Anthropic Cut Opus 5's Price in Half. Here's the Trick.

Anthropic just cut the price of its best model in half, and the interesting part isn't the discount. It's how they did it.

On July 24, Anthropic shipped Claude Opus 5 at $5 per million input tokens and $25 per million output — the same rate as the outgoing Opus 4.8, not a premium increase you'd expect from a flagship refresh. Compare that to Fable 5, the model Anthropic released as its top-of-line option back in June: $10 in, $50 out. Opus 5 gets you roughly Fable-5-level intelligence, according to Anthropic's benchmark claims, at half that cost. Somewhere in Anthropic's model-training pipeline, the frontier got cheaper to run. That's the actual story.

Here's the mechanism, as far as the released details make it possible to reconstruct. Opus 5 ships with a low/medium/high effort toggle, a parameter that trades inference cost for reasoning depth on a per-request basis. Low effort is fast and cheap, suitable for the bulk of workaday requests — drafting an email, summarizing a document, answering a factual question. High effort spends more compute per token generated, working through a problem the way Fable 5 does by default on every request, whether it needs to or not. Fable 5 doesn't have this knob. It runs at one intensity, always, which is a large part of why it's expensive: you're paying premium rates even when the task doesn't require premium reasoning.

This is the same insight that's been reshaping every corner of the frontier-model market for the past year: most queries don't need the full model. Google split Gemini into Pro and Flash tiers for exactly this reason. OpenAI's GPT-5.6 line ships with its own reasoning-effort parameter. What's different about Anthropic's approach is that they didn't build effort-tiering as separate models with separate names and separate release cycles. They built it as a runtime parameter on one model. You don't choose between "Opus 5 Lite" and "Opus 5 Pro" in a pricing table. You set effort=high in the API call and pay for what you asked for.

There's a second lever here too: a "fast mode" that doubles the per-token rate to $10/$50 (Fable 5's exact pricing, notably) in exchange for roughly 2.5x lower latency. That's a separate axis from effort — you can run a low-effort request in fast mode when you need a quick, simple answer instantly, or a high-effort request in standard mode when you're willing to wait for a harder problem solved well. Two dials instead of one flat price point is a meaningfully different pricing architecture than what Claude offered a year ago, when a model was a model and you paid its listed rate regardless of what you asked it to do.

Picture a customer support tool built on Claude. Most tickets are routine: password resets, order status, plan questions. Those go out at low effort, standard mode, at the cheapest rate Anthropic offers. The rare ticket that requires the model to read a contract, cross-reference three prior support threads, and draft a carefully worded escalation response gets bumped to high effort. Under the old one-size model, that entire ticket volume ran through Fable 5's rate card regardless of which bucket each ticket actually fell into. Under Opus 5, the easy 90% costs a fraction of what it used to, and the hard 10% still gets the reasoning it needs. That's not a marginal efficiency gain. At scale, it changes which applications are economically viable to build on Claude at all.

The numbers behind this matter for anyone building on the API, not just enthusiasts comparing benchmark charts. Opus 5 carries a 1-million-token context window and a 128,000-token maximum output, with training data through May 2026, per the full spec breakdown. Reported benchmark gains cluster around agentic coding and long-horizon, multi-step tasks — the categories where a model has to hold a plan in its head across dozens of tool calls without losing the thread. That's a specific kind of capability, and it's not the same thing as "smarter at everything." A model can post state-of-the-art numbers on agentic coding while being unremarkable at, say, creative writing or casual conversation. The marketing rarely draws that distinction; the benchmark suites usually do, if you read past the headline number.

None of this means Fable 5 is obsolete, and it doesn't touch Sonnet 5, Anthropic's mid-tier default, which occupies a different price point entirely. Fable 5 still exists, still costs twice as much, and Anthropic hasn't discontinued it. What Opus 5 does is remove the reason most developers had to reach for Fable 5 in the first place: needing frontier-tier reasoning without wanting to pay a flat premium on every single call, including the easy ones. If your workload is genuinely uniform in difficulty, a single fixed-rate model is simpler to reason about. If it's bursty and mixed, like most real applications actually are, a tunable model wins on cost without giving up ceiling.

The bigger pattern worth watching is what this signals about where the price-per-intelligence curve is headed. A year ago, "the best model" and "the expensive model" were the same sentence. Anthropic just spent a release proving that gap isn't structural, it's a temporary state that gets closed the moment a lab figures out how to serve the same intelligence more efficiently. Expect Fable 5's price to look less like a permanent ceiling and more like a marker of how good frontier models were before someone worked out how to make them cheaper. For a full rundown of what Claude currently offers across the lineup, the model's profile is the place to check what's changed since.