Claude Opus 5 and the Shift to Long-Running Agents

Claude Opus 5 is Anthropic’s attempt to bring near top-tier capability to everyday agentic work at a lower price. Anthropic says it comes close to the frontier intelligence of Claude Fable 5 at half the price, and the company frames it around longer-horizon, multi-step tasks in coding and professional work. That makes it a natural candidate for agents that run for many minutes or hours. It also makes the controls around those agents more important, not less.

What changed

Anthropic announced Claude Opus 5 on July 24, 2026, and made it available the same day. According to the announcement:

• Pricing is $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8.
• The API model ID is claude-opus-5, and the model is available on all platforms.
• It is the new default model on Claude Max and the strongest model on Claude Pro.
• A fast mode runs about 2.5x the default speed at twice the base price on the Claude Platform, and through usage credits in Claude Code.
• Beta features include changing tools mid-conversation without invalidating the prompt cache, and automatic API fallbacks to another model when requests are flagged.

The announcement does not state a context window, so check the current model documentation before planning around one.

What Anthropic claims on performance

Anthropic says Opus 5 is the new state of the art on Frontier-Bench and GDPval-AA, while remaining behind Mythos 5 on cybersecurity tasks. Selected vendor-reported results:

• Frontier-Bench v0.1: more than double Opus 4.8’s performance at a lower cost per task.
• CursorBench 3.2 at max effort: within 0.5% of Fable 5’s peak score at half the cost per task.
• OSWorld 2.0: surpasses Fable 5’s best result at just over a third of the cost.
• Zapier AutomationBench: a pass rate about 1.5x the next-best model at the same cost per task.
• Life sciences: better than Opus 4.8 on every evaluation listed, including 10.2 points higher on organic chemistry.

Early-access customers also reported gains. Lovable cited a 22% improvement over Opus 4.7 on its hardest agentic coding tasks, and Box reported an 8% improvement over Opus 4.8. These are vendor and customer figures. We have not reproduced them.

The pattern in these claims is cost per task, not only peak score. Anthropic is pitching Opus 5 as the efficient choice for agentic work, and the announcement says the largest gains appear on longer-horizon and vaguer tasks. It also describes effort settings as a dial between intelligence and token spend.

Safety and safeguards

Anthropic says its automated behavioral audit rates Opus 5 as its most aligned model to date, with an overall misaligned-behavior score of 2.3, the lowest of recent models. It also describes Opus 5 as its safest model yet at avoiding reckless actions with hard-to-reverse side effects. For long-running agents, that last claim is the one to test hardest.

On cybersecurity, Opus 5’s classifiers are less restrictive than Fable 5’s, and Anthropic expects them to intervene about 85% less often. They allow vulnerability finding in source code but block binary vulnerability scanning, penetration testing, and exploit generation. Flagged requests fall back to Opus 4.8 by default in Claude.ai, Claude Code, and Claude Cowork, and fallbacks can be enabled on the API.

That fallback has a practical consequence. If a long agent run triggers a classifier partway through, part of the work may be done by a different model. Teams should log which model handled each step.

Why it matters

A ten-minute agent run creates different risks from a ten-second response. It can accumulate wrong assumptions, modify more files, consume more budget, and encounter untrusted instructions in documents or websites.

The right safety question is not whether the model is generally capable. It is whether the specific workflow has checkpoints that limit the cost of being wrong. A cheaper, faster model makes it tempting to run more agents with less review. That is exactly when boundaries need to be designed in advance.

Who should use it

Opus 5 looks most relevant for teams that already run agentic coding or document workflows and want lower cost per task. Claude Max users get it by default. API users weighing it against Fable 5 should run their own comparison, since Anthropic’s positioning is “close to” Fable 5, not equal across every task. For a look at the next Fable generation, see our guide to Claude Fable 5.1 and Mythos 5.1.

Security teams should note the cyber limits and the fallback behavior before building workflows that depend on vulnerability work.

Design the boundary first

Classify tools into read, prepare, and commit groups. Reading a repository is different from pushing a release. Drafting an email is different from sending it. Looking up an invoice is different from paying it.

Start Opus 5 with broad read access, narrow prepare access, and no irreversible commit access. Add permissions only after logs show reliable behavior on repeated tasks.

What to test

Use a task that takes a skilled person at least thirty minutes and has an objective finish line. For code, require tests and a diff review. For research, require primary sources and a claim table. For a document, require a rendered output and a checklist against the brief.

Record pauses, recoveries, unnecessary tool calls, and changes made outside scope. The model should not receive credit merely for producing a plausible final message.

Before adopting, verify:

• total cost per completed task at each effort level you plan to use, not just token price;
• how often fallbacks to Opus 4.8 occur in your workload;
• whether fast mode’s speed gain justifies its doubled price for your use case;
• how the model behaves when a tool returns an error or a document contains injected instructions;
• whether its self-verification, such as writing tests, actually catches its own mistakes.

The conclusion

Claude Opus 5 is designed for work where persistence matters, and Anthropic’s pricing makes long runs cheaper to attempt. That makes it interesting for complex production tasks and a poor excuse for removing oversight. Longer runs should produce stronger evidence, not longer chains of invisible decisions.