Claude Opus 5 Lands: The Frontier AI Race Is Now a Price War

Anthropic released Claude Opus 5 yesterday, its fourth model in under two months, pitched as coming close to the frontier intelligence of Fable 5 at half the price. The individual release matters less than the pattern: every major lab has stopped competing purely on capability and started competing on what intelligence costs.

What actually shipped

Opus 5 takes over as the everyday flagship: state-of-the-art results on coding and knowledge-work benchmarks like Frontier-Bench and GDPval-AA, the default model on Claude Max and the strongest on Pro, while still trailing the restricted Mythos 5 on offensive cybersecurity [1] [2]. Pricing holds at $5 per million input tokens and $25 per million output, unchanged from Opus 4.8, so the improvement arrives free, and a Fast mode runs about 2.5 times quicker at double the price [3] [1].

The tell is where the new features point. There's an effort toggle, low, medium, or high, that lets users trade capability against cost per task, a direct answer to enterprise complaints about runaway AI bills [4]. The model is built to need less back-and-forth, verify its own work, and recover from errors without hand-holding, which is efficiency framed as intelligence [4]. And two beta platform features ship alongside: developers can change a conversation's available tools without invalidating the prompt cache, and API requests flagged by safety classifiers can auto-route to another model instead of being blocked [1] [3]. Every one of those is a cost or reliability feature, not a capability flex.

Why efficiency suddenly leads the spec sheet

Context explains it. Fable 5, the model Opus 5 undercuts, drew loud criticism for its burn rate: it's brilliant, and it eats token budgets like a teenager eats groceries, with business customers blowing through allotments and running up bills [4]. Opus 5 also skips Fable's mandatory 30-day data retention, removing a second enterprise objection [5]. The message underneath the launch is that the premium model exists for days-long autonomous projects, and this one exists for the invoice you can predict.

The competition is telling the same story. OpenAI shipped its GPT-5.6 family earlier this month with its own emphasis on economical token use [5] [4]. China's Moonshot released the open-weight Kimi K3 last week, claiming near-parity with both Fable 5 and GPT-5.6's top variant, at open-weight prices, which is to say approaching free for anyone with hardware [5]. And both Anthropic and OpenAI filed confidentially to go public, Anthropic in June with a listing possibly this year, which means every one of these releases is also an argument to future shareholders about unit economics [5].

When every lab's headline feature is a cost dial, the race has changed. Peak capability is table stakes; the fight is over what a completed task costs.

The read for web professionals

Three practical takeaways. First, if you locked an AI toolchain around one model's pricing more than a month ago, your math is stale: near-flagship capability at half price resets what content generation, audits, and agent workflows should cost you, and this will keep happening. Second, the metric that matters is cost per completed task, not per-token sticker price; effort dials, burn rates, and tokenizer changes all live in the gap between those two numbers, which is exactly where the Sonnet 5 launch hid its September price step-up. Third, build for substitution. The labs themselves now ship automatic fallbacks that swap models mid-request; workflows hardwired to a single model are more fragile than the vendors' own infrastructure assumes. Fable 5's 19-day government-ordered disappearance last month made the same point less politely.

Cheap, swappable intelligence is the environment now. Tools priced as if the model behind them were scarce are the ones the market keeps repricing, which is a dynamic we have written about before and build our own pay-per-use pricing around.