Anthropic and OpenAI cut their frontier prices on the same day - the 40 and the 50 are the sellers' numbers

On Tuesday 22 September Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5, with cache reads at $0.20 against $0.50. Anthropic says that at default settings it costs 40% less than Opus 5 on typical workloads and writes output more than 30% faster; Sonnet 5.5 and Haiku 5.5 follow in the coming weeks. The same day OpenAI released GPT-6 Sol at $2 in and $10 out, and GPT-6 Luna at $0.10 and $0.50, which it calls 50% cheaper than the GPT-5.6 versions' promotional pricing; its business pricing page carries those figures today.

What the pages say

Anthropic's per-token cut is 20% on input and output and 60% on cache reads; its 40% is "at default settings ... on typical workloads" and, by its own account, "nets out" from lower prices plus fewer tokens per task. OpenAI gives the GPT-5.6 base as $4 and $20 for Sol and $0.20 and $1.20 for Luna, names that base as promotional, and does not give a non-promotional 5.6 price. Anthropic's page also says that "benchmark margins have become a less reliable guide to real-world differences" and that in its own use the gap between Opus 5.5 and Claude Fable 5.1 "is narrower than these scores suggest." Each lab's comparison table takes the rival's figures from the rival's published reports.

The check

List prices fell - by a fifth at Anthropic on input and output, and by half against a discount at OpenAI. The same token counts on Opus 5.5 come in 20% under Opus 5 on a plain workload and about 26% under on a cache-heavy one; the rest of Anthropic's 40 is tokens per task, which only a buyer's own usage can show.