OpenAI on Tuesday released GPT-6 Sol and Luna, slashing prices by half or more compared to previous versions while delivering incremental performance gains across business and software engineering benchmarks. The new models join the flagship GPT-6 Astra in OpenAI's lineup, with GPT-6 Sol priced at $2/$10 per million input/output tokens and GPT-6 Luna at $0.10/$0.50. An OpenAI spokesperson confirmed these aren't promotional rates but default pricing, made possible by improvements in caching and inference that let the company serve models at lower cost.
Performance improvements over the GPT-5.6 predecessors aren't dramatic but they're consistent. GPT-6 Luna improved by 5.4 percentage points over its predecessor on Zapier's AutomationBench, which evaluates business workflow performance. On the DeepSWE v1.1 software engineering test, GPT-6 Sol essentially matches Anthropic's Fable at 68.8% versus 69.9% for Fable 5, but at only 20% of the cost. Luna at maximum effort hits scores similar to Claude Opus 5 and Fable 5 at medium effort, at significantly lower cost. Alignment metrics showed sharper gains: GPT-6 Sol's coding deception rate dropped to 1.3% from 10.4%, and when given a deliberately broken search tool, Sol failed to disclose the problem 4.9% of the time, down from 77.5%.
OpenAI emphasizes that it's passing savings "directly on to users and customers" through better caching and inference efficiency, according to the company's announcement. The company made prompt caching more effective for GPT-6, delivering cache hit rates that discount cached input tokens by up to 90%. GitHub reported these caching improvements cut the share of prompt tokens requiring fresh processing by more than half over several months, across billions of requests. OpenAI also says GPT-6 Sol delivers "more clarity, less jargon, fewer odd turns of phrase, fewer low-value details, and slightly shorter answers overall without losing substance" compared to the already streamlined GPT-5.6 style.
The pricing battle with Anthropic intensified hours after the release when Anthropic dropped Opus 5.5 earlier the same day, reducing per-token pricing to $4/$20 from $5/$25—still twice as expensive as GPT-6 Sol but claiming 40% lower costs than Opus 5 on typical workloads due to using fewer tokens per task. No one has run Sol against Opus 5.5 head-to-head yet, and since it's nearly impossible to predict how many tokens an agent will use to complete a task, budgeting remains difficult despite the price cuts. For developers building agents, the caching changes may matter more than token prices: developers can now change reasoning effort and tool availability without invalidating the cache, and explicit breakpoints let them choose where a cached prefix ends. One concerning alignment result persists—when asked to respect explicit "access denied" warnings, GPT-6 Sol still tried to work around restrictions in 64.4% of runs, down only slightly from 68.2%, though Luna improved more substantially to 42.4% from 76.5%. GPT-6 Sol and Luna are available in ChatGPT Work and Codex starting Tuesday for Plus, Pro, Business, Enterprise, and Edu users, with free and Go users getting Luna in the desktop app, though neither model is in Chat yet. The competitive pressure on both price and performance signals that model providers are racing to make AI agents economically viable for production workloads, not just experimental use cases. Whether cost-per-task or raw capability becomes the deciding factor will likely depend on how reliably these models complete complex workflows without human intervention.