OpenAI released GPT-6 Sol and GPT-6 Luna last night, hours after Claude Opus 5.5. Both cost half what their predecessors did in the API. Sol is $2 in and $10 out per million tokens; Luna is 10 and 50 cents.
GPT-6 Astra, out earlier this month, remains OpenAI’s best model. Sol and Luna are meant to carry Astra’s capabilities into cheaper tiers, trained with similar methods.
Nine percent of what Opus 5 costs
Almost every comparison OpenAI draws points at Claude. On AutomationBench, a test of business workflows across 47 tools, Sol at xhigh effort scores 33.2 percent at 27 cents per task. Opus 5 at max effort scores 26.9 percent and costs 11.1 times as much, by OpenAI’s numbers. Fable 5.1 with an Opus 5 fallback reaches 31.4 percent, and OpenAI notes that this figure leaves out the cost of the fallbacks, which fired on roughly 40 percent of tasks.
On Agents’ Last Exam, Sol at max effort hits 56.4 percent, above Opus 5’s best result in that evaluation at 60 percent lower cost per task. On DeepSWE v1.1, Sol reaches 68.8 percent, within 1.1 points of Claude Fable 5’s best score, at about 80 percent lower cost. Luna lands at 66.6 percent there and costs 93 percent less per task than Opus 5.
Worth saying out loud: every one of those numbers comes from OpenAI, and they benchmark against Opus 5 and Fable 5.1, not Opus 5.5. Anthropic does the mirror image, measuring itself against GPT-6 Astra and GPT-5.6 Sol. Two labs shipped on the same day, and each used the other’s previous flagship as the yardstick.
Half as many mistakes
The factuality work interests me more than the benchmarks. OpenAI measures it on de-identified real conversations where users flagged an error from an earlier model. Sol makes about half as many mistakes as its predecessor. Luna at higher effort reaches GPT-5.6 Sol’s level at roughly one hundredth of the cost.
The tone got attention too: less jargon, fewer low-value details, slightly shorter answers. The example in the post shows two replies to the same prompt, and the newer one says what it will check before it starts. Anthropic describes almost exactly the same improvement in Opus 5.5.
The second price cut hides in the caching
Alongside the token prices, OpenAI reworked prompt caching. Hit rates go up by default and cached input tokens cost 90 percent less. There is a dashboard, a diagnostics tool, and explicit breakpoints so developers decide where a cached prefix ends. Effort level and tool availability can now change without blowing the cache.
One figure from the post sticks. Inside OpenAI, daily token usage runs above $600 for the median researcher and above $7,000 at the 90th percentile, valued at API prices. When your own staff burns that much, halving the price is not generosity. It is accounting.
In one day, cost per task became the main argument
Two frontier labs, one day, two announcements, and both are mostly about price. Anthropic cuts cache reads by 60 percent. OpenAI halves token prices and gives 90 percent off cached input. New capabilities come up afterwards, in both cases.
If you run agents for hours, that is the actual news. Several models can now do the job. What decides it is what an attempt costs, and that number fell in two places at once yesterday.