Claude Opus 5.2: availability and setup for Hermes Agent & OpenClaw

What is Claude Opus 5.2?

Opus 5.2 is the name circulating in developer reports about a possible Claude Code rollout. As of this check, we could not verify a public Anthropic release announcement or API model ID for that name. Anthropic's official model catalog lists Opus 5 with the API ID claude-opus-5. Its newsroom did not establish an Opus 5.2 launch in the sources we checked.

What's circulating: developer reports, not a release note

The news source is 新智元's WeChat article 突发！Opus 5.2深夜上线，RSI真来了？. These are unverified reports, not Anthropic documentation:

• Some developers reportedly saw routing or slug clues pointing to 5.2 while the interface still said Opus 5, prompting speculation that 5.1 was skipped.
• They described faster responses, cleaner output, less stopping on long tasks, and a “gauntlet loop” of autonomous testing and revision.
• A reported probe asks who “重置哥 Tibo” is with search off: the alleged 5.2 route answers while web Opus 5 does not. That cannot reliably identify a backend model.

The supplied references include @notjazii on X and @NFT_Chen on X. We have not independently verified their routing evidence. Fable 5.2, an internal “Model 2,” and recursive self-improvement (RSI) belong to this rumor context; they are not confirmed OpenClaw Launch product capabilities.

Why the reports matter for agents

For a Hermes or OpenClaw agent, the useful improvement would be finishing a task with fewer interventions: reading the relevant files, making a change, checking it, and correcting a failed attempt. Faster individual replies help only if the complete task also finishes sooner. More autonomous retries can increase token use.

Treat these as things to measure when access is confirmed. A loop that edits and tests a project can improve that project without changing the model itself; it is not, by itself, evidence of RSI. Keep your current working model until a comparison on your own tasks supports switching.

Prepare the hosted setup: identical steps for both frameworks

Hermes Agent and OpenClaw instance cards share the same model selector on OpenClaw Launch. The Anthropic BYOK path below applies to both. BYOK means using your own provider API key, with model usage billed to that provider account.

1. Confirm access first. Check the Anthropic model catalog and your account's available models. Copy the exact ID. For Opus 5.2, proceed only when the provider actually exposes it to you.
2. Connect your key. On an eligible plan, open API Keys, add your Anthropic API key, and save it. Check your provider balance and usage limits.
3. Select the model on your running instance. Choose its Anthropic entry, or use the custom fields: Platform anthropic, Model ID set to the exact bare API ID. For example, claude-opus-5 is the documented Opus 5 ID, not an Opus 5.2 switch. The selector adds the provider prefix.
4. Verify the change. Wait for the save to finish and confirm the displayed provider and model. Try a small task in your own test conversation before moving scheduled work. If the provider rejects the ID or your account lacks access, return to a known working model and check the provider listing.

If you use OpenRouter

Check the actual OpenRouter catalog for availability, tool support, and pricing. Select a listed model through the OpenRouter route. With custom fields, choose Platform openrouter and copy the provider's complete model slug into Model ID. Do not translate an Anthropic ID into a guessed OpenRouter slug or assume a Claude Code rollout is available through platform credits.

Claude subscriptions and Claude Code routing

If you subscribe to Claude, check the models available to your account inside Claude or the official Claude Code client. In Claude Code, /model opens model selection and /status helps inspect your current setup. Anthropic's model configuration docs currently map the Anthropic opus alias to Opus 5 and explain that alias resolution depends on the provider and can change over time. They do not document a guaranteed Opus 5.2 route.

A subscription rollout in an Anthropic app does not grant the same access to your hosted agent. This guide's hosted route uses an Anthropic API key; it does not promise a generally available Claude subscription connection on OpenClaw Launch. Account eligibility and billing must be checked for each route.

Self-hosted Anthropic configuration

These examples configure the documented Opus 5 baseline via Anthropic, not an unverified Opus 5.2 model. If a future model becomes available, substitute its exact API ID and confirm that your installed framework supports it. Hosted customers should use the dashboard steps above.

Hermes Agent

In ~/.hermes/config.yaml, Hermes uses separate provider and model fields, as shown in its upstream example configuration and native Anthropic provider guide:

model:
  provider: anthropic
  default: claude-opus-5

Set your own key in ~/.hermes/.env:

ANTHROPIC_API_KEY=YOUR_ANTHROPIC_API_KEY

You can also use hermes model for interactive configuration. The provider is not prefixed onto the native Anthropic model.default. Model changes apply to new sessions; use /new for a fresh conversation. If a gateway restart is needed after changing credentials, use hermes gateway restart. A restart preserves conversation history, as the Hermes session docs explain.

OpenClaw

In ~/.openclaw/openclaw.json, merge this model setting into your existing configuration. Make ANTHROPIC_API_KEY available to the gateway process through its environment or supported secret configuration, following the OpenClaw Anthropic provider guide:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-opus-5"
      }
    }
  }
}

OpenClaw needs the anthropic/ prefix. If your configuration restricts selectable models through agents.defaults.models, include the selected model there too. Existing conversations can have model pins: changing the default or restarting does not necessarily clear them. Use /model default to clear a session override and inspect the active model, following upstream model-selection documentation.

Evaluate an available model before switching your bot

• Use repeatable tasks: one bug fix with a failing test, one multi-step tool task, and one structured report checked against its source.
• Keep conditions comparable: same prompt, workspace, tools, search setting, and effort setting where supported; start fresh conversations.
• Record completed outcomes: pass rate, manual interventions, elapsed time, retries, and total provider cost. Include failed attempts.
• Set a stopping point: cap task time or budget and define the acceptance check. More self-iteration is useful only when it improves the result.

There is no verified Opus 5.2 price in this guide. Before a paid trial, use the provider's current pricing. Hosting charges and model usage are separate; do not carry Opus 5 rates forward as a quote for a rumored model.

Frequently asked questions

Related setup guides

• Hermes Agent with Anthropic · Hermes hosting
• OpenClaw with Anthropic · Opus 5 guide
• Hermes and Claude Code · OpenClaw and Claude Code
• Connect an API key · Read this guide in Chinese