability and cost, not context.
Benchmarks Across the Three Tiers
| Benchmark | Luna | Sol | Astra |
|---|
| DeepSWE v1.1 | 66.6% | 68.8% | 74.1% |
| FrontierCode 1.1 | 42.4% | 49.3% | 53.3% |
| AutomationBench | 20.7% | 33.2% | 41.4% |
| Agents' Last Exam | 50.9% | 56.4% | 59.3% |
| OSWorld 2.0 | 52.7% | 64.4% | 73.5% |
| Factual error rate | 7.6% | 4.6% | 3.9% |
The Luna Number Nobody Expected
Luna scores 66.6% on DeepSWE v1.1. Sol scores 68.8%. Astra scores 74.1%.
Luna costs one twentieth of Sol and one hundredth of Astra on input. A 2.2 point drop from Sol for a 95% cost reduction is the most consequential line in this launch, and it is not the line OpenAI led with.
The catch is in the other rows. Luna falls off a cliff on AutomationBench at 20.7% against Sol's 33.2%, and on OSWorld 2.0 at 52.7% against 64.4%. Its factual error rate is 7.6%, nearly double Astra's 3.9%.
Read together those numbers describe a model that writes decent code when the task is handed to it in one piece, and loses its way when it has to decide what to do next. Luna is a worker, not a planner.
Sol's Actual Pitch
OpenAI put a cost-per-task figure on Sol: 33.2% completion on AutomationBench at $0.27 per task. That is the right way to publish an agentic number and more vendors should do it, because per-million-token pricing tells you almost nothing about what an agent run costs.
Sol also halves the error rate of its predecessor and improves substantially on FrontierCode. It is the general-purpose tier: good enough for most agentic work, a fifth of Astra's price.
Where You Can Use It
- API:
gpt-6-sol and gpt-6-luna
- ChatGPT Plus, Pro, Business, Enterprise, Edu: both models, through ChatGPT Work and Codex
- Free and Go: Luna only, in the desktop app
OpenAI described the rollout as gradual through launch day, so availability may lag the announcement on some accounts.
Which Tier Fits Your Workload
- High-volume, well-specified, repetitive tasks - Luna. Classification, extraction, summarisation, bulk transforms, first-pass code generation with a human or a stronger model reviewing.
- General agentic work with tool calls - Sol. The AutomationBench and OSWorld gaps make Luna a false economy here.
- Computer use, browser automation, long-horizon tasks - Astra. 73.5% on OSWorld 2.0 against Sol's 64.4% is the widest gap in the table.
- Accuracy is the product - Astra. A 7.6% factual error rate on Luna is not acceptable for anything customer-facing without a verification layer.
- Coding specifically - benchmark Sol against Claude Opus 5.5 at $4 in / $20 out. Opus 5.5 leads Terminal-Bench 4.0 at 66.4% against Astra's 57.9%, and Sol is cheaper than both.
The Two-Model Stack
The cheapest sensible configuration this launch enables is Luna doing the volume and Sol or Astra doing the judgement. Route the bulk work to Luna, have Sol check or plan, and only escalate to Astra when the task needs computer use or the error rate has to be under 4%.
At $0.10 in and $0.50 out, Luna is cheap enough that running it twice and comparing outputs still costs less than one Sol call. That is a new option and it is worth testing before you assume you need the bigger model.
Sources