The Launch and the Pricing
OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026, expanding a GPT-6 line that began earlier that month with Astra, which the lab had billed as its most powerful model and the "world's best model" for computer work and coding. Sol is priced at $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 and $0.50 on the same measure. Both sit at half the promotional rates of their predecessors, a cut sqmagazine.co.uk confirms at the API level.
Sol's rate is exactly half of Opus 5.5's and one fifth of Astra's, according to eu.36kr.com. Luna drops to one percent of Astra's price. OpenAI attributes the reductions to improvements in caching and inference rather than to a cheaper training run — the models were trained with methods similar to the flagship, per finance.biggo.com. The stated targets are professional work, coding, automation and computer-use tasks, the same categories Astra was promoted against.
Astra remains the top of the GPT-6 series. OpenAI's guidance, as reported by eu.36kr.com, is that tasks which are critical, highly complex and intolerant of capability compromises still belong on the flagship. That positions Sol and Luna as volume tiers rather than replacements: the lab keeps a premium ceiling while cutting the floor price by two orders of magnitude. The structure is standard tiering; the depth of the discount is not.
The 90-Minute Rivalry with Anthropic
Anthropic released Claude Opus 5.5 priced 20 percent below its predecessor. Ninety minutes later, OpenAI launched Sol and Luna. Both finance.biggo.com and eu.36kr.com record the gap and read it the same way: a deliberate undercut, executed before buyers could benchmark either release. The timing compresses the usual evaluation window to nothing. A procurement decision that would normally take weeks of testing now gets made against press-release numbers.
The two companies are now benchmarking against each other's previous-generation models, per finance.biggo.com — OpenAI pricing Sol against Opus 5.5, Anthropic pricing Opus 5.5 against its own prior release. That is a shift from capability competition to rate-card competition. The products differ; the numbers are the message. When a lab prices its new mid-tier at exactly half a rival's flagship, the claim being made is about cost per unit of work, not intelligence.
eu.36kr.com's account is blunt: throughout the night, before model capabilities could be fully verified, a price war had already broken out. The open question is whether the 90-minute gap reflects preparation or reaction. OpenAI had Astra in the market since earlier in the month, so the cheaper models plausibly existed before Anthropic moved. The launch date, however, is a choice, and it was made the same evening.
What OpenAI Claims About Performance and Cost
OpenAI reports that Sol achieves roughly half the error rate of its predecessor, per finance.biggo.com. techcrunch.com carries the same claim in its headline — lower cost and fewer mistakes. If both hold, the price cut is not a quality trade: a model costing half as much and failing half as often changes the cost-per-correct-answer calculation by more than either figure alone. That is the arithmetic the pricing is built on.
The cost side rests on engineering, not scale. OpenAI cited improvements in caching and inference as the basis for the reductions, per finance.biggo.com. Caching lowers the effective price for repeated context; inference gains lower it for everything. Neither is a claim about training efficiency, and the evidence does not say what Sol and Luna cost to run. The published numbers describe margin policy as much as technology.
The record is silent on independent verification: no source in the evidence set publishes third-party benchmarks for Sol or Luna, and none reports a quality comparison against Opus 5.5 at matched tasks. What exists is OpenAI's error-rate figure, the pricing grid, and the stated use cases. Buyers pricing workloads can act on the numbers. Buyers assessing capability cannot yet, and the 90-minute launch window guaranteed it would stay that way for a while.
Sources
- techcrunch.com — OpenAI Launches GPT-6 Sol and Luna at Half the Price — 90 Minutes After Anthropic's Answer
- biggo.com — OpenAI Slashes GPT-6 Prices With Sol and Luna, Undercutting ... (2026-09-22)
- co.uk — OpenAI Launches GPT-6 Sol and Luna at Half the API Price (2026-09-22)
- 36kr.com — GPT-6 Sol and Luna are launched, offering even steeper discounts ... (2026-09-23)
- New Horizon daily digest, 2026-09-23 (topic provenance)