OpenAI’s GPT-6 Sol and Luna cut API prices, but the benchmark picture is not a clean sweep

OpenAI has added two models below GPT-6 Astra: GPT-6 Sol, aimed at complex coding and agent workflows, and GPT-6 Luna, aimed at focused, high-volume work. The concrete change is price. Sol lists at $2 per million input tokens and $10 per million output tokens. Luna lists at $0.10 and $0.50.

That puts Sol at half the published GPT-5.6 Sol tariff, and Luna below the already inexpensive GPT-5.6 Luna. It also gives OpenAI a clearer three-tier API family: Astra for the most demanding work, Sol for expensive agentic work that needs a lower bill, and Luna for volume. The launch announcement makes a broad 50% price-cut claim, and the underlying rate card mostly supports it. The important exception is Luna output, which falls from $1.20 to $0.50 per million tokens, a steeper 58.3% drop.

The performance case needs more care. OpenAI publishes useful numbers, especially on AutomationBench, DeepSWE and OSWorld. But those are OpenAI-reported comparisons, often at different reasoning-effort settings, and some peer cost figures are incomplete. Independent measurement is beginning to appear at Artificial Analysis; the public Agent Arena agent leaderboard snapshot available at review time did not yet list either new model.

What is actually shipping

Both models use the Responses API and Chat Completions API, accept text and image input, return text, and expose reasoning.effort settings from none through max. The API documents list a 1.05 million-token context window, a 922,000-token maximum input and a 128,000-token maximum output for each model. They support streaming, structured outputs, function calling, file search, web search, prompt caching and the Responses API tool set, including hosted shell, code interpreter, computer use and MCP.

The OpenAI launch page says Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. It also says Free and Go users can access Luna in the desktop app, while the models are not yet available in Chat. That is a staged product rollout, not a claim that every ChatGPT surface has already switched.

The price card, including the parts that change the bill

The table below is OpenAI’s Standard short-context pricing, per million tokens. Cached input is the rate for a reused prompt prefix; cache writes are the higher-priced input class used to create a reusable prefix.

Model	Fresh input	Cached input	Cache write	Output
GPT-6 Sol	$2.00	$0.20	$2.50	$10.00
GPT-5.6 Sol	$4.00	$0.40	$5.00	$20.00
GPT-6 Luna	$0.10	$0.01	$0.125	$0.50
GPT-5.6 Luna	$0.20	$0.02	$0.25	$1.20

The new prices are simple on an ordinary request. A synthetic short-context agent turn with 100,000 fresh input tokens, 100,000 cached input tokens and 50,000 output tokens costs $0.720 on Sol instead of $1.440 on GPT-5.6 Sol. On Luna, the same turn costs $0.036 instead of $0.082, a 56.1% reduction because the cheaper output rate contributes to the mix. That example remains below OpenAI’s 272,000-token input threshold.

There are several terms behind the headline rates in the API pricing table:

• Batch and Flex cost half of Standard rates. That makes Sol $1/$5 and Luna $0.05/$0.25 for input/output, before any other modifier.
• Fast mode costs twice the applicable rate. For short-context requests, that is $4/$20 for Sol and $0.20/$1 for Luna. OpenAI says Fast mode targets lower latency, but a request that is downgraded to Standard because traffic ramps too quickly is charged at Standard rates. The Fast mode documentation is explicit about that distinction.
• A request with more than 272,000 input tokens is charged long-context prices for the entire request: twice the input and cache rates, and 1.5 times the output rate. A million-token context window does not mean a million tokens always receive the short-context tariff.
• Cached input costs 10% of the uncached input rate, while a cache write costs 1.25 times the fresh-input rate. The prompt-caching guide says a matching rendered prefix is required, so a cache discount should be treated as a workload outcome, not guaranteed savings.
• Eligible data-residency processing carries a 10% uplift. For Sol and Luna, EU data residency is available only with Standard processing, not Batch, Flex or Fast.

Token rates are not the whole bill for an agent. OpenAI also prices web search at $10 per 1,000 calls plus search-content tokens at model rates. Hosted Shell and Code Interpreter containers range from $0.03 for 1 GB to $1.92 for 64 GB per 20-minute session, and the model tokens used inside those tools are billed at the selected model rate. File search costs $0.10 per GB-day after the first free GB and $2.50 per 1,000 tool calls. Those are platform tool costs, not Sol or Luna token tariffs, but they matter as soon as an application starts browsing or executing code.

OpenAI’s benchmark case: good data, uneven comparisons

OpenAI’s clearest comparable table is AutomationBench, a benchmark of end-to-end business workflows across 47 tools in sales, marketing, operations, support, finance and HR. Its published row for Sol is specific enough to be useful:

OpenAI-reported AutomationBench result	Score	Reported cost per task
GPT-6 Sol, xhigh effort	33.2%	$0.27
GPT-6 Astra, low effort	30.3%	3.9x Sol’s cost
Claude Opus 5, max effort	26.9%	11.1x Sol’s cost
Claude Fable 5.1 with Opus 5 fallback, max effort	31.4%	More than 8.9x Sol’s cost, fallback cost not reported

That table lets OpenAI make a fair narrow point: at the selected settings, Sol scored above the listed Opus 5 and Fable 5.1 rows, while costing less. It does not establish a universal lead. The Fable row is particularly awkward: OpenAI says its stated cost omits Opus 5 fallbacks that occurred on roughly 40% of tasks. The displayed score may be useful, but the displayed cost is not a fully reproducible all-in comparison.

Luna’s AutomationBench claim is more limited. OpenAI says Luna at high effort improved 5.4 percentage points over its predecessor at 58% lower cost per task, without supplying the same full peer table in the launch text.

The coding and computer-use results also deserve their labels:

• On Agents’ Last Exam, OpenAI reports Sol at maximum effort scoring 56.4%, above the cited top Opus 5 result at 60% lower cost per task. The launch text does not provide the comparison model’s exact score and cost beside that claim.
• On DeepSWE v1.1, Sol at maximum effort scored 68.8%. Claude Fable 5 at xhigh effort scored 69.9%, so Sol was within 1.1 points, not ahead. OpenAI says Sol cost about 80% less per task in that comparison. Luna at maximum effort scored 66.6% and OpenAI says it cost 93% less than Opus 5 and 96% less than Fable 5 in selected medium-effort comparisons.
• On the OSWorld 2.0 offline set, Sol at xhigh effort scored 60.5%, versus 60.3% for Claude Opus 5 at medium effort. OpenAI says the Sol run cost about 80% less. OSWorld itself describes 108 long-horizon computer-use workflows, with a median human completion time of about 1.6 hours, so these are not lightweight browser tasks. Luna at maximum effort is said to exceed GPT-5.6 Sol at medium effort for one-tenth of the cost, but the launch text does not print both raw scores.

OpenAI also claims better factuality and lower rates of misleading coding-work claims than the GPT-5.6 models. Those results come from internal evaluations. The factuality set draws on de-identified conversations where users had flagged factual errors, and the system-card alignment tasks deliberately seek difficult failure conditions. Neither is a production error-rate estimate.

The independent picture is useful, but not settled

Artificial Analysis currently lists GPT-6 Sol at 48 on its Intelligence Index and GPT-6 Luna at 37. Its index combines ten evaluations, including AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity’s Last Exam and GDPval-AA. That makes it a valuable separate signal, not a replication of OpenAI’s own effort-specific scorecard.

The same Artificial Analysis profiles place the listed raw rates at $2/$10 for Sol and $0.10/$0.50 for Luna, and estimate index-task costs of $1.06 and $0.07 respectively. Those index-task figures are benchmark-run accounting, not a quote for a production task.

There is one specification mismatch worth watching. OpenAI documents a 1.05M-token context window for both APIs. The captured Artificial Analysis model profiles list 872K for Sol and 1M for Luna. The profiles do not explain the difference, so OpenAI’s API documentation is the right source for the advertised product limit, while the independent catalog’s context fields should not be silently substituted for it.

Agent Arena’s public agent leaderboard snapshot was dated September 15 and did not list either GPT-6 Sol or GPT-6 Luna. That is not evidence that the models lack agentic performance. It means there was no public Agent Arena result to cite at the time of review.

The practical read

Sol is the interesting option for teams that were already spending heavily on agentic coding or multi-tool workflows and can verify quality at the chosen reasoning effort. Its tariff is not merely cheaper than GPT-5.6 Sol; it is half the list price before caching, Batch or Flex enter the picture.

Luna is the sharper cost play. At $0.10 per million input tokens and $0.50 per million output tokens, it can change the economics of high-volume classification, extraction, routing and narrow tool-use loops. Its published benchmark results are not a substitute for testing a specific workload, especially when effort level, tool calls, long-context billing and output length can dominate the final bill.

OpenAI has made a credible case that the GPT-6 family now has a lower-cost middle and a much cheaper volume tier. It has not supplied a clean, independently replicated verdict that either model wins every frontier-agent benchmark. For buyers, the next useful comparison is a production-like evaluation with fixed prompts, fixed tools, a disclosed effort setting and an all-in bill that includes search, containers and cache misses.