Anthropic Ships Claude Opus 5 at Half of Fable 5’s Price

Anthropic released Claude Opus 5 this morning and priced it at $5 per million input tokens and $25 per million output, which is exactly what Opus 4.8 costs. The company is not claiming this is the smartest model it makes. Fable 5 still holds that spot. The claim is narrower and more interesting: Opus 5 comes "close to the frontier intelligence of Claude Fable 5 at half the price," in Anthropic's own words, and it does that without charging a cent more than the model it replaces at the top of the Opus line.

It went live the same day across claude.ai, Claude Code, Claude Cowork, the API, Pro and Max. The API id is claude-opus-5. A Fast mode runs it at roughly 2.5 times default speed for twice the base price, the same deal Opus 4.8 had.

RelatedClaude Sonnet 5: Near-Opus Coding at Half the Price

What did Anthropic actually claim?

The launch post leans almost entirely on partner and internal harnesses rather than the public benchmarks most readers know. Here is the full set of performance claims it makes:

• Frontier-Bench v0.1: Opus 5 surpasses every other model and more than doubles Opus 4.8's score.
• CursorBench 3.2: lands within 0.5% of Fable 5's peak at half the cost, per Cursor co-founder Sualeh Asif.
• ARC-AGI 3: three times the score of the next-best model.
• OSWorld 2.0: passes Fable 5 at just over a third of the cost.
• Zapier AutomationBench: Zapier CEO Wade Foster says previous models did not pass at all and Opus 5 hit 100%.
• Science tasks: 10.2 percentage points above Opus 4.8 on organic chemistry and 7.7 points above on protein work.

Notice what those have in common. Only the science figures are stated as a gap in points, and even those are relative to Anthropic's own previous model. Everything else is a comparison to a competitor or a multiple of somebody else's result. Cognition CEO Scott Wu adds that inside Devin, Opus 5 "shows particular strength on difficult debugging and root-cause analysis tasks," which reads as the most concrete practitioner claim in the whole post.

Why is there no SWE-bench Verified number?

Because Anthropic did not publish one. That is unusual for an Opus release, and it matters more than it sounds. SWE-bench Verified is the closest thing the industry has to a shared yardstick for agentic coding, and it is the metric our own AI Coding Leaderboard ranks on. Without it, Opus 5 cannot be placed against Opus 4.8 at 88.6%, Fable 5 at 95.0% or GPT-5.6 Sol at 96.2% on any common footing.

Frontier-Bench v0.1 is version 0.1 of something most developers have never run. CursorBench and AutomationBench are proprietary to the companies quoting them, and both of those companies are Anthropic customers with a commercial interest in the model looking good. None of that makes the numbers false. It does mean nobody outside those companies can check them today. We have added Opus 5 to the leaderboard as an unranked "verifying" row with no score attached, which is the honest state of the evidence until an independent harness gets to it. vals.ai runs every model itself on one neutral bash-only harness, and historically its numbers have landed 2.6 to 11.6 points below what vendors claim, because SWE-bench scores the scaffolding as much as the model.

Where does Opus 5 sit in the Claude lineup?

The awkward model here is Fable 5. If Opus 5 really does sit within half a percent of it on Cursor's harness while costing half as much, the reason to reach for Fable 5 shrinks to the specific long-horizon agent workloads it was built for. Anthropic has not deprecated Opus 4.8 and has not said anything about retiring it, so for now the Opus tier holds two models at identical prices, one of which is claimed to roughly double the other on Anthropic's own frontier benchmark. Most teams will not spend long deciding between those.

RelatedLaguna S 2.1: 8B Active Params, 70% Terminal-Bench

What did the safety testing find?

Anthropic published a system card alongside the launch and says the evaluations were run with private-sector and government partners. The headline finding is that Opus 5 "does not advance the frontier in risky, dual-use capabilities" and "remains behind Mythos 5 in both biology research and offensive cybersecurity." Mythos 5 is the invitation-only defensive-security model inside Project Glasswing, so the comparison is internal rather than reassuring in isolation. The launch post does not name an ASL tier, which leaves the system card as the only place to check what threshold the model was cleared under.

What it means for the market

Anthropic is private, so there is no ticker to watch directly. The read runs through the companies that resell and host it. Amazon and Alphabet both hold Anthropic stakes and both serve Claude through Bedrock and Vertex, and a flagship that halves the price of frontier-adjacent output pulls inference volume onto those platforms rather than pushing revenue per token up. For the coding-tool layer, Cursor, Cognition and Zapier all went on the record on launch day, which tells you where the distribution fight is. The signal for investors is on unit economics: when the top of a lab's range gets cheaper without the range getting shorter, margin pressure moves to whoever is selling per-seat subscriptions on top of it.

Our take

This is a real release with a real price advantage and a thin evidence package. Both things are true at once. The pricing move is aggressive and easy to verify, and if the CursorBench claim holds up in practice it changes the default model for a lot of agentic coding work overnight. The benchmark story is weaker than Anthropic usually ships. A launch built on version 0.1 of an internal benchmark plus three customer testimonials is a launch that expects to be taken on trust, and the absence of a SWE-bench Verified figure on a model sold primarily to developers is conspicuous. We are not calling it a red flag. We are saying the number should show up, and when it does we will rank it.

Original analysis by GenZTech. Primary source: Anthropic.