Claude Opus 5 holds Opus pricing flat while chasing Fable 5's coding lead

Anthropic released Claude Opus 5 on July 24 as its new flagship, and priced it exactly where Opus 4.8 sat: $5 per million input tokens and $25 per million output. The pitch is less a new capability ceiling than a claim about value, that a smaller, cheaper model can trade blows with the current frontier on the work developers actually point agents at.

What shipped

Opus 5 is available today across Anthropic’s platforms and through the API as claude-opus-5, with reasoning on by default at configurable effort levels. The pricing holds flat against Opus 4.8, and a fast mode runs at roughly 2.5 times the speed for double the base rate. Two new beta features target agent reliability rather than raw output: mid-conversation tool changes, which let a running session swap its tool set without restarting, and automatic fallbacks that route a safety-flagged request to a less powerful model instead of returning an error.

The benchmark claims are Anthropic’s own, and they cluster around coding and agentic work. On the company’s Frontier-Bench v0.1, Anthropic reports Opus 5 more than doubles Opus 4.8’s score at a lower cost per task. On CursorBench 3.2 at maximum effort, it says the model lands within half a percent of Fable 5, the current frontier, at half the cost. It also reports gains on computer-use and automation suites, and single-digit-percentage-point improvements over Opus 4.8 on chemistry and protein tasks. Vendor benchmarks deserve the usual caution, but the shape of the claim is consistent: a mid-weight model priced below the frontier that Anthropic positions as close enough on the tasks that pay for themselves.

For a working developer, the practical read is where that lands day to day. Anthropic and its launch partners describe Opus 5 as stronger on long-horizon agentic coding, multi-file features, larger refactors, and coordinating teams of subagents, the exact workloads that make a Claude Code session or a CI agent expensive to leave running. If the CursorBench parity holds outside the benchmark, the calculation for a multi-file refactor or an overnight agent loop shifts from “reach for the frontier model and pay for it” to “run Opus and keep the budget.”

Where this lands in the market

The more interesting positioning is what Opus 5 gives up and what it drops. TechCrunch, reporting independently on the launch, notes that Opus 5 is smaller than Fable 5 yet still outperforms it on several benchmarks, and frames the model as cheaper and less restrictive, “likely making it preferable in most use cases.” Part of that looseness is a data point: Opus 5 is not bound by the 30-day retention policy that applies to Anthropic’s Fable and Mythos tiers, which matters to teams that cannot send code and prompts into a longer retention window.

That reframes the competition the same way Sonnet 5 did a month earlier, around cost and usability rather than a top-line score. Anthropic appears to be betting that, for many developer workflows, the next gains come less from another capability jump than from making a capable-enough model cheap, fast, and unencumbered enough to run at scale. Holding Opus pricing flat while claiming frontier-adjacent coding, and stripping the retention strings, is a coherent version of that bet aimed squarely at teams deciding which model their agents default to.

What’s worth watching

1. Whether the CursorBench parity survives real repositories. A within-half-a-percent claim against Fable 5 at half the price is the load-bearing number here. The test is whether teams running Opus 5 in agent loops see completion rates that justify not escalating to the frontier tier.
2. How the automatic fallbacks behave in production. Silently routing a flagged request to a weaker model avoids an error, but it also means an agent can quietly change quality mid-task. Developers building on Opus 5 should confirm they can see when a fallback fired rather than discover it in the output.
3. Whether the retention carve-out becomes a durable product line. Opus 5 sitting outside the Fable and Mythos retention policy is a design signal as much as a privacy one. Watch whether Anthropic keeps its highest-capability tiers on stricter terms while positioning Opus as the model teams are meant to actually deploy.

The plain read is that the flagship label is doing less work than the price tag. The number that decides adoption is no longer the top benchmark but the cost at which an agent reliably finishes, and Opus 5 is the clearest sign yet that Anthropic is designing to that number. Stackmaven’s follow-up coverage will revisit how Opus 5 holds up in production agent workloads on or around October 23.

Sources cited

1. Anthropic: Introducing Claude Opus 5 www.anthropic.com
2. TechCrunch: Anthropic launches Opus 5 techcrunch.com
3. Vercel Changelog: Claude Opus 5 now available on AI Gateway vercel.com