When Anthropic shipped Claude Opus 5 back in July, it kept the exact same price tag as the model before it. That streak ended on September 22. Opus 5.5 launched with an immediate 20% price cut — input tokens down to $4 per million, output down to $20, and cache reads slashed by 60%. Because the new model burns through fewer tokens to finish the same task, actual bills for developers tend to drop by around 40%.
Anthropic says Opus 5.5 outperforms Fable 5.1 — a pricier model the company shipped just three weeks earlier — on the benchmarks that matter for coding agents: 66.4% versus 55.8% on Terminal-Bench 4.0, and a higher Elo score on GDPval-AA. Box VP Yashodha Bhavnani said the model's answers came back 40% shorter without losing accuracy, while GitHub's chief product officer, Mario Rodriguez, noted it solved more terminal tasks than Opus 5 in less than half the steps.
OpenAI wasn't about to let that pass. The same day, it rolled out two new models, GPT-6 Sol and GPT-6 Luna, both priced at half of what their predecessors cost. Sol now runs $2 per million input tokens and $10 for output; Luna comes in at $0.10 and $0.50. OpenAI told reporters these are permanent rates, not an introductory discount, and cached-token reads got a 90% cut on top of that.
The timing isn't a coincidence. Anthropic is heading toward a stock listing that could land within weeks, and showing the market an efficiency curve bending downward matters more right now than another point on a leaderboard. OpenAI, meanwhile, is telling investors it expects to spend $750 billion on compute by 2030 — cutting the price per token isn't charity, it's a bet that usage will grow faster than revenue per request shrinks.
The two companies used to fight over benchmark scores. Now the argument has shifted to what it actually costs to get a task done — not abstract intelligence, but a finished code migration or a closed support ticket. For any business paying for these APIs, the takeaway is simple: frontier model prices are falling faster than they climbed over the past two years, and the real competition between Anthropic and OpenAI is no longer just about who answers better, but who finishes the job cheaper.
