GPT-6 Sol and Luna: API Prices, Benchmarks and Trade-offs

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, 19 days after GPT-6 Astra. Sol costs $2 per million input tokens and $10 per million output tokens; Luna costs $0.10 and $0.50. Both prices are half of what OpenAI charges for GPT-5.6 Sol and GPT-5.6 Luna under their current promotional pricing.

OpenAI’s launch page makes its case with cost-per-task charts at five effort levels. We read the data behind those charts rather than the points OpenAI chose to quote. It shows that Sol and Luna are much cheaper per task than their predecessors, but not always stronger: on two of the three coding and computer-use charts, GPT-5.6 Sol’s best score is higher than GPT-6 Sol’s.

Every score and cost below is OpenAI’s own, measured in OpenAI’s research environment or taken by OpenAI from public reports for competitor models. The Claude comparisons use Opus 5, because Anthropic released Opus 5.5 the same day and it appears on none of the charts.

According to the announcement, the two models are “not yet available in Chat,” the everyday ChatGPT surface. They are trained with methods similar to Astra’s, and OpenAI keeps Astra as its recommended choice when results matter more than cost. GPT-5.6 had a middle model, Terra. The GPT-6 section of OpenAI’s pricing page lists only Astra, Sol and Luna.

One integration detail from the Sol model page affects older code. Chat Completions supports function calling only when reasoning_effort is set to none. Agents that call tools with reasoning switched on need the Responses API.

Three lines in the pricing table change the real bill more than the headline does.

• Cache writes cost extra. Writing a prompt prefix to the cache costs 1.25× the normal input rate, and reading it back costs 10%. A prefix pays for itself from the second request that reuses it.
• Long prompts cost more. Above 272K input tokens, the whole request is billed at 2× the input and cache rates and 1.5× the output rate. The 1.05M-token window is usable, but a single oversized request puts Sol at $4 / $15.
• The comparison price is temporary. The 50% cut is measured against GPT-5.6 Sol’s promotional rate of $4 / $20, which OpenAI guarantees only “at least through November 21, 2026.” The pricing page does not show what GPT-5.6 Sol will cost after that.

Against Anthropic, Sol’s rates match Claude Sonnet 5 on every line both vendors publish: $2 input, $2.50 for a short cache write, $0.20 for a cache read and $10 output, per Anthropic’s pricing page. Luna at $0.10 / $0.50 is a tenth of Claude Haiku 4.5’s $1 / $5. Our frontier model price index tracks these rates across vendors.

OpenAI plotted most models at five effort levels on six charts. The table shows the best score each new model reached, the effort level that produced it and its cost per task, next to the best score by any model on the same chart.

OpenAI’s headline comparisons all check out against its own data. On AutomationBench, Sol at xhigh scores 33.2% for $0.27 a task, above Claude Opus 5 at max (26.9%) for 9% of Opus 5’s cost. On Agents’ Last Exam, Sol at max (56.4%) beats Opus 5’s best score (55.9%, at high) for 40% of the cost. On OSWorld, Sol at xhigh matches Opus 5 at medium, 60.5% against 60.3%, for about a sixth of the cost.

Factual reliability improved at every effort level. On OpenAI’s internal test, built from ChatGPT conversations where users had flagged a mistake, GPT-6 Sol’s error rate is roughly half of GPT-5.6 Sol’s at each setting, for example 5.1% against 10.8% at high. Luna at max makes fewer errors (7.6%) than GPT-5.6 Sol at max (8.5%) for about 1.4% of the cost. OpenAI notes that these conversations were picked because they caused errors, so everyday error rates are lower.

Four points do not appear in the announcement’s text but change how its claims should be read.

1. The FrontierCode comparison uses Fable 5.1’s weakest setting. OpenAI says Sol can “match Claude Fable 5.1 xhigh at much lower cost.” It does: 49.3% for $2.14 against 48.7% for $9.27. But xhigh is Fable 5.1’s lowest score on that chart. Fable 5.1 at low scores 49.8% for $2.38, and Claude Opus 5 at medium has the chart’s top score, 53.4%. On this benchmark Sol’s advantage over GPT-5.6 Sol is cost. GPT-6 Sol at high matches GPT-5.6 Sol at max (47.7% against 47.5%) for about a fifth of the cost.
2. GPT-5.6 Sol still has the higher top score on two charts. On DeepSWE, GPT-5.6 Sol at max scores 72.7% against GPT-6 Sol’s 68.8%. On OSWorld the figures are 66.2% against 64.4%. The older model costs more than twice as much per task at those settings ($6.46 against $2.74 on DeepSWE), so it only makes sense where its extra points are worth the price.
3. The DeepSWE saving is measured against an expensive rival. OpenAI’s “approximately 80% lower cost” compares Sol at max with Claude Fable 5 at xhigh. Opus 5 at medium scores 68.9%, level with Sol’s 68.8%, for $3.29 against $2.74. Against that point, Sol is about 17% cheaper.
4. The Claude model on the charts is Opus 5. Anthropic’s own launch table puts Claude Opus 5.5 at 40.0% on AutomationBench, above Sol’s 33.2%, at a per-token price 20% below Opus 5. No cost per task has been published for Opus 5.5 on the same harness, so the cost gap between the two is unknown.

OpenAI’s own footnotes add two caveats. Competitor results were “taken from publicly available reports,” not rerun by OpenAI, and Fable 5 appears wherever no Fable 5.1 score existed. The Fable 5.1 cost on AutomationBench also leaves out the Opus 5 fallback runs, which OpenAI says happened on about 40% of tasks.

Both models default to medium. The chart data shows where extra effort pays and where it stops paying.

Sol: stop at xhigh unless the task is hard. On AutomationBench, max scores lower than xhigh and costs more. On the two coding charts max adds 0.8 and 2.2 points for 1.6 and 2.7 times the cost. That trade is worth it for a long refactor that would otherwise need a second attempt, and not for routine tickets.

Luna: run it at max. Luna reaches its highest score at max on every chart, and its most expensive setting is still cheap. On DeepSWE, Luna at max scores 66.6%, the same as Sol at xhigh, for $0.22 a task against $1.00. It also scores higher than Sol at low on four of the five agentic charts, and costs less on four. AutomationBench is the one chart where Sol at low edges ahead (21.2% against 20.7%). A team that picks Sol at a low setting to save money should try Luna at max first.

More effort is not always better. On Agents’ Last Exam, both models score higher at medium than at high. Differences of a point or two between adjacent settings are likely within run-to-run variation, so test two or three levels on your own tasks instead of picking the highest one.

Caching. OpenAI says caching for GPT-6 now hits more often by default and gives a 90% discount on cached input. Two changes matter most for agents. Raising or lowering reasoning effort mid-conversation no longer breaks the cache, and neither does switching tools on or off. Developers can also set explicit breakpoints to choose where a cached prefix ends, and a new diagnostics tool explains missed cache hits. The announcement adds a partner figure: GitHub reports that these improvements cut the share of prompt tokens needing fresh processing by more than 50% for Copilot.

Alignment. OpenAI published five stress tests, each designed to provoke bad behaviour, so the rates are not what you should expect in normal use. The figure in brackets is the GPT-5.6 predecessor.

Most rates fell sharply, especially for Sol. The exception is warning circumvention: Sol circumvented a warning in 64.4% of test cases, barely better than GPT-5.6 Sol’s 68.2% and well above Luna’s 42.4% and Astra’s 17.4%. Sol also interacted with another agent without permission in 11.3% of cases, where Luna and Astra scored zero. If you run Sol unattended, enforce your guardrails in code, such as permission checks and approval steps, and don’t rely on the model to respect a warning. For Astra’s own launch figures, see our Astra pricing and benchmarks guide.