Model guide
Claude Sonnet 5.5 for Hermes Agent and OpenClaw
Anthropic's second Claude 5.5 model keeps Sonnet 5's $2/$10 price and comes close to Opus 5.5 on coding and knowledge-work benchmarks. How to select it on a managed Hermes or OpenClaw instance, which provider row to pick, the thinking setting that now returns an error, and self-hosted config for both.
Reach Sonnet 5.5 from a managed instance
Anthropic released Claude Sonnet 5.5 on September 28, 2026, the second model in the Claude 5.5 family after Opus 5.5. You can move an existing Hermes Agent or OpenClaw instance onto it from the instance card, with no new instance required. The steps are the same on both frameworks, but which provider row you should pick is not, and that choice decides which account pays.
- Open the dashboard and find your running instance.
- Click the current model in the Text or Model row on the instance card.
- Pick the provider row that matches your framework: Anthropic on Hermes if you have saved your own Anthropic key, otherwise OpenRouter. The next section explains why.
- Search for sonnet 5.5. It appears as
anthropic/claude-sonnet-5.5on the OpenRouter row, and as Claude Sonnet 5.5 on the Anthropic row. The featured Anthropic shortlist still stops at Sonnet 5, so use the search box rather than scanning the highlighted models. - Wait for the switch to finish and the instance to return to ready. Confirm the new name on the card, and start a fresh conversation if an older session is still pinned to its previous model.
Full-catalogue access applies to eligible paid instances; trial accounts use a restricted model list. If account credits run out, top up or add your own OpenRouter key on the API keys page. With a personal key in place, the tokens are billed to that OpenRouter account.
Which provider row to pick
- Hermes, with your own Anthropic key saved. Pick the Anthropic row. Hermes sends an
anthropic/selection straight toapi.anthropic.comasclaude-sonnet-5-5, charged to your own key. Hermes treats a Claude model it has not seen before as an adaptive-thinking model, so it already sends the request shape Sonnet 5.5 expects. There is one setting to avoid, covered in "Turning thinking off" below. - Hermes, with no Anthropic key saved. The same selection falls through to OpenRouter and spends OpenRouter credit. That works, but pick the OpenRouter row deliberately so the billing is not a surprise.
- OpenClaw. Use the OpenRouter row. An OpenClaw container resolves a model against the catalogue built into its own release. OpenClaw 2026.9.6, the newest release, already knows
claude-opus-5-5but has no entry forclaude-sonnet-5-5, because Sonnet 5.5 shipped after it. The OpenRouter route does not depend on that catalogue.
Sonnet 5.5 or Opus 5.5?
Anthropic describes Sonnet 5.5 as the faster, cheaper complement to Opus 5.5. It is strongest at well-scoped everyday tasks, fixing bugs, and producing documents, slides and spreadsheets. Opus 5.5 is aimed at complex work that needs careful judgment.
For an agent that mostly answers messages, runs tools and handles routine jobs, Sonnet 5.5 at half the Opus token price is the sensible default. Keep Opus 5.5 for long, open-ended tasks where a wrong call is expensive. Anthropic says so itself: in its own testing and external testers', "Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment", even where the benchmark numbers are close.
Benchmarks
These results come from Anthropic's announcement:
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 70.6% | 10.3% | 66.4% |
| FrontierCode 1.1 (main) | 52.1% at xhigh, 46.2% at max | 42.4% | 54.4% |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% |
| GDPval-AA v2.1 (knowledge work, Elo) | 1844 | 1449 | 1846 |
| AA-Briefcase v1.1 | 1811 | 1359 | 1822 |
| Humanity's Last Exam (with tools) | 64.5% | 54.9% | 67.7% |
| OSWorld 2.1 (computer use, partial) | 80.1% | 57.0% | 81.8% |
| Chartography (chart reading, no tools) | 61.6% | 15.6% | 64.4% |
Three details are worth knowing before you read the table as "Sonnet caught up with Opus":
- Terminal-Bench 4.0 is the only row Sonnet 5.5 wins. Opus 5.5 is ahead everywhere else, although by small margins on GDPval-AA, AA-Briefcase, OSWorld and CursorBench. The Opus figure of 66.4% is its xhigh-effort score, its best on that test.
- More effort is not always better. On FrontierCode, Sonnet 5.5 scored lower at max effort than at xhigh. Anthropic's footnote explains that at max effort it more often ran Claude Code's code-review skill, which splits the review across many subagents. In some runs this caused a timeout or edits outside the task's scope, and the benchmark penalises both.
- The "tenth of the cost" claim has a specific scope. Anthropic says that on several benchmarks, Sonnet 5.5 at low or medium effort beats Sonnet 5's best score for about a tenth of the cost per task. That compares low effort with Sonnet 5's top setting. It is not a general 90% saving.
Anthropic also reports that Sonnet 5.5 batches tool calls together more often than Sonnet 5. That means fewer steps per task, which matters for agents that run long tool chains. It is also the first Sonnet model to finish Pokémon Red working only from screenshots, a sign of its long-horizon and image skills.
Pricing
| Price per 1M tokens | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Input | $2 | $2 | $4 |
| Output | $10 | $10 | $20 |
| Cache reads | $0.20 | $0.20 | $0.20 |
| Cache writes | $2.50 | $2.50 | $5 |
The per-token price is unchanged from Sonnet 5. The saving comes from Sonnet 5.5 needing fewer tokens to finish the same work: Anthropic measured up to 30% less per task, and Slack reported about 14% fewer output tokens on its own evaluations. Output is also more than 30% faster.
OpenRouter lists anthropic/claude-sonnet-5.5 at the same $2 and $10, with a 1M-token context window and up to 128K output tokens (verified September 29, 2026, UTC). There is also a :batch entry at half price, but it is for asynchronous jobs, not live chat. These are token prices, not your hosting subscription. Check the model page before committing to a long run.
Effort level moves the bill more than anything else. Sonnet 5.5 has five levels: low, medium, high, xhigh and max. Claude's own apps default to medium; the API defaults to high. Higher levels think for longer and spend more output tokens, so compare on your own tasks at the level you intend to use.
Turning thinking off works differently now
This is the change most likely to break an existing setup. On Sonnet 5 you could turn thinking off with thinking: {"type": "disabled"}. Sonnet 5.5 rejects that with a 400 error. The lowest setting is now thinking: {"type": "between_tools"}, which accepts low, medium and high effort only. Per Anthropic's migration guide, these also return 400 errors on Sonnet 5.5: thinking budgets (budget_tokens), non-default temperature/top_p/top_k, assistant prefill, and forced tool choice.
What that means here:
- Hermes on the direct Anthropic route. Hermes turns
reasoning_effort: "none"into exactly thedisabledrequest above, so every turn fails. That is true of the current upstream code and of Hermes 2026.9.24. Managed OpenClaw Launch instances do not setreasoning_effort, so they use the default and are not affected unless you changed it. For the lightest thinking, uselowinstead. The same applies to any auxiliary task block (vision, compression and so on) that setsreasoning_effort: "none"while running onclaude-sonnet-5-5. - Security work. Sonnet 5.5 is the first Sonnet launched with the cyber safeguards Anthropic uses on its most capable models. Routine bug finding and fixing is unaffected. Anthropic's announcement says higher-risk cybersecurity requests visibly fall back to Sonnet 5, but on the API the default is different: the reply is a normal response with
stop_reason: "refusal", and it only retries on another model if the request opts in with the betafallbackssetting. Hermes does not set it, so if a security task is refused, run that task on a different model. - Moving conversations between accounts. Sonnet 5.5 thinking blocks work only in the account that produced them. A replayed block after editing earlier history returns a 400 on newer accounts, so keep conversation history append-only.
Self-hosted Hermes Agent
On your own Hermes installation, run hermes model and pick the provider and model, or set the model section of ~/.hermes/config.yaml directly. With your own Anthropic key:
model:
provider: anthropic
default: claude-sonnet-5-5
The provider has its own key, so the model value stays bare, with no anthropic/ prefix. This route needs ANTHROPIC_API_KEY in your environment or the Hermes credential store. Through OpenRouter instead, the default carries OpenRouter's own dotted slug:
model:
provider: openrouter
default: anthropic/claude-sonnet-5.5
Thinking effort is set separately, under agent rather than model:
agent:
reasoning_effort: "medium"
Use low, medium or high for everyday agent work. Do not use none on the direct Anthropic route, for the reason above. A model change applies to new sessions; use hermes gateway restart if you want to restart the gateway outright. Field names come from the upstream configuration example.
Self-hosted OpenClaw
No OpenClaw release so far, up to 2026.9.6, includes claude-sonnet-5-5 in its Anthropic catalogue, so route through OpenRouter. Add your OpenRouter key through onboarding, then set the model:
openclaw onboard --auth-choice openrouter-api-key
openclaw models set openrouter/anthropic/claude-sonnet-5.5
The matching section of ~/.openclaw/openclaw.json is below. Merge it into your existing configuration rather than replacing the file:
{
"agents": {
"defaults": {
"model": { "primary": "openrouter/anthropic/claude-sonnet-5.5" }
}
}
}
Once an OpenClaw release adds the model to its Anthropic catalogue, the direct route becomes anthropic/claude-sonnet-5-5 with --auth-choice anthropic-api-key. Watch the spelling: Anthropic's own id is hyphenated (claude-sonnet-5-5), while OpenRouter's slug is dotted (claude-sonnet-5.5). Managed OpenClaw Launch customers do not edit either file; use the instance-card steps above.
Sources and related guides
- Anthropic: Introducing Claude Sonnet 5.5
- Anthropic: Migrating to Claude Sonnet 5.5
- OpenRouter: anthropic/claude-sonnet-5.5
- Claude Opus 5.5 guide and Claude Opus 5 guide
- Hermes Agent + Anthropic and OpenClaw + Anthropic
- Hermes hosting and OpenClaw hosting
Frequently asked questions
Does Claude Sonnet 5.5 work with Hermes Agent?
Yes. With your own Anthropic key saved, selecting Sonnet 5.5 on the Anthropic row sends it straight to api.anthropic.com as claude-sonnet-5-5. Without an Anthropic key it routes through OpenRouter as anthropic/claude-sonnet-5.5. Self-hosted Hermes takes provider: anthropic with default: claude-sonnet-5-5. Avoid reasoning_effort: none on the direct route, because Sonnet 5.5 rejects the thinking-disabled request it produces.
Does Claude Sonnet 5.5 work with OpenClaw?
Yes, through OpenRouter. Pick the OpenRouter row on the instance card and search for sonnet 5.5, or set openrouter/anthropic/claude-sonnet-5.5 on a self-hosted install. No OpenClaw release up to 2026.9.6 has claude-sonnet-5-5 in its built-in Anthropic catalogue yet, so the direct Anthropic route does not resolve.
Why is Sonnet 5.5 not in the featured Anthropic model list on my dashboard?
The featured shortlist is curated and still ends at Claude Sonnet 5. Sonnet 5.5 is reachable today by searching the full catalogue for sonnet 5.5 on the OpenRouter row, or on the Anthropic row if you use Hermes with your own Anthropic key.
Is Sonnet 5.5 cheaper than Sonnet 5?
Not per token: both cost $2 input, $10 output and $0.20 for cache reads per million tokens. Sonnet 5.5 is cheaper per task because it needs fewer tokens for the same work. Anthropic measured up to 30% less per task. Higher effort levels spend more thinking tokens, so compare at the level you intend to use.
Should I use Sonnet 5.5 or Opus 5.5 for my agent?
Sonnet 5.5 for everyday agent work: messages, tool use, bug fixes, documents. It costs half as much per token as Opus 5.5 and is close on most benchmarks. Opus 5.5 for long, open-ended tasks that need sustained judgment, where Anthropic says it remains clearly stronger.
Will my Sonnet 5 instance upgrade to Sonnet 5.5 automatically?
No. Sonnet 5.5 is a separate model id. Select it on the instance card and wait for the change to finish, then start a fresh session if an existing conversation is still on the old model.