Claude Fable 5.1 on OpenClaw and Hermes: Setup, Pricing, and the 75% Cache Discount

What Is Claude Fable 5.1?

Claude Fable 5.1 is Anthropic's newest frontier model, announced in September 2026 alongside Claude Mythos 5.1. The two are the same underlying model with different safeguards: Fable 5.1 is generally available on every Anthropic surface and on AWS, Google Cloud, and Azure, while Mythos 5.1 ships only through vetted access programs aimed at cybersecurity and life-sciences work.

The headline capability claim is long-horizon reliability. Anthropic reports that Fable 5.1 avoids the shortcuts that leave work looking finished but subtly broken, and that it chases root causes instead of symptoms. The concrete example: at the investment firm Millennium it identified the cause of a rare crash that the firm's own engineers — and every other model they tried — had failed to explain across several years.

For anyone running an agent rather than a chatbot, that is the property that matters. An agent that reads a repo, plans, edits, runs tests, reads the failure, and loops is only as good as its worst step forty turns in.

The Price Change Is the Real News

Fable 5.1 costs exactly what Fable 5 cost per input and output token. What changed is the cache read — the rate you pay to re-read context the API has already processed.

Anthropic puts the effect at roughly 25% lower total cost for typical workloads and up to about 45% for context-heavy, tool-heavy agentic work — measured against four weeks of real August 2026 usage, not a synthetic benchmark.

The reason the discount lands so unevenly is structural. A chat turn sends a short prompt and pays mostly for output. An agent sends the same system prompt, the same loaded skills, the same tool definitions, and a growing transcript on every turn — so after the first call almost all of its input is a cache read. Cut that line by 75% and you cut the dominant term in the bill. If you run scheduled jobs, long coding sessions, or a bot with a large skill set loaded, you are in the 45% bucket, not the 25% one.

Benchmarks

The Terminal-Bench-Science jump is the outlier worth pausing on: 24.7% to 52.6% in one generation, against 22.4% for GPT-5.6 Sol. Anthropic notes a standard error of roughly 3.5–4.5 points on that benchmark, so treat the gap as large rather than exact. AutomationBench nearly doubling (17.1% to 31.4%) is the second one to note if your agent's job is finishing business workflows rather than writing code.

One caveat Anthropic states directly: Fable 5.1 was evaluated with production safeguards on, and tasks where those safeguards intervened scored zero. The published numbers are therefore a floor, not a ceiling.

What It Did Outside Benchmarks

Three results shipped with the announcement, and they are useful mainly as a read on what a long-horizon agent can now be pointed at:

• Venus elevation map — Fable 5.1 trained a neural network on 30-year-old NASA Magellan radar data and produced a high-resolution elevation map covering about a third of the planet, resolving 2–3km features instead of 10–20km, with heights up to 25% more accurate. Anthropic is releasing it under a Creative Commons license ahead of the NASA VERITAS and ESA EnVision missions.
• Protein binder design — Mythos 5.1, given open-source design and folding tools, produced binders whose affinities were 10× the best entries in Adaptyv Bio's competitions on three targets, with a hit rate near 50% across 12 targets. Typical hit rates in the field are 10–15%. Designs were validated experimentally by external organizations.
• GPU kernels for computational biology — Mythos 5.1 wrote custom GPU kernels for seven open-source genomics and protein models, hitting up to 2.5× inference speedups on an H100 with identical outputs, cutting estimated GPU costs on genome-wide analyses by 30–60%. It worked from public source code in days, against the weeks a performance-engineering team would normally need.

Safeguards: Looser Where It Was Wrong, Tighter Where It Counts

If you have ever had a model refuse to look at your own codebase because the word "exploit" appeared in a comment, this is the release that addresses it. Anthropic reports 60% fewer false positives in cybersecurity, and Fable 5.1 may now be used to discover software vulnerabilities. The line it will not cross is writing working exploits for them.

For agents that review code, triage dependency advisories, or run security linting on a schedule, that materially reduces the class of "my bot just refused a legitimate task" incidents. Deeper offensive work — penetration testing and the like — still sits behind Mythos 5.1's vetted access program, which is currently limited to a set of US organizations.

Anthropic also announced Enterprise Frontier Safeguards, which gives customers zero-data-retention-equivalent privacy by keeping data in infrastructure the customer controls. It rolls out to enterprise customers in phases from later this fall; eligible customers can use Fable 5.1 with zero data retention until then.

Option 1: Run Fable 5.1 on OpenClaw Launch (Easiest — Either Framework)

On OpenClaw Launch you deploy a managed OpenClaw or Hermes Agent instance in about 30 seconds — no server, no config files, no Docker. Fable 5.1 launched in September 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. The steps below are identical on both frameworks — the model selector is the same component:

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-fable-5-1 as the model ID.
4. Click Apply. The saved model becomes anthropic/claude-fable-5-1 and your agent uses it on the next message.

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

If you run your own OpenClaw instance, connect Fable 5.1 through either the Anthropic API directly or through OpenRouter. Direct Anthropic access gives you the effort settings and the fastest path to new features; OpenRouter gives you one balance across every provider.

Direct through the Anthropic API:

"models": {
  "providers": {
    "anthropic": {
      "apiKey": "sk-ant-..."
    }
  }
},
"agents": {
  "defaults": {
    "model": {
      "primary": "anthropic/claude-fable-5-1"
    }
  }
}

Or through OpenRouter:

"models": {
  "providers": {
    "openrouter": {
      "apiKey": "sk-or-..."
    }
  }
},
"agents": {
  "defaults": {
    "model": {
      "primary": "openrouter/anthropic/claude-fable-5.1"
    }
  }
}

Option 3: Self-Hosted Hermes Agent

Hermes Agent ships a native Anthropic provider — it talks the Anthropic Messages API directly with the x-api-key header, rather than tunnelling through an OpenAI-compatible shim. That matters for a model like Fable 5.1, because prompt caching and thinking blocks are Messages-API features and a compatibility shim is where they usually get dropped.

Hermes reads its default model from the model mapping in ~/.hermes/config.yaml — not from an environment variable. An older LLM_MODEL env var still appears in some tutorials; upstream stopped reading it. Note that the mapping keeps the provider in its own key and the model name bare, so Hermes does not use OpenClaw's provider/model prefix form here:

# 1. Give Hermes an Anthropic key (any of these env vars is read)
export ANTHROPIC_API_KEY=sk-ant-...

# 2. Set the default model — interactive picker
hermes model

# …or edit ~/.hermes/config.yaml directly:
#   model:
#     default: claude-fable-5-1
#     provider: anthropic

# 3. New sessions pick it up immediately. Restart the gateway
#    only if you want already-running clients to re-spawn:
hermes gateway restart

Through OpenRouter instead, set OPENROUTER_API_KEY, put openrouter in model.provider, and use OpenRouter's own slug — anthropic/claude-fable-5.1 — as model.default. Full walkthroughs live in the Hermes + Anthropic guide and the Hermes BYOK guide.

Watch the Model ID — the Two Routes Spell It Differently

This is the single most likely thing to break your first deploy, on either framework. Anthropic's own model id uses hyphens; OpenRouter's slug uses a dot in the version number.

Note that the prefix rule is not the same on both sides. OpenClaw always wants provider/model and never the bare name. Hermes puts the provider in model.provider and keeps model.default bare on a native provider — the prefix reappears only when the aggregator's own slug contains one. A wrong slug fails at the relay and reads to the user as "the bot is broken", so confirm it on OpenRouter's model list before deploying. Providers do occasionally adjust slugs after launch.

Fable 5.1 vs Fable 5 vs Opus 5: Which Should Your Agent Run?

Fable 5.1 replaces Fable 5 outright. It is better on every published benchmark, costs the same per token, and costs less per task. There is no scenario where staying on Fable 5 is the cheaper choice — if you are on it, switch.

Against Opus 5, the trade is real. Opus 5 is half the per-token price ($5/$25) and still strong on agentic coding. Fable 5.1 wins on the benchmarks, sometimes by a lot — AutomationBench 31.4% vs 26.9%, Terminal-Bench-Science 52.6% vs 29.0%. Run Opus 5 for high-volume routine agent work where cost dominates, and Fable 5.1 where the task is long, unfamiliar, or expensive to get wrong. Compare everything on the models page.

Best Use Cases for Claude Fable 5.1

• Long-running coding agents — the closed loop of read, plan, edit, test, diagnose, repeat is exactly what the Terminal-Bench and CursorBench gains measure. Pair with a hosted coding workspace.
• Scheduled and background jobs — the cache-read discount compounds on anything that reruns with the same system prompt. See the cron guide.
• Business workflow automation — AutomationBench nearly doubled, which is the benchmark closest to "did the bot actually finish the task".
• Security review and vulnerability triage — now permitted, with 60% fewer false refusals. Exploit development is still out of scope.
• Research and analysis agents — Terminal-Bench-Science more than doubled generation over generation.

FAQ

How much does Claude Fable 5.1 cost?

$10 per million input tokens and $50 per million output tokens — identical to Fable 5. Cache reads dropped 75%, from $1.00 to $0.25 per million, which Anthropic estimates cuts total cost by about 25% on typical workloads and up to about 45% on highly agentic ones. On OpenClaw Launch, managed hosting is billed separately and starts at $3 for the first month, then $6/mo Lite or $20/mo Pro.

What is the difference between Fable 5.1 and Mythos 5.1?

They are the same model with different safeguards. Fable 5.1 is generally available. Mythos 5.1 has safeguards tuned for cybersecurity and life-sciences work and is available only through vetted access programs, currently limited to a set of US organizations. Mythos 5.1 scores higher where its looser guardrails matter — 60.9% vs 55.8% on Terminal-Bench 4.0.

Can I select Fable 5.1 on OpenClaw Launch?

Yes, on both OpenClaw and Hermes instances. It is not yet a one-click preset in the model dropdown, but you can set it today with the custom Model ID field: platform anthropic, model ID claude-fable-5-1. It saves as anthropic/claude-fable-5-1 and applies on your next message.

Does Fable 5.1 work on Hermes Agent?

Yes. Hermes ships a native Anthropic provider that speaks the Anthropic Messages API directly, so prompt caching — the thing that makes this release cheap — works the same way it does on OpenClaw. Self-hosted, set model.default to claude-fable-5-1 and model.provider to anthropic in ~/.hermes/config.yaml (or run hermes model), and provide ANTHROPIC_API_KEY. On OpenClaw Launch, use the same Model ID field described above. See the Hermes + Anthropic guide.

Is the OpenRouter model id different?

Yes, and this trips people up. Anthropic uses claude-fable-5-1 with hyphens; OpenRouter uses anthropic/claude-fable-5.1 with a dot. In OpenClaw config that becomes openrouter/anthropic/claude-fable-5.1.

Should I upgrade from Fable 5?

Yes. Fable 5.1 beats Fable 5 on every published benchmark, charges the same per input and output token, and charges a quarter as much for cache reads. There is no published metric on which Fable 5 is the better buy.

Can I use it for security work?

For defensive work, yes — Fable 5.1 can now be used to discover software vulnerabilities, with 60% fewer cybersecurity false positives than before. It will not write working exploits. Offensive security work such as penetration testing runs through Mythos 5.1's vetted access program instead.

Can I bring my own Anthropic API key?

Yes, on both frameworks. BYOK is supported on eligible plans, so Fable 5.1 usage bills against your own Anthropic account instead of platform credits — usually what you want at this price point. It has to be a pay-as-you-go API key: Anthropic does not permit driving a Claude Pro or Max subscription through third-party agent software. See the OpenClaw BYOK guide or the Hermes BYOK guide.

What's Next?

• Anthropic setup guide — full walkthrough for wiring Anthropic models into OpenClaw
• Claude Opus 5 guide — the cheaper Anthropic option, and when it is the better pick
• BYOK guide — bill model usage to your own provider account
• Hermes + Anthropic setup — the same wiring on the Hermes side
• Managed Hermes hosting — run a Hermes agent without a server
• Connect to Telegram — put your Fable 5.1 agent in a chat app
• Compare all models — pricing and capabilities across every supported model