OpenAI halves the API price of Sol and Luna against GPT-5.6's promotional rates

Somewhere in a backlog there is a feature with a note beside it: too expensive to run at 5.6 prices. Re-costing that note takes four numbers, and the announcement supplies one of them cleanly.

List price is the clean one. OpenAI says it reduced API prices for Sol and Luna by 50% against GPT-5.6 promotional pricing, crediting improvements in caching and inference ^[1]. Tokens per task is the second number, and the ChatGPT desktop release notes say the GPT-6 models use fewer of them than the GPT-5.6 models of the same class ^[14]. Third is which tier runs the work. Fourth is the effort setting, and the large ratios come from there: the hundredfold figure for Luna is qualified "at higher effort levels" ^[23], and on OSWorld 2.0 offline the pairing OpenAI published is Luna at max against GPT-5.6 Sol at medium ^[7]. A team that takes only the price cut gets half. A team that gets tenfold or hundredfold has also changed tier and effort, and both of those land on output quality.

The same variability runs through Sol's numbers against Anthropic. On AutomationBench, OpenAI reports Sol scoring 6.3% better than Claude Opus 5 at 9% of the cost per task ^[15]. The OSWorld 2.0 offline figures put Sol at xhigh effort on 60.5% against Opus 5 at medium on 60.3%, at roughly 80% lower cost per task ^[8]. Those are cost reductions of 91% and about 80% against the same competitor model ^[20]. DeepSWE 1.1 went the other way, where Sol did not beat Claude Fable 5 and landed within 1.1 percentage points of Fable's highest score ^[18]. ZDNET, whose parent Ziff Davis sued OpenAI in April 2025 over copyright in the training and operation of its systems ^[12], notes that OpenAI is benchmarking against its competitors' generationally previous version ^[9].

The reliability claim comes with its own caveat. TechCrunch quotes the announcement: "On our internal factuality evaluation, which is based on de-identified real-world conversations where users flagged mistakes by our models, GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability at much lower cost" ^[4]. OpenAI did not publish the underlying error rate, so the halving is a ratio without a level. ZDNET illustrates it with a hypothetical 20% dropping to 10% ^[5].

OpenAI's own bill points the same way as the pricing. The company says daily token usage, valued at API prices, has exceeded $600 for the median researcher and $7,000 for researchers at the 90th percentile ^[19]. The 90th percentile spends about 11.7 times the median ^[21]. At 22 working days, the median researcher runs about $13,200 a month, and halving prices takes roughly $6,600 off that ^[22].

OpenAI describes Luna's jobs as "high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions" ^[10], which is most of what gets shelved on cost. The question for the shelved feature is whether Luna, at an effort setting the budget allows, clears the error rate the team's users already tolerate, measured on the team's own tasks. The 50% price cut is the only number in that calculation nobody has to measure ^[1].