Anthropic released Claude Opus 5 on July 24, 2026, positioning it as the model most paid Claude users and enterprise teams should reach for before paying for Fable 5.
The pitch is simple: Opus 5 comes close to Claude Fable 5's frontier capability at half the token price, while keeping the same base rate as Opus 4.8. Anthropic says it is available now on Claude.ai, Claude Code, Claude Cowork, the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. It is also the new default model for Claude Max subscribers and the strongest model on Claude Pro.
Watch the demo
Anthropic's launch page includes a short demo of Opus 5 producing an interactive 3D animal-cell artifact. The point is not that Opus 5 is a video model; it is that Anthropic is pushing the model's ability to reason through visual, scientific, and code-heavy artifact generation.
What changed
Opus 5 is not being framed as Anthropic's absolute top model. Fable 5 remains the choice for the hardest long-running autonomous work, and Mythos 5 still leads in restricted cybersecurity and biology settings. Opus 5 is the more practical release: broadly available, cheaper than Fable, less burdened by Fable's heavy safety fallbacks, and strong enough to win several work-oriented benchmarks outright.
Anthropic says Opus 5 is especially better at verifying its work, iterating when it fails, and building tools it needs along the way. One launch example had the model receive a machine-part drawing without direct image-viewing access, then write its own computer-vision pipeline to extract geometry and rebuild the part as a 3D FreeCAD model. Another example involved finding the root cause of a real package-manager bug and fixing an edge case missed by the community patch.
That is the central product story: less "answer this prompt" and more "keep working until the job is checked."
The benchmark numbers
Anthropic's official page emphasizes cost-performance charts more than raw tables, but the public figures tell a clear story. Opus 5 leads the reported field on agentic coding, novel problem-solving, and knowledge-work Elo, while trailing Fable/Mythos slightly on some specialist coding and safety-sensitive domains.
| Benchmark | Claude Opus 5 | Claude Fable 5 | GPT-5.6 Sol | Why it matters |
|---|---|---|---|---|
| Frontier-Bench v0.1 | 43.3% | 33.7% | 34.4% | Agentic terminal coding tasks |
| ARC-AGI-3 | 30.2% | - | 7.8% | Novel problem-solving |
| GDPval-AA v2 | 1,861 Elo | 1,747 Elo | 1,736 Elo | Knowledge-work task quality |
| SWE-bench Pro | 79.2% | 80.0% | - | Software engineering issue resolution |
| DeepSWE v1.1 | 68.8% | 69.7% | 72.7% | Agentic coding on harder software tasks |
The standout is Frontier-Bench v0.1, where Opus 5 beats Fable 5 and GPT-5.6 Sol while more than doubling Opus 4.8's reported result. On ARC-AGI-3, Anthropic says Opus 5 scores about three times higher than the next-best model in its comparison set. On GDPval-AA v2, a knowledge-work evaluation from Artificial Analysis, Opus 5 also leads Fable 5 and GPT-5.6 Sol.
The caveat: these are launch-day benchmark claims and should not replace private evals. Teams should test Opus 5 against their own codebases, documents, workflows, and refusal-sensitive tasks before switching production traffic.
Pricing and model details
Opus 5 keeps Opus 4.8's base pricing:
| Model or mode | Input | Output |
|---|---|---|
| Claude Opus 5 standard | $5 / million tokens | $25 / million tokens |
| Claude Opus 5 Fast mode | $10 / million tokens | $50 / million tokens |
| Claude Fable 5 | $10 / million tokens | $50 / million tokens |
The API model ID is claude-opus-5. Anthropic's docs list a 1 million-token context window, 128K max output in synchronous Messages API calls, and up to 300K output tokens through the Message Batches API beta header.
The model also supports Anthropic's effort setting. On Opus 5, effort defaults to high on the Claude API and Claude Code, but developers can set it lower when they want faster and cheaper results. That matters because reasoning tokens are part of the bill: a routing policy that sends routine tasks to lower effort and escalates only hard work can change real operating cost more than the headline per-token price suggests.
Safeguards
Opus 5 uses a lighter-touch safety profile than Fable 5, but Anthropic still adds stronger guardrails around a narrow range of cybersecurity tasks. The company says Opus 5 can help find vulnerabilities in source code, but blocks binary-based vulnerability scanning, penetration testing, and exploit generation.
Anthropic says its cyber classifiers should intervene about 85% less often than they do on Fable 5. When Opus 5 is flagged in Claude.ai, Claude Code, or Claude Cowork, requests fall back to Opus 4.8 by default. API users can also enable Automatic Fallbacks so a flagged request routes to another model instead of failing outright.
For privacy-sensitive teams, there is another practical difference: Anthropic says Opus 5, like Opus 4.8, does not carry the general-access data retention requirements that made Fable 5 more awkward for some users.
Why this release matters
Opus 5 looks like Anthropic's answer to a new market reality: frontier models are no longer judged only by peak intelligence. Teams now care about cost per task, latency, refusal rate, long-context reliability, and whether a model can keep checking its own work without constant human steering.
For engineers, the strongest signal is agentic coding. For researchers and analysts, the GDPval-AA and life-sciences gains make Opus 5 a stronger everyday reasoning model. For lawyers, finance teams, and other professional users, the more interesting detail is consistency: early customers repeatedly describe better self-checking, cleaner handoffs, and fewer missed edge cases.
The short version: Fable 5 remains Anthropic's highest-end public model for the hardest jobs, but Opus 5 is now the more obvious default for most professional Claude work.