Models Claude Opus 5 Costs Half of Fable 5, With One Caveat Published 25 July 2026 | Last updated 26 July 2026 Verified by Grok 26 July 2026 • AI Race Facts verification protocol Anthropic released Claude Opus 5 on 24 July 2026 at $5 per million input tokens and $25 per million output tokens, unchanged from its predecessor and exactly half the rate of Claude Fable 5, the company's most capable widely released model. Anthropic says the new model comes close to Fable 5 on intelligence while costing half as much per token. The headline coding result that supports that claim carries a footnote. Anthropic discloses that a different model, Claude Opus 4.8, stood in whenever a safety classifier blocked a request during the benchmark run, for both Opus 5 and Fable 5. TL;DR Claude Opus 5 shipped 24 July 2026 at $5 and $25 per million input and output tokens , the same rate as Claude Opus 4.8. Claude Fable 5 is documented at $10 and $50 , so Opus 5 is exactly half on both figures. Opus 5 in Fast mode is priced at $10 and $50 , the same as Fable 5's standard rate. Anthropic expects Opus 5's cybersecurity classifiers to trigger around 85 percent less often than Fable 5's. Blocked requests no longer simply fail. They fall back to another model, including inside the benchmark run Anthropic published. Disclosure. Anthropic makes the AI model used to draft copy for this publication, and the model in question is Claude Opus 5, the subject of this article. Every performance figure below is Anthropic's own and is labeled accordingly. None of it has been independently benchmarked here. This story was verified by Grok on 26 July 2026 against primary Anthropic sources using the AI Race Facts verification protocol (duplication check + six-field verification log + pre-publication gate). All key claims trace to Tier 1 primary documents. What Anthropic shipped on 24 July Claude Opus 5 is available on the Claude API as claude-opus-5 , and through Amazon Bedrock, Google Cloud, and Microsoft Foundry, according to Anthropic's platform documentation . It carries a one million token context window as both the default and the maximum, with no smaller variant, and up to 128,000 output tokens per request. Anthropic states in its launch post that Opus 5 is the new default model on Claude Max and the strongest model available on Claude Pro. It also states that Opus 5 remains behind Claude Mythos 5, the restricted model offered through Project Glasswing, on both biology research and offensive cybersecurity. Does this apply to you? The pricing below is API pricing, billed per token to developers calling the model directly. Most people reading this will encounter Opus 5 through a Claude subscription instead, where it is included in a plan allowance rather than metered. The rate card matters if you are building on the API, budgeting an agent workload, or comparing vendors. It does not describe what a Pro or Max subscriber pays. How the price compares Model Input per 1M Output per 1M Access As of Claude Opus 5 $5 $25 General 2026-07-25 Claude Opus 5, Fast mode $10 $50 Claude API only, research preview 2026-07-25 Claude Opus 4.8 $5 $25 General, still available 2026-07-25 Claude Fable 5 $10 $50 General 2026-07-25 Claude Mythos 5 $10 $50 Limited, Project Glasswing 2026-07-25 DERIVED Opus 5 is exactly half of Fable 5 on both figures: $5 times two is $10, and $25 times two is $50. Opus 5 in Fast mode, at $10 and $50, matches Fable 5's standard rate exactly. Fast mode runs at roughly 2.5 times default speed and is available on the Claude API only, not through Amazon Bedrock, Google Cloud, or Microsoft Foundry. One documented difference does not show up in price. Fable 5 and Mythos 5 carry 30-day data retention and are not available under zero data retention agreements. Anthropic states Opus 5 has no data retention requirement for general access, which matters for compliance-sensitive workloads. The benchmark number comes with a footnote CLAIM Anthropic reports that on Frontier-Bench v0.1, a software engineering evaluation, Opus 5 surpasses all other models and more than doubles Opus 4.8's performance at a lower cost per task. It reports similar results on CursorBench, ARC-AGI 3, OSWorld 2.0, and Zapier's AutomationBench. A footnote at the base of the launch post qualifies the Frontier-Bench figure. Anthropic states the results come from an internal run on the mini-SWE-agent harness with a Google Kubernetes Engine backend, averaged over five attempts per task, and that "Opus 4.8 served as fallback on safety-classifier refusals for Opus 5 and Fable 5" . That qualifier is doing real work. When a safety classifier blocked a request mid-benchmark, an older model answered instead, and the result counted. The reported figure therefore describes a system consisting of a model plus a fallback, not a model on its own. It also means Opus 4.8 sits inside the measurement that Opus 5 is being compared against. Anthropic disclosed this in its own copy. Nobody has to infer it. Anthropic has not published how often the fallback fired during the run. Without that figure, the size of the effect on the reported scores cannot be calculated from the material released. What happens now when Claude refuses The footnote points at a broader shift. Refusal is becoming a routing decision rather than a dead end, and the routing now runs in several directions at once. On Fable 5, a declined request returns a successful response carrying a refusal status and the name of the classifier that declined it, rather than an error. Developers can retry on another model server-side, client-side, or manually, and Anthropic refunds the prompt-cache cost of the switch. Opus 5 adds a mode in which Anthropic picks the fallback model itself, by refusal category, instead of the developer maintaining a list. Opus 5's cyber classifiers are described as less restrictive than Fable 5's, allowing vulnerability finding in source code while blocking binary-based scanning, penetration testing, and exploit generation. PROJECTION Anthropic expects them to intervene around 85 percent less often than Fable 5's, based on its own testing. Flagged requests in Claude.ai, Claude Code, and Claude Cowork fall back to Opus 4.8 by default. In the other direction, biology requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8. This is the commercial face of the same control problem that produced federal sign-off on frontier model releases in June, the export curbs on Fable 5 , and the restricted clearance of Mythos 5 for critical-infrastructure defenders . Capability is being partitioned by who is asking and what they are asking for, and the partition now has an API parameter. One breaking change developers have to handle The launch post does not mention it, but the documentation does. On Opus 5, thinking is on by default, and disabling it is accepted only at effort level high or below. A request that disables thinking at effort xhigh or max returns a 400 error. Anthropic labels this a breaking change from Opus 4.8, where the two settings were independent. Two smaller changes also land. The minimum cacheable prompt drops to 512 tokens from 1,024, and tools can be added or removed between turns without invalidating the prompt cache, in beta. Anthropic also advises removing verification instructions carried over from earlier models, since Opus 5 verifies its own work and the old instructions cause it to over-verify. What cannot be checked yet Every capability and safety figure in this piece originates with Anthropic. The company publishes a system card for Opus 5 covering the biology and cybersecurity evaluations it references, and that document was not retrieved for this article. No independent benchmark of Opus 5 has been published at the time of writing, and the 24 customer testimonials in the launch post come from early-access partners, each an interested party. The alignment figure deserves particular care. Anthropic reports that Opus 5 scores 2.3 on overall misaligned behavior in an automated behavioral audit, which it describes as the lowest of its recent models. That audit is Anthropic's own instrument, run on Anthropic's own model, and the scale has no external referent published alongside it. It is a vendor measurement, not a third-party result. Model pricing, availability, and plan terms change frequently. Verify with the vendor before relying on any figure here. Figures labeled CLAIM are operator-stated and not independently audited. Figures labeled DERIVED are calculated here. Sources Anthropic, "Introducing Claude Opus 5" , 24 July 2026. Anthropic, "What's new in Claude Opus 5" , Claude Platform Docs, retrieved 25 July 2026. Anthropic, "Introducing Claude Fable 5 and Claude Mythos 5" , Claude Platform Docs, retrieved 25 July 2026. Anthropic, "Refusals and fallback" , Claude Platform Docs, retrieved 25 July 2026. Anthropic, "Redeploying Fable 5" , 30 June 2026. Published 25 July 2026 | Last updated 26 July 2026 • Verified by Grok