Anthropic shipped Claude Opus 5.5 on September 22, and roughly 90 minutes later OpenAI released GPT-6 Sol and GPT-6 Luna at half the token pricing of their GPT-5.6 predecessors. It was the sharpest single-day compression of frontier serving costs in 2026, and it happened on a Tuesday afternoon.

The numbers do the work. Sol lists at $2 per million input tokens and $10 per million output; Luna, the smaller sibling, at $0.10 and $0.50. Anthropic's Opus 5.5 came in at $4 and $20, a 20% per-token cut from Opus 5 that Anthropic says lands at 40% less on typical workloads once shorter outputs are factored in. Cache economics moved even harder: Anthropic cut Opus 5.5 cache reads by 60% to $0.20 per million tokens, and OpenAI is discounting cached input reads by 90%, with hit rates improved so coding agents can change reasoning effort and tool sets without breaking the cache.

Benchmarks arrived with the price cuts, as they always do. Opus 5.5 posts 66.4% on Terminal-Bench 4.0 against Fable 5.1's 55.8%, and jumps to 40.0% on AutomationBench from Opus 5's 26.9%. It runs over 30% faster than its predecessor. OpenAI's counter-metric is per-task cost: Sol at xhigh effort clocks in at 9% of Opus 5's per-task cost on AutomationBench. Both companies, in other words, are arguing that the cheaper model is also the more capable one, which is what a serious price war sounds like.

The distribution layer moved in lockstep. Sol and Luna landed inside GitHub Copilot the same day, across VS Code, JetBrains, and Xcode, with Sol available on Copilot Pro+, Max, Business, and Enterprise, and Luna extended down to Copilot Pro. GitHub's Mario Rodriguez, quoted in the Anthropic release, said Opus 5.5 "solved more terminal tasks than Opus 5 in less than half the steps" inside VS Code. Anthropic also flagged Sonnet 5.5 and Haiku 5.5 as forthcoming; Opus 5.5 is already live across the Claude Developer Platform, AWS, Google Cloud, and Azure.

This is the second frontier-pricing event of the month. xAI's Grok 4.7 shipped at 2.1T parameters and the same $2/$6 line earlier the same day, following Anthropic's Fable 5.1 nearly doubling its AutomationBench score and the September 3 launch of GPT-6 Astra with computer use and agentic professional work. Three weeks, four price and capability moves, all in the same direction.

The reflex reference is the 2015–2016 AWS-Azure-Google Cloud price war, where the cost of a compute-hour halved in eighteen months and the vendors who had priced products on year-old assumptions found their margins gone. That's the position any AI product built on August pricing is in this morning. The floor moved. The question is whether the ceiling did too.

Sources

https://openai.com/index/introducing-gpt-6-sol-and-luna/
https://www.anthropic.com/claude-opus-5-5
https://techcrunch.com/2026/09/22/openai-launches-gpt-6-sol-and-luna/
https://siliconangle.com/2026/09/22/anthropic-releases-claude-opus-5-5-and-openai-counters-with-two-cheaper-gpt-6-models/
https://github.blog/changelog/2026-09-22-openais-gpt-6-sol-and-gpt-6-luna-now-available/