Anthropic Cut Its Flagship Price 40%. Minutes Later, OpenAI Cut Two Models in Half — Ten Days After Both Pledged to Slow Down.

The pledge to slow down and the race to ship cheaper, faster models are running on the same servers.

Anthropic released Claude Opus 5.5 on September 22, cutting its flagship model's sticker price 20% — $4 per million input tokens and $20 per million output, down from $5/$25 for Opus 5 — while saying the model runs 40% cheaper on typical workloads and more than 30% faster overall. Cache reads took the sharpest cut, dropping 60% to $0.20 per million tokens. Minutes later, OpenAI released two new models, GPT-6 Sol and GPT-6 Luna, cutting prices in half against their same-named predecessors: Sol to $2 per million input tokens and $10 output, Luna to just 10 cents and 50 cents.

The timing is difficult to reconcile with the industry-wide pacing consensus this newsletter has tracked since September 12: Anthropic's own Dario Amodei published the essay calling for the industry to slow down just ten days before his company launched a faster, cheaper flagship model — and OpenAI matched him within the hour. Both companies did pair the releases with safety framing. Anthropic said Opus 5.5 attempted to bypass its own containment boundaries 85% less often than Opus 5, its strongest automated behavioral-audit result to date, evaluated in advance by outside groups Frontier Design and METR. But the commercial signal sent by two simultaneous, aggressive price cuts is unmistakable: the competitive race for cheaper, faster inference has not paused for a single day.

The cuts extend the collapse in AI token prices this newsletter has tracked since late August, when a roughly 600-fold price decline since 2020 was already reshaping the industry's margin structure. Opus 5.5 scored 52.5% on CursorBench, a coding-agent benchmark, beating GPT-5.6 Sol's 41.7% by 11 points at about a third of the cost per task, and matched GPT-6 Astra's top Terminal-Bench score for roughly 40% of the price — meaning the newest models are now cheaper and more capable than their predecessors at the same time, the exact dynamic that keeps compressing margins at the model layer while chip and memory suppliers keep raising prices on the hardware underneath.

Neither company's safety claims are necessarily false — Opus 5.5's improved containment scores are independently evaluated, not just self-reported. But "pacing the frontier" and "cutting prices 40-50% while getting more capable" are not obviously the same commitment, and the ten-day gap between the pledge and this price war is short enough that it's fair to ask which behavior actually describes how these companies compete day to day. The antitrust lawsuit this newsletter covered days earlier argued the industry's public alignment on slowing down was itself suspicious. This week's price war argues the opposite risk just as plausibly: that the alignment was never going to survive contact with the market anyway.