OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, completing the GPT-6 family under the flagship GPT-6 Astra that shipped on September 3. The pitch is not a smarter model than Astra but the same generation at a fraction of the price: Sol halves GPT-5.6 Sol’s promotional token rates, and Luna halves input and cuts output pricing by 58%.
For people who code, research, or process documents at volume, that changes which model is worth reaching for. It does not by itself change what a task costs: reasoning effort, output length, caching, long-context surcharges, and tool calls can move the bill either way, and OpenAI’s launch benchmarks show good price-performance without proving broad superiority. Below: what changed, what the numbers mean, what each model costs in the API and in ChatGPT, and who should use which. We have not tested the models; no independent benchmark data existed at launch.
GPT-6 Sol and Luna at a glance
- Release date: September 22, 2026
- GPT-6 Sol is for: complex coding and agentic workflows
- GPT-6 Luna is for: focused, high-volume tasks such as summarization and extraction
- API model names: gpt-6-sol and gpt-6-luna
- Standard API price: Sol $2 per million input tokens and $10 output; Luna $0.10 and $0.50
- Context window: 1,050,000 tokens, 128,000 max output, on both
- Reasoning effort: none, low, medium (default), high, xhigh, max
- Availability: ChatGPT Work and Codex on Plus, Pro, Business, Enterprise, and Edu; the API; Luna in the desktop app for Free and Go
- Not included: a GPT-6 Terra, a Chat surface rollout, or independent benchmarks
What is new in GPT-6 Sol and Luna?
1. Token prices halved, with surcharges that still apply
| Price per million tokens | GPT-6 Sol | GPT-5.6 Sol (promo) | GPT-6 Luna | GPT-5.6 Luna |
|---|---|---|---|---|
| Input | $2.00 | $4.00 | $0.10 | $0.20 |
| Output | $10.00 | $20.00 | $0.50 | $1.20 |
| Cached input | $0.20 | $0.40 | $0.01 | $0.02 |
| Cache writes | $2.50 | $5.00 | $0.125 | $0.25 |
| Long context (over 272K input), input and output | $4 and $15 | $8 and $30 | $0.20 and $0.75 | $0.40 and $1.50 |
Sources: OpenAI API pricing page and model pages, September 22, 2026. Fast mode costs 2x, Batch and Flex 50%, regional data-residency processing adds 10% (EU residency for Sol and Luna only with Standard processing), and built-in tools such as web search, file search, and containers carry per-call fees. GPT-6 Astra stays at $10 and $50.
The 50% is measured against GPT-5.6 promotional pricing. GPT-5.6 Sol’s standard price was $5 and $30, with the $4 and $20 promotion running at least through November 21, 2026, so against the standard rate GPT-6 Sol is 60% cheaper on input and 67% on output. GPT-5.6 Luna’s $0.20 and $1.20 was itself an 80% cut in July, covered in our piece on the GPT-5.6 price cut; Luna has gone from $1 and $6 to $0.10 and $0.50 in under three months.
At equal token usage, GPT-6 Sol costs half as much as its predecessor. Actual cost per completed task still needs measurement: OpenAI’s own guidance says model choice, context, reasoning, tool use, retrieval, and caching all affect consumption. Caching pulls the other way: GPT-6 achieves higher cache hit rates by default, cached reads cost 10% of input, and changing effort or enabling tools mid-conversation no longer breaks the cache.
2. Higher usage limits in ChatGPT Work and Codex
ChatGPT Work and Codex share one allowance, metered in credits. OpenAI publishes estimated ranges of local messages per five-hour window; they are estimates, and cloud chats and image generation draw on the same pool.
| Model | Plus | Pro 5x | Pro 20x | Standard Business |
|---|---|---|---|---|
| GPT-6 Sol | 15 to 150 | 70 to 700 | 300 to 3,000 | 15 to 150 |
| GPT-6 Luna | 350 to 3,000 | 1,750 to 14,000 | 7,000 to 56,000 | 350 to 3,000 |
| GPT-5.6 Sol | 10 to 100 | 50 to 500 | 200 to 2,000 | 10 to 100 |
| GPT-5.6 Luna | 250 to 2,000 | 1,250 to 10,000 | 5,000 to 40,000 | 250 to 2,000 |
Source: ChatGPT Learn pricing page, September 22, 2026. Enterprise and Edu with flexible pricing have no fixed limits; without it they match Plus. Credit rates mirror the API cut: Sol costs 50 credits per million input tokens and 250 per million output against 100 and 500 for GPT-5.6 Sol; Luna 2.5 and 12.5 against 5 and 30. Two differences from the API: credit billing has no separate cache-write charge, and Fast mode costs 2.5x the credit rate rather than 2x.
3. Coding and agent claims
OpenAI says Sol improves substantially over GPT-5.6 Sol on FrontierCode, the merge-readiness benchmark, and reports near-parity with Claude Fable 5 on DeepSWE and with Claude Opus 5 on OSWorld at about 80% lower cost per task; the scores are in the table below. One change matters more than any score for teams that build agents. Both model pages state that built-in tools and function calling should use the Responses API; in Chat Completions, function calling works only with reasoning effort set to none. A pipeline on Chat Completions running GPT-5.6 Sol at medium or high effort cannot swap the model id and keep the same behavior. It moves to the Responses API or loses reasoning on tool-calling turns.
4. Fewer factual mistakes and a clearer style
On OpenAI’s internal factuality evaluation, GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol, approaching Astra-level reliability, and Luna at higher effort matches GPT-5.6 Sol at about a hundredth of the cost. The evaluation is built from de-identified conversations where users flagged errors, which OpenAI says are not representative of typical use.
Both models inherit Astra’s communication style: more clarity, less jargon, fewer low-value details, slightly shorter answers. OpenAI’s side-by-side shows Sol stating what it did and did not check instead of narrating implementation details. This is a vendor description with no independent test yet.
5. Alignment results and safeguards
OpenAI reports that Sol and Luna improve on their GPT-5.6 predecessors in its alignment evaluations, including lower rates of misleading claims about their own coding work, across deliberately adversarial scenarios: coding deception, broken search, reviewer bypass, warning circumvention, and unauthorized interaction. Those are the failure modes that matter when an agent runs unattended. OpenAI notes the tests do not measure rates in typical use and points to the system card for full results.
GPT-6 Sol and Luna benchmark results
OpenAI publishes selected comparisons rather than a full table. Its footnotes qualify each one: competitor scores come from public reports, Claude Fable 5 stands in where Fable 5.1 was unavailable, and evaluations ran in OpenAI’s research environment rather than production ChatGPT.
| Benchmark | What it measures | OpenAI’s claim | Footnote |
|---|---|---|---|
| AutomationBench 1.0.6 | Business workflows across 47 tools | Sol (xhigh) 33.2% at $0.27 per task; Astra (low) 30.3% at 3.9x; Claude Opus 5 (max) 26.9% at 11.1x; Fable 5.1 with Opus 5 fallback (max) 31.4% at more than 8.9x | Fable 5.1’s cost omits fallback runs on about 40% of tasks |
| Agents’ Last Exam V1 | Long-horizon professional tasks across 55 sub-industries | Sol (max) 56.4%, above Opus 5’s best at 60% lower cost per task | Opus 5 effort level not stated |
| DeepSWE v1.1 | Software engineering in real codebases | Sol (max) 68.8% vs Fable 5 (xhigh) 69.9%; Luna (max) 66.6% | Fable 5, not 5.1 |
| OSWorld 2.0 offline | Computer use | Sol (xhigh) 60.5% vs Opus 5 (medium) 60.3% | Partial reward on the offline set |
| FrontierCode 1.1 Main | Merge-ready code | Sol matches Fable 5.1 (xhigh) at much lower cost | No score published |
Read as price-performance claims, these hold up on their own terms: pairing Sol at xhigh with a Claude model at medium is a fair way to ask which configuration reaches a result more cheaply. They do not establish that Sol is the more capable model, and the weaknesses lie elsewhere: the benchmarks are OpenAI’s selection, competitor numbers come from other companies’ reports under harnesses that may differ, no confidence intervals are given, one comparison drops the cost of fallback runs, and the Claude set ends at Opus 5 and Fable 5.1. Claude Opus 5.5, released by Anthropic the same evening at $4 and $20, appears nowhere, and Artificial Analysis had no pages for either model at launch. Our Claude Opus 5.5 explainer covers the omitted model; our GPT-6 Astra benchmarks piece covers how OpenAI documents the flagship.
What GPT-6 Sol and Luna mean for everyday users
You do not need an API key to benefit. On Plus, Pro, Business, Enterprise, and Edu, both models appear in ChatGPT Work and Codex on the web, in the CLI, in the IDE extension, on iOS, and in cloud tasks, with a gradual rollout and, on Enterprise and Edu, only after an administrator enables them. Free and Go users get Luna in the desktop app at Standard speed. Neither model is in the Chat surface yet. One more date for Codex users: GPT-5.5 retires from ChatGPT, ChatGPT Work, and Codex on October 14, 2026, though not from the API.
Good uses for Sol include:
- Multi-file code changes, migrations, and bug investigations across repositories
- Agent workflows that call tools, browse, and run in the background
- Drafting technical documents where the model needs to check its own work
Good uses for Luna include:
- Summarizing documents, threads, and transcripts at volume
- Extracting structured fields from forms, emails, and reports
- Routine rewriting and reformatting where speed and cost matter more than depth
Using Luna for classification and other short, high-volume calls is our recommendation from its price and positioning, not a result OpenAI has published.
Should you switch to GPT-6 Sol or Luna?
Choose Sol when the task is multi-step and the cost of an incomplete result is high: agent runs, codebase-wide changes, research that must be sourced. Choose Luna when the task is narrow and repeated thousands of times. Choose Astra only where the outcome justifies five times Sol’s rate.
- If you use GPT-5.6 Sol or Luna through the API: switch model ids at the default medium effort and compare tokens billed per completed task before touching effort; OpenAI’s published scores come from xhigh and max. Check first whether your pipeline uses Chat Completions with function calling.
- If you compare against Claude: run your own tasks. Neither vendor’s launch table includes the other’s newest model.
- If you use ChatGPT only for short chats: the new models are in Work and Codex, not Chat, so for now the change you will notice is the higher five-hour allowance. Our GPT-5.6 Luna guide still describes where the cheapest tier breaks down.
The bigger picture
Two frontier labs cut prices on the same evening: Anthropic’s Opus 5.5 at 20% below Opus 5, then OpenAI’s Sol and Luna at 50% below GPT-5.6 promotional rates. The competition has moved from who has the top score to who delivers a given result for the least money, and both launch tables are built to win that argument while leaving out the other side’s newest model. The gap in OpenAI’s family is the middle: GPT-5.6 Terra has no successor. For Astra itself, see our GPT-6 Astra coverage.