OpenAI Launches GPT-6 Sol and Luna: API Prices Cut in Half, and a New Fight Over Cost per Task

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, cutting API prices by 50% relative to the GPT-5.6 promotional pricing. Sol drops from $4 to $2 per million input tokens and from $20 to $10 per million output tokens, while Luna drops from $0.20 to $0.10 for input and from $1.20 to $0.50 for output. Both models were trained with methods similar to GPT-6 Astra, released earlier this month, and OpenAI reports that on the AutomationBench workflow benchmark Sol beats Claude Opus 5 at max effort at 9% of its cost per task. ASAP reproduces the price table and benchmarks from the official announcement, then reads separately what the comparisons establish and what they leave out.

Two rows of pricing are the real headline of this launch

OpenAI itself frames GPT-6 Sol and Luna not as new frontier models but as Astra's training advances carried down into cheaper tiers. The announcement says the most demanding projects "still call for Astra's full depth" and that Sol and Luna advance "the frontier on cost efficiency." The price cut is attributed to caching and inference improvements that lowered serving costs, with the savings passed directly to customers.

Model	Input (per 1M tokens)	Output (per 1M tokens)	OpenAI label
GPT-5.6 Sol → GPT-6 Sol	$4 → $2	$20 → $10	50% cheaper
GPT-5.6 Luna → GPT-6 Luna	$0.20 → $0.10	$1.20 → $0.50	50% cheaper

Both models are rolling out gradually in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users starting on launch day. Free and Go users get Luna in the desktop app, and neither model is yet available in Chat. In the API they are named gpt-6-sol and gpt-6-luna.

Luna's output price actually fell by more than 58%

OpenAI labels both rows "50% cheaper," but Luna's output price moved from $1.20 to $0.50, a reduction of roughly 58.3%. In agent workloads the cost driver is usually long outputs and reasoning tokens rather than input, so for teams calling Luna at volume, 58% may be closer to the felt reduction than the headline 50%. Luna's input price is cut by exactly half.

The table also contains a caveat running the other way. The baseline is GPT-5.6's "promotional pricing," not a list price, as the announcement states. Organizations that used GPT-5.6 under different terms may see a different reduction, so the accurate baseline is the unit price on their own invoice. A "percent cheaper" claim only means something alongside the baseline it is measured from.

OpenAI moved the comparison unit from token price to cost per task

Nearly every benchmark sentence in this announcement pairs a score with a cost per task, which sets it apart from earlier model launches. On AutomationBench 1.0.6, Sol at xhigh effort scores 33.2% at $0.27 per task; on the same table Astra at low effort scores 30.3% at 3.9 times Sol's cost, and Claude Opus 5 at max effort scores 26.9% at 11.1 times the cost. The benchmark tests end-to-end workflows using 47 tools across sales, marketing, operations, support, finance, and HR.

The other rows follow the same pattern. On DeepSWE v1.1, Sol at max effort scores 68.8%, within 1.1 percentage points of Claude Fable 5's best score of 69.9% at xhigh, at approximately 80% lower cost per task. Luna at max effort scores 66.6%, comparable to Opus 5 and Fable 5 at medium effort, at 93% lower cost than Opus 5 and 96% lower than Fable 5. On the OSWorld 2.0 offline set, Sol at xhigh reaches 60.5% against Opus 5 at medium effort with 60.3%, again at roughly 80% lower cost. On Agents' Last Exam, Sol at max effort scores 56.4%, above Opus 5's highest score, at 60% lower cost per task.

Cost per task is closer to operational reality than token price. A cheap token rate can still be expensive if a model burns several times more tokens per job, and the reverse also holds. OpenAI putting this unit up front signals that price competition is shifting from "how much per million tokens" to "how much to finish one piece of work." Cost per task, however, depends heavily on the length and shape of benchmark tasks, so it does not transfer to a team's own work without re-measurement on its own task mix.

Many comparisons pair different effort settings

Collected together, the comparison pairs often put Sol and Luna at high effort (xhigh or max) against competitors at medium or a specific setting. OSWorld 2.0 is Sol xhigh against Opus 5 medium; DeepSWE is Luna max against Opus 5 and Fable 5 at medium. That pairing answers "who reaches the same score more cheaply," which is a fair question, but it does not answer "who does better at equal effort."

The announcement also spells out its measurement caveats. Competitor numbers come from public reports, and Fable 5 scores were used where Fable 5.1 scores were unavailable. OpenAI notes that the AutomationBench row for "Fable 5.1 with Opus 5 fallback" understates true cost because it omits the Opus 5 fallbacks that occurred on about 40% of tasks, so that cell is marked only as ">8.9x." OpenAI models were evaluated in a research environment or via the API, which may differ slightly from production ChatGPT. Disclosing these footnotes is to OpenAI's credit, but the fact remains that cells in the same table come from different measurement conditions.

The 90% cache discount is the real variable in agent costs

GPT-6 prompt caching is the second cost lever in this launch: OpenAI raised default cache hit rates and applies a 90% discount on cached input-token reads. The additions are a Prompt Caching Dashboard, a diagnostics tool that explains missed caching opportunities, a change that preserves earlier context for cache reuse when reasoning effort or available tools change, and explicit breakpoints that let developers choose where cached prefixes end. GitHub reports that over the past several months these improvements cut the share of prompt tokens requiring fresh processing by more than 50% across billions of requests.

This item is less visible than the price table but can matter more to agent operating costs. Agents resend the same system prompt, tool definitions, and accumulated conversation on every call, so much of their input repeats. Fixing the problem of cache breaks on effort changes in particular makes it possible to route easy follow-ups to low effort and hard steps to high effort without paying a cache penalty. For teams adopting Sol and Luna, the first metric to watch is their own workload's cache hit rate, not the token price.

What teams can decide now and what should wait

The decision available immediately is swapping pipelines that ran GPT-5.6 Sol or Luna: changing the model name to gpt-6-sol or gpt-6-luna cuts unit prices by half or more. OpenAI reports that Sol makes about half as many factual mistakes as its predecessor, and that Luna at higher effort matches GPT-5.6 Sol's factuality at about a hundredth of the cost. The announcement adds that this factuality evaluation is built from de-identified conversations where users flagged errors, so it does not represent error rates in typical use.

Other decisions should wait. All benchmarks in the announcement are English-language, with no figures on non-English document handling or factuality. And because competitor comparisons pair mismatched effort settings, a team on Claude models that wants to conclude "switching to Sol gets the same quality for less" needs to re-measure under matched conditions on its own tasks. OpenAI's own figure that daily token usage, valued at API prices, exceeds $600 for its median researcher and $7,000 at the 90th percentile shows coding-agent spend already rivals labor costs, which makes monthly cost simulation a better basis for model choice than a benchmark table.

Open questions

The first open question is the gap between Astra and Sol. OpenAI reports that Sol at xhigh beats Astra at low effort on AutomationBench, meaning a smaller model at high effort can outscore a larger one at low effort within the same family. If so, the case for choosing Astra may be narrower than "when you want the best results," and there is not yet a map of which tasks favor Sol at high effort over Astra at low effort.

The second is how much alignment data is public. The announcement says Sol and Luna improve on their GPT-5.6 counterparts, including lower rates of misleading claims about their coding work, but defers the numbers to the system card. The third is independent replication. Cost per task is sensitive to evaluation environment and harness, and whether gaps like 80% and 93% hold when outside evaluators run the same benchmarks at matched effort will be the real test of this launch.

Source: OpenAI, "Introducing GPT-6 Sol and Luna" (openai.com/index/introducing-gpt-6-sol-and-luna, September 22, 2026). ASAP verified the price table, benchmark statements, and caching and availability details against the announcement directly.