Last month I argued that the price of intelligence is collapsing even as AI bills climb. Last week both halves of that argument got sharper on the same four days. On July 24, Anthropic shipped Claude Opus 5 at exactly the same list price as Opus 4.8 — $5 per million input tokens, $25 per million output — while positioning it as “near Fable 5 intelligence at half the price.” Three days later, Moonshot dropped Kimi K3’s weights and made near-frontier capability free to anyone with the racks to serve it. The squeeze is now coming from both ends: the closed frontier is halving its own prices, and the open ecosystem is commoditizing the tier just below it.
What actually shipped
- Same SKU, different model. Opus 5 replaces Opus 4.8 at an unchanged $5/$25. No new pricing tier — the capability jumped underneath a fixed price point. On Anthropic’s published numbers, the deltas are not subtle: Frontier-Bench coding goes from 21.1% to 43.3%, ARC-AGI-3 from 1.5% to 30.2%, OSWorld computer use from 55.7% to 70.6%.
- The freshest cutoff in the lineup. Opus 5 ships with reliable knowledge through May 2026, versus January 2026 for both Fable 5 and Opus 4.8. More on why that matters below.
- Fast mode: roughly 2.5× output speed at 2× the base price, on the same model — a latency dial, not a different brain.
- Distribution: default model on Max, selectable on Pro,
claude-opus-5in the API from day one.
The price curve is the product
The interesting move here is not the benchmarks; it is the packaging. Anthropic did not launch a cheaper tier or a premium tier. It moved the capability curve under a fixed price and let “per task” economics do the marketing: if a model one notch below the flagship solves your task just as well, your effective frontier price was just cut in half without you touching anything.
This is what maturing markets look like. GPUs stopped competing on clock speed and started competing on performance-per-watt; models are now competing on capability-per-dollar at fixed price points. The strategic consequence for anyone building on these APIs: the model-selection decision has quietly become a routing decision that expires quarterly. The evals you ran in May that concluded “we need the flagship for tasks A and B” are stale twice over — the mid-tier moved, and the flagship’s price premium now buys a smaller delta.
The cutoff nobody prices in
The spec I would actually re-run evals for is the May 2026 knowledge cutoff. Teams spend enormous effort bolting retrieval onto models to compensate for stale world knowledge — library versions, API surfaces, framework idioms that changed after training. A four-month-fresher cutoff shrinks that compensation layer in ways that benchmark suites barely measure but production systems feel immediately: fewer hallucinated deprecated APIs, less RAG scaffolding for “what is the current way to do X.” When two models are close on capability, the fresher cutoff is often the better production model even if it loses a benchmark or two.
Fast mode and the shape of latency pricing
Fast mode — 2.5× speed for 2× price — is worth reading as a pricing experiment, not a feature. It unbundles a thing that used to be implicit: how much is a second worth in your product? For background batch work, nothing. For an interactive agent holding a user’s attention, potentially everything. Expect this dial to become standard across vendors, and expect your cost model to need a latency column it probably does not have today.
What I would do this week
- Re-run your routing evals. If you route premium traffic to Fable-class models, test how much of it Opus 5 now absorbs at half the per-task cost. On Anthropic’s own numbers, the answer for agentic and coding workloads is: a lot.
- Audit your RAG for staleness compensation. Anything you built specifically to patch a January cutoff may be dead weight against a May cutoff.
- Add latency to the cost model. Fast mode makes the speed–price trade explicit; decide deliberately which flows deserve it rather than toggling it globally.
- Assume this repeats. Between K3 on the open side and Opus 5 on the closed side, the half-life of a “which model” decision is now measured in weeks. Build the eval harness once; run it on every release.
The frontier used to be a place. Increasingly it is a curve — and the curve is bending toward whoever re-evaluates fastest.