Anthropic keeps most of its model history live in the API. That means there are currently ten Claude models you can technically use with OpenClaw — and only three you should actually reach for.
Short version: Sonnet 4.6 for everything you do daily, Opus 4.6 for the problems Sonnet can’t solve, Haiku 4.5 when cost matters more than quality. Opus 4.7 was announced yesterday but isn’t widely available yet; this guide reflects what you can deploy today.
The quick answer
| Model | Input/Output Cost | Context | Best For |
|---|---|---|---|
| Claude Sonnet 4.6 | $3 / $15 | 1M | Default for daily coding |
| Claude Opus 4.6 | $5 / $25 | 1M | Hard reasoning, architecture |
| Claude Haiku 4.5 | $1 / $5 | 200K | Budget Claude |
| Claude Sonnet 4.5 | $3 / $15 | 1M | Stable fallback if 4.6 is unavailable |
| Claude Opus 4.5 | $5 / $25 | 200K | Context-bounded Opus work |
| Claude 3.7 Sonnet | $3 / $15 | 200K | Legacy — use 4.6 |
| Claude Opus 4.1 | $15 / $75 | 200K | Legacy — use 4.6 |
Most people should start with Sonnet 4.6 and only reach for something else when the task actually needs it.
Claude Sonnet 4.6 — the default pick
Sonnet 4.6 is where I’d put almost any OpenClaw user. $3/M input, $15/M output, 1M token context, 128K output cap. The context window is the headline — a full million tokens means you can drop a mid-sized codebase in and let the model actually reason across files.
Tool calling is the other reason. OpenClaw leans heavily on the model to produce well-formed function arguments, and Sonnet 4.6 does this more reliably than any GPT model I’ve tested at the same price point. Refusals on code modification tasks are noticeably less frequent than in the 3.7 and 4.5 eras — Anthropic tuned the 4.6 release for agent use.
The 128K output cap deserves a callout. You can ask Sonnet 4.6 to regenerate an entire file in one pass and actually get the full file back, not a truncated middle. For large refactors this is the difference between “the model helps” and “the model works.”
Where it falls short: very hard reasoning problems. If you’re debugging a race condition, designing a new algorithm, or holding five competing hypotheses in your head, Opus 4.6 is noticeably better. But that’s maybe 5% of daily coding work.
Claude Opus 4.6 — the hard-problems model
$5/M input, $25/M output, 1M context, 128K output. Opus 4.6 is what you reach for when Sonnet 4.6 keeps getting it wrong.
Three scenarios where Opus earns its 1.67x premium over Sonnet:
- Architecture reviews. Opus trades off across competing constraints more thoughtfully than Sonnet does. It’ll point out that your proposed database schema breaks an invariant three layers up.
- Hard debugging. Race conditions, memory leaks, deadlocks — the kind of bug where the model needs to trace through several hypotheses before landing on the right one.
- Novel algorithm work. Sonnet can implement algorithms you describe. Opus can reason about which algorithm to pick.
Older Opus versions (4.1 and earlier) were priced at $15/$75 — five times Sonnet. At that tier, Opus was hard to justify. The 4.5 generation dropped Opus pricing to $5/$25, and that’s where it now sits. Worth it for the narrow set of problems Sonnet won’t handle; not worth it as a default.
The 1M context window on 4.6 matters because Opus workloads are often context-heavy. If you’re feeding it an entire service’s worth of code to review, you want the full window.
Claude Haiku 4.5 — the budget Claude pick
$1/M input, $5/M output, 200K context, 64K output. Haiku 4.5 is the first Haiku that feels like a real Claude model, not a shrunken curiosity.
Tool calling works. Long outputs work (the 64K cap means it can generate non-trivial files). The 200K context is narrower than Sonnet’s 1M, but 200K is still enough for most single-task work.
Use Haiku 4.5 when the task is simple enough that Sonnet would be overkill:
- Commit message generation
- File summarization
- Mechanical refactors (rename this variable across the codebase)
- First-pass triage before handing work to Sonnet
The 3x cost gap over Sonnet matters if you’re running OpenClaw all day every day. If you’re using it for occasional pair-programming sessions, just use Sonnet and move on.
Honorable mentions
Claude Sonnet 4.5 ($3/$15, 1M context). Same price and context as 4.6. Slightly less refined on tool calling, slightly more prone to refusals. Keep it in your config as a fallback for when 4.6 hits a capacity issue — same price, no reason not to.
Claude Opus 4.5 ($5/$25, 200K context). Identical pricing to 4.6 but with a 200K context instead of 1M. If you’re running Opus on small-scope problems where 200K is plenty, either works. For anything context-heavy, use 4.6.
Claude 3.7 Sonnet ($3/$15, 200K). Was the daily driver a year ago. Sonnet 4.6 is better on every axis for the same price. No reason to deploy 3.7 new.
Claude Opus 4.1 and older ($15/$75). These are still in the API at the old price tier. Anthropic hasn’t deprecated them, but at 3x the cost of 4.6 with worse performance, there’s no scenario where they’re the right pick.
Claude 3 Haiku and 3.5 Haiku ($0.25–$0.80 input). Cheap, but the 4K–8K output caps make them useless for real coding work. Haiku 4.5’s 64K output cap is the minimum you want for OpenClaw.
Claude Opus 4.7 (just-announced)
Anthropic announced Opus 4.7 on April 16, 2026 — one day before this guide was written. Stronger reported performance on coding, agents, vision, and multi-step tasks. I haven’t tested it with OpenClaw yet and it isn’t in OpenRouter’s model catalog as of writing. Check back in a week. If the jump from 4.6 to 4.7 is similar to the jumps we’ve seen in the 4.x line, it’ll become the new default for hard problems within a month.
Setup in OpenClaw
Anthropic is a built-in provider. You don’t need a providers block — just export your API key:
export ANTHROPIC_API_KEY="your-anthropic-api-key"
Then add the models you want to the allowlist in ~/.openclaw/openclaw.json:
{
"agents": {
"defaults": {
"models": {
"anthropic/claude-sonnet-4-6": {},
"anthropic/claude-opus-4-6": {},
"anthropic/claude-haiku-4-5": {}
}
}
}
}
Apply with openclaw gateway config.apply and switch models with /model during a session.
Running through haimaker.ai
All Claude models are also available through haimaker.ai with a single API key. If you’re already routing other providers through haimaker, you don’t need a separate Anthropic account:
{
"models": {
"providers": {
"haimaker": {
"baseUrl": "https://api.haimaker.ai/v1",
"apiKey": "your-haimaker-api-key",
"api": "openai-completions"
}
}
}
}
This gives you Claude alongside GPT, Gemini, DeepSeek, Grok, and dozens of other models through one provider — useful when you want to A/B Claude against alternatives without standing up a new account each time.
What I’d do
Set Sonnet 4.6 as your default. Keep Opus 4.6 configured as a second option for hard problems. Add Haiku 4.5 if you find yourself burning money on tasks that don’t need a flagship model.
Don’t bother with the legacy versions unless you have a specific reason. Anthropic pricing has compressed over the last year — Opus 4.6 at $5/M is basically the same tier that Sonnet used to occupy. Pick based on capability, not on historical pricing.
And keep an eye on Opus 4.7. If it delivers what the announcement promises, it’ll change this recommendation within a month.
For a complete model comparison, see our guide to all models in OpenClaw. For API key setup, see the OpenClaw API key guide.