The decision: start with GPT-6 Luna for bounded extraction, triage, and repeatable automation; use GPT-6 Sol for coding and work that needs judgment; reserve GPT-6 Astra for the hardest end-to-end jobs. OpenAI’s current picker starts Sol at Medium, Luna at High, and Astra at Light. Lower Luna’s effort only after an acceptance check shows that the cheaper setting still meets the task’s quality bar. Sol and Luna launched on September 22, 2026, and their strongest immediate advantage is price and capacity. (OpenAI model guidance)
OpenAI offers the models through the API as gpt-6-sol and gpt-6-luna, and through ChatGPT Work and Codex. They are not available in ordinary ChatGPT chats in the current rollout. Plus, Pro, Business, Enterprise, and Edu users can receive both in Work and Codex; Free and Go users can try Luna in the desktop app. Workspace settings and gradual rollout still control what an account can select. Both models accept text and image input, support none, low, medium, high, xhigh, and max reasoning effort, and have a 1.05 million token context window with 128,000 maximum output tokens. (OpenAI model catalog, release notes)
API prices are the cleanest starting point
The standard API card is quoted per 1 million tokens. These are dollars, separate from any ChatGPT subscription or purchased product credits.
| Model | Input | Cached input | Cache write | Output | Long prompt input / cached / write / output* |
|---|---|---|---|---|---|
| GPT-6 Sol | $2.00 | $0.20 | $2.50 | $10.00 | $4.00 / $0.40 / $5.00 / $15.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.125 | $0.50 | $0.20 / $0.02 / $0.25 / $0.75 |
| GPT-5.6 Sol | $4.00 | $0.40 | $5.00 | $20.00 | $8.00 / $0.80 / $10.00 / $30.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $0.25 | $1.20 | $0.40 / $0.04 / $0.50 / $1.80 |
*The last column applies when a prompt exceeds 272,000 input tokens; OpenAI bills the full request at twice the input and cache rates and 1.5 times the output rate. GPT-6 Sol and Luna therefore halve the current GPT-5.6 promotional input rate, while Luna’s output rate falls from $1.20 to $0.50. OpenAI describes the launch as 50% lower than GPT-5.6 promotional pricing, so treat that comparison as a dated baseline rather than a permanent claim about the older models.
Batch and Flex processing are priced at 50% of Standard rates, with Batch intended for asynchronous jobs and Flex trading response speed and availability for lower cost. Fast mode is priced at twice the applicable rate where the model and account support it. Regional processing adds 10%; EU data residency for Sol and Luna is available only with Standard processing. Cache writes cost 1.25 times uncached input at the Standard tier, while cached reads are 10% of uncached input. (API pricing, Sol model details, Batch, Flex)
At list rates, a 100,000-input/10,000-output turn costs about $0.30 on Sol and $0.015 on Luna before cache or processing discounts. A single 300,000-input/30,000-output codebase review crosses the 272,000-token threshold: it costs about $1.65 on Sol ($1.20 input + $0.45 output) and $0.0825 on Luna ($0.06 input + $0.0225 output), before cache writes. A ten-million-input/one-million-output day costs about $30 on Sol and $1.50 on Luna when it is an aggregate of individual requests that each stay at or below 272,000 input tokens. These are arithmetic illustrations, not AIHackers task results: tool calls, retries, cache misses, and review time can dominate the token bill.
Credits are a different ledger
ChatGPT Work and Codex use credits per 1 million tokens rather than API dollars. The current Standard-speed card is:
| Model | Input credits | Cached input credits | Output credits |
|---|---|---|---|
| GPT-6 Astra | 250 | 25 | 1,250 |
| GPT-6 Sol | 50 | 5 | 250 |
| GPT-6 Luna | 2.5 | 0.25 | 12.5 |
| GPT-5.6 Sol | 100 | 10 | 500 |
| GPT-5.6 Luna | 5 | 0.5 | 30 |
Fast mode uses 2.5 times the Standard credit rate for GPT-6 models where available. Included plan usage, purchased credits, API dollars, and any temporary promotion are separate balances. Credit prices also do not predict a fixed number of messages: context, reasoning effort, tools, retrieval, and caching change consumption. Check the account usage dashboard before treating a plan as capacity. (ChatGPT pricing and credits, Business and Enterprise rate card)
What the benchmark record actually says
OpenAI’s launch page reports strong vendor evaluations. Sol scored 33.2% on AutomationBench at xhigh for a stated $0.27 per task, versus 26.9% for Claude Opus 5 at max; Luna improved 5.4 percentage points over GPT-5.6 Luna at high effort with 58% lower cost per task. OpenAI reports 68.8% for Sol and 66.6% for Luna on DeepSWE v1.1, and 60.5% for Sol on OSWorld 2.0 offline. The same page says its GPT evaluations ran in OpenAI’s research environment or API and that competitor scores came from public reports. The cited competitor is the older Opus 5; this row does not compare Sol with Opus 5.5. These are vendor launch results, with different harnesses across some comparisons. (OpenAI launch evaluation table)
The independent record is more specific than a single leaderboard. Artificial Analysis’ v4.3.2 comparison uses ten evaluations at max effort: Sol scores 48 versus GPT-5.6 Sol at 47 on the Intelligence Index, while Luna scores 37 versus GPT-5.6 Luna at 37. Its Coding Agent Index article reports Sol at 57 versus 55 and Luna at 41 versus 43. On the separate Intelligence Index task set, it reports costs of $1.06 versus $1.99 for Sol and $0.07 versus $0.18 for Luna. These are point measurements from Artificial Analysis’ harness, with its own token and provider assumptions. (Artificial Analysis Sol comparison, Luna comparison, AA Sol and Luna analysis)
LiveBench’s June 25 release publishes the task rows and category map. Recomputing the unweighted mean of its seven category means gives Sol 79.2 versus GPT-5.6 Sol 81.1, and Luna 72.0 versus GPT-5.6 Luna 73.6. That is a dated CSV snapshot, not a confidence interval. ARC Prize’s verified Luna result reports 86.7% on ARC-AGI-1 and 59.3% on ARC-AGI-2 at max reasoning, with the harness and effort shown on the result page. Those scores describe specific task families; they do not settle the model choice for a different workflow. (LiveBench task table, LiveBench category map, ARC Prize verified result)
Our Opus 5.5 review compares its independent evidence and effort tradeoffs against the launch enthusiasm. AIHackers has not run a controlled comparison of accepted results for Sol or Luna; CAR is not-run.
That leads to a practical routing rule:
| Workload | Start with | Reconsider when |
|---|---|---|
| High-volume extraction, triage, routine transformations | Luna at high effort | A representative acceptance check passes at a lower effort, or review time erases the saving |
| Multi-file coding, agent workflows, ambiguous implementation | Sol at medium effort | The task needs deeper verification or Astra’s computer-use and judgment |
| Ambiguous, high-value, end-to-end work | Astra at light effort | The task needs more depth, tool use, or sustained review |
| Long prompts | Either model below 272K input tokens | The full-request long-context multiplier changes the budget |
| Offline evaluations and enrichment | Batch or Flex | The delayed completion or occasional unavailability misses the deadline |
Use these point comparisons as routing evidence, then measure accepted results on the same task set, effort, tools, provider route, and retry budget before changing a production default. For model availability, promotions, reset events, and unannounced plan chatter around Tuesday’s DevDay, keep the dated watch separate from this price and benchmark decision; the September launch and offers watch owns that volatile record. OpenAI’s official DevDay page confirms the event for Tuesday, September 29, 2026, but it has not announced an additional “o” model or a new Pro tier in the sources reviewed here.
For a small cross-provider evaluation, use the September value shortlist: Luna alongside MiMo V2.6 Flash, DeepSeek V4.1 Flash and Opus 5.5. Apply the cornerstone cost method to the work you accept, including retries and review time.
Reviewed September 27, 2026 (+08:00). Prices, rollout status, credits, and benchmark leaderboards can change; recheck the linked owner pages before committing to a route.