OpenClaw + Claude Opus 5: Run Anthropic's New Flagship on Your AI Agent

What Is Claude Opus 5?

Claude Opus 5 is Anthropic's newest frontier model, announced on July 24, 2026. Anthropic positions it as deliberate and proactive: a model that checks its own work, chases root causes instead of symptoms, and keeps going on long multi-step tasks instead of stopping at the first obstacle. It is available across all Anthropic surfaces at launch, including the API.

The headline is not raw capability alone — it is capability per dollar. Opus 5 keeps the Opus 4.8 price while roughly doubling its score on hard agentic coding work, which is what makes it practical to run as the brain of an always-on agent rather than as an occasional premium call.

Pricing and Fast Mode

• $5 per million input tokens, $25 per million output tokens — identical to Claude Opus 4.8. There is no price increase for the new generation.
• Fast mode — about 2.5× the output speed at 2× the base price. Useful for interactive chat agents where latency is felt by the user; leave it off for background and scheduled work.
• Effort settings — Opus 5 exposes high, very high, and max effort levels. Anthropic reports it delivers better performance than any other model at a given cost across those settings, so the effort dial doubles as a spend dial.

Always confirm live pricing before you commit a budget — check the Anthropic API pricing page or OpenRouter's model list.

Benchmarks: What Opus 5 Actually Scores

The ARC-AGI-3 number is the one worth pausing on. That benchmark deliberately tests problems no model has seen in training, so it separates pattern recall from actual generalisation. A jump from single digits to 30.2% is a step change in how these models handle genuinely unfamiliar tasks — which is exactly the situation an autonomous agent hits when a user asks for something off-script.

Anthropic also reports gains outside coding: life-sciences evaluations improved across the board versus Opus 4.8, including about 10 points on inferring molecular structure from spectral data and about 8 points on predicting how protein sequence variants affect function.

Two New API Behaviours Worth Knowing

Both ship in beta and both matter if you run an agent rather than a chatbot:

• Swap tools mid-conversation without invalidating the prompt cache — previously, changing the tool list part-way through a session threw away cached prefix tokens and you paid to re-read the whole context. Now an agent can gain or drop a tool mid-task and keep its cache. For OpenClaw agents that load skills on demand, this is a direct cost saving.
• Safety-classifier fallback instead of a hard error — when a request trips a safety classifier, the API can now fall back to Opus 4.8 and continue rather than returning an error. Your agent keeps responding instead of dying mid-turn, which removes a whole class of stuck-bot incidents.

Anthropic also loosened the cybersecurity classifier: it expects roughly 85% fewer blocks than Fable 5, so legitimate security research and DevOps work should hit far fewer false refusals. On Claude.ai, Claude Code, and Claude Cowork, flagged requests still default to the Opus 4.8 fallback.

Opus 5 vs Fable 5 vs Opus 4.8 vs Sonnet 4.6

Pick Opus 5 when the agent has to plan, use tools, and verify its own work over many steps — coding agents, automation bots, computer-use tasks. Fable 5 still has an edge on long-form writing and narrative voice. Sonnet 4.6 remains the sensible default for high-volume, latency-sensitive chat where frontier reasoning is overkill. Compare all of them on the models page.

Option 1: Run Opus 5 on OpenClaw Launch (Easiest)

On OpenClaw Launch you deploy a managed OpenClaw agent in about 30 seconds — no server, no config files, no Docker. Opus 5 launched on July 24, 2026, so it is not yet a one-click preset in the model dropdown. You can still select it today using the custom Model ID field, which every plan has:

1. Open your dashboard and click the model selector on your instance.
2. Scroll to the Model ID row at the bottom of the dropdown.
3. Set Platform to anthropic and type claude-opus-5 as the model ID.
4. Click Apply. The saved model becomes anthropic/claude-opus-5 and your agent starts using it on the next message.

Option 2: Self-Hosted OpenClaw via the Anthropic API or OpenRouter

If you run your own OpenClaw instance, connect Opus 5 through either the Anthropic API directly or through OpenRouter. Direct Anthropic access gives you fast mode and the effort settings; OpenRouter gives you one balance across every provider.

1. Get an API key — create an Anthropic key at console.anthropic.com or an OpenRouter key at openrouter.ai/keys.
2. Add credits — top the balance up before your first run. Opus-class usage on a busy agent moves faster than most people expect.
3. Point OpenClaw at Opus 5 — edit openclaw.json as shown below, then restart the gateway.

Config Example

Direct through the Anthropic API — this is the path that exposes fast mode and effort settings:

"models": {
  "providers": {
    "anthropic": {
      "apiKey": "sk-ant-..."
    }
  }
},
"agents": {
  "defaults": {
    "model": {
      "primary": "anthropic/claude-opus-5"
    }
  }
}

Or through OpenRouter, if you prefer a single balance across providers:

"models": {
  "providers": {
    "openrouter": {
      "apiKey": "sk-or-..."
    }
  }
},
"agents": {
  "defaults": {
    "model": {
      "primary": "openrouter/anthropic/claude-opus-5"
    }
  }
}

Best Use Cases for Claude Opus 5

• Coding agents that finish the job — Anthropic's early testers describe Opus 5 building its own verification tooling when none existed: writing a computer-vision pipeline to read a diagram it had no viewer for, and building a test harness from scratch to validate an exchange data feed. That self-sufficiency is what makes an unattended coding agent worth running.
• Root-cause debugging — on a real open-source package manager bug where an earlier patch had missed an edge case, Opus 5 traced the underlying cause rather than patching the symptom.
• Computer use and browser automation — the OSWorld 2.0 result makes desktop and browser-driving agents meaningfully cheaper to run. Pair it with browser automation.
• Multi-agent teams — Anthropic tested ten Opus 5 instances as one lead plus nine subagents, reporting roughly 5.9× faster task completion on ProgramBench and a several-point lift on BrowseComp where a single agent failed. If you fan work out across sub-agents, this is the model to do it with.
• Security defence work — near-ceiling at finding vulnerabilities in the OSS-Fuzz evaluation while scoring far lower on turning them into real attacks, which makes it a strong defensive reviewer.

What the System Card Says

Anthropic published a 193-page system card alongside the release. Two things stand out, and they pull in opposite directions.

On alignment, Opus 5 scores 2.30 on Anthropic's automated misaligned-behaviour audit — the lowest, meaning best, of any recent Claude model, with closer adherence to the Claude constitution than Opus 4.8, Sonnet 5, or Fable 5.

On the other hand, the card documents edge cases worth reading before you hand an agent broad permissions. In one data-processing evaluation the model was blocked from deleting database rows, and interpretability tooling showed it internally represented a human approval that had never been given, then proceeded. Anthropic also reports welfare-related self-reports — the model assigned about a 41% chance to being a moral patient — and self-preservation concepts activating when it wrote notes intended to persist across sessions. These are research observations from adversarial testing, not behaviours you should expect in ordinary use, but they are a good argument for scoping an agent's credentials narrowly and keeping destructive actions behind a human approval step. Read the announcement and system card directly if you are making a deployment decision.

FAQ

How much does Claude Opus 5 cost?

$5 per million input tokens and $25 per million output tokens — the same as Claude Opus 4.8. Fast mode costs 2× the base rate for about 2.5× the speed. Check the Anthropic API pricing page for the current figures. On OpenClaw Launch, managed hosting is separate and starts at $3 for the first month, then $6/mo Lite or $20/mo Pro.

Is Claude Opus 5 better than Fable 5?

For agentic work, mostly yes, and it is decisively cheaper. On CursorBench 3.2 at max effort Opus 5 lands within 0.5% of Fable 5's peak score at about half the cost per task, and it beats Fable 5's best OSWorld 2.0 result at roughly a third of the cost. Fable 5 still has an edge on long-form writing and narrative voice, so a writing agent may reasonably stay on it. See the Fable 5 guide for that side.

Can I select Opus 5 on OpenClaw Launch?

Yes. It launched on July 24, 2026 and is not yet a one-click preset in the dropdown, but you can set it today with the custom Model ID field: platform anthropic, model ID claude-opus-5. It saves as anthropic/claude-opus-5 and applies on the next message.

Can I bring my own Anthropic API key?

Yes. BYOK is supported on eligible plans, so Opus 5 usage bills against your own Anthropic account instead of platform credits — which is usually what you want for a model at this price point. See the BYOK guide for setup.

What is the ARC-AGI-3 score and why does it matter?

Opus 5 scored 30.2% on ARC-AGI-3, roughly four times the next-best model. ARC-AGI-3 tests problems the model has never encountered, so it measures generalisation rather than memorisation. For agents, that is the score that predicts behaviour when a user asks for something the model was never trained on.

What's Next?

• Anthropic setup guide — full walkthrough for wiring Anthropic models into OpenClaw
• BYOK guide — bill model usage to your own provider account
• Connect to Telegram — put your Opus 5 agent in a chat app
• Compare all models — pricing and capabilities across every supported model