Ten days ago the heads of the largest AI labs asked the industry to slow down. On 22 September two of them shipped new frontier models ninety minutes apart, and both cut the price.

Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output, down from $5 and $25, with cache reads cut by 60% to $0.20 and output generated more than 30% faster than Opus 5. Anthropic says it performs at the level of Claude Fable 5.1 on most work while costing about 40% less to run, reports 66.4% on Terminal-Bench 4.0 against 52.3% for Opus 5, and says it attempts to work around its boundaries 85% less often. About an hour and a half later OpenAI released GPT-6 Sol at $2 and $10 and GPT-6 Luna at $0.10 and $0.50, each half the price of the GPT-5.6 tier it replaces, both with a 1.05M token context window. OpenAI says Sol gets 4.6% of answers wrong at maximum effort where GPT-5.6 Sol got 8.5% wrong, and told VentureBeat the prices are permanent, not promotional.

Why this one is different

Frontier releases usually arrive at the price of the model they replace, or above it, and the cheaper tier arrives months later. This is the first time both leading labs cut the headline price of their flagship tier on the same afternoon. It is also the first release week where the pitch was not a new capability. Anthropic's own sentence for Opus 5.5 is that it matches the model above it and costs less; OpenAI's is that the same work costs half.

Nobody announced a capability. Both announced a discount.

How we got here

1. 12 Sep 2026Amodei, Altman and Musk all call in public for the frontier to be paced.
2. 17 Sep 2026Anthropic publishes the first measurement of how much of its own R&D Claude leads.
3. 18 Sep 2026Anthropic hires an embedded evaluator; California studies requiring one by law.
4. 22 Sep 2026Claude Opus 5.5 at $4 and $20, then GPT-6 Sol and Luna at half the old tier price.

What it does and does not mean

Lower list prices are not lower spending, and neither company published what a task costs rather than what a token costs. A model that is cheaper per token and used for longer agent runs can raise the bill, which is the pattern the last two years have followed every time. The benchmark numbers are each lab's own, measured under each lab's own harness, so the 66.4% and the 58.2% currently on the public Terminal-Bench board are not the same measurement. What is plain is the direction of the market. Two weeks after agreeing in public that the frontier should be paced, the two labs that said it competed on price on the same day, which is what competition does to a voluntary agreement that has no mechanism in it.