When David Ondrej posted a clip on April 25 showing "Gemini 3.1 Pro" repeating ten messages in a row in OpenClaw — quoting itself, not stopping, and eventually freezing — he hit the nail on the head with the question every team using OpenClaw eventually asks: Which model really works in this environment, and which ones quietly fall apart? Marketing benchmarks don't answer this question. Leaderboards don't either. Only real-world deployment does.
We've been running OpenClaw in customer projects for several months, swapping out models as new ones become available. This is the practical guide we wish we had at the beginning: which models we trust in OpenClaw today, which ones we no longer use, and how to decide for your own workload.
What "Best" Means in OpenClaw
OpenClaw isn't a chat box. It's an agentic framework: tool usage, file editing, long task loops, persistent context, hooks, and a CLI for Cron-like automation. A model with 90 points in a coding benchmark can still be the wrong choice if it ignores OpenClaw's hook conventions, breaks tool-call contracts, or burns the context window by reading the same file four times in a row.
Three qualities matter more than benchmark numbers:
- Tool-Call Discipline — does it call the right tool with the right schema on the first try?
- Stop Discipline — does it know when the task is finished, or does it loop?
- Context Economy — does it re-read what's already in the context, or does it trust it?
Almost everything else follows from this. We evaluate each model in our deployment notes based on these three points. Here's how the most important frontier options look at the end of April 2026.
Sonnet 4.6 — the Default That Earned Its Place
Anthropic Claude Sonnet 4.6 is the model we set as the default in almost every OpenClaw deployment — and it earned this position, it wasn't given. Tool-call discipline: excellent. Stop discipline: best in class — when a Sonnet 4.6 task is finished, it stops. It rarely re-reads files that it has just been shown. The cost-per-task ratio for typical agent workflows is roughly where Haiku was before the price adjustment.
Where it falters: deep multi-step refactors in unfamiliar codebases sometimes need a sharper blade. For code reviews, architectural decisions, or debugging traces with long cause-and-effect chains, Sonnet 4.6 occasionally gives up too early. That's exactly when we turn to Opus.
For a deeper understanding of why agentic work has shifted towards this model, see our post on the agentic working model that OpenAI delivered against the Claude myth.
Opus 4.7 — When Reasoning Depth Really Pays Off
Opus 4.7 is the heavyweight. We don't set it as the default because the cost per task quickly adds up — but it's our escalation path for three job classes:
- Complex Debugging, where the cause-and-effect chain spans multiple files and the symptoms are misleading
- Architectural Decisions, where the model must honestly weigh trade-offs instead of choosing the first plausible answer
- High-Risk One-Shot Tasks such as migration scripts, schema changes, or anything that touches production data
Opus 4.7 is also the model we trust most when adaptive thinking counts — that is, when the model should invest reasoning tokens before committing to a tool call. The costs are real, but the success rate on hard tasks justifies them. Rule of thumb: If a Sonnet run fails twice with similar errors, escalate to Opus instead of retrying.
GPT-5.5 in OpenClaw — Strong Coder, Wrong Tool for Now
GPT-5.5 in OpenClaw is both interesting and frustrating. As a pure coder, it's strong, and OpenAI's confirmation that GPT-5.5 = Codex (Romain Huet, April 25) means there's no longer "Codex for coding, GPT-5.5 for general tasks." One model, two frameworks.
But OpenClaw isn't its framework. We see two recurring error patterns when incorporating GPT-5.5 into OpenClaw:
- Tool-Call-Schema-Drift — it occasionally invents tool fields that don't exist, especially in longer contexts
- Looser Stop Discipline — it produces filler text of the "I'm continuing to work on it" type more frequently than Sonnet 4.6
For OpenClaw, our recommendation today is: Leave GPT-5.5 in the Codex CLI, where its conventions match its training, and keep Anthropic models in OpenClaw. This may shift as the framework evolves around other providers. We test with every minor release.
DeepSeek V4 — Cost Disruption That Needs Real Testing
DeepSeek V4 (1.6 Trillion parameters, MIT license, drastically cheaper than Opus for comparable tasks) has just landed. We covered the price implications extensively in our DeepSeek V4 Price Earthquake post.
In OpenClaw, our initial tests show that DeepSeek V4 Flash handles 70-80% of typical Haiku workloads at a fraction of the cost. V4 Pro is seriously competitive with Opus on isolated reasoning tasks, but its stop discipline lags behind Anthropic models. We're not ready to recommend it as the default in customer deployments yet — too early to know how it behaves in long agent loops, and the open-weights variant requires self-hosting infrastructure that most teams don't have.
If you're cost-sensitive and willing to invest in evaluation: Start with V4 Flash on low-risk OpenClaw cron jobs (Intel scans, summaries, content quality checks) and measure stop discipline and tool-call accuracy. Don't deploy it in production-like jobs until you have a stability history of several weeks.
Models We Tested and Do Not Recommend
Some signals from the last six months of deployment work:
- Gemini 2.5 Pro and 3.x Variants: Inconsistent in OpenClaw. The Ondrej report matches our own tests. The model is strong in its own environment (Vertex, AI Studio) but does not reliably respect OpenClaw's tool conventions. We are no longer accessing it.
- Nemotron and Qwen Mid-Tier: Good as a fallback in our Cost-Cascade, but timeout rates increase in longer contexts. Only suitable for short bursts.
- Older Claude Versions (3.5, 4.0, 4.5): Outdated. No reason to use these unless for cost reasons.
More provider context can be found in our analysis of agentic Compute Pricing.
How to Choose: A Decision Matrix
Here's how we proceed when a customer asks: "Which model should we use?"
| Workload | Default | Escalate to | Why |
|---|---|---|---|
| Daily Cron Jobs (Audit, Scan, Summarize) | Sonnet 4.6 | Opus 4.7 if accuracy is critical | Sonnet's stop discipline keeps costs predictable |
| Code Generation and Review | Sonnet 4.6 | Opus 4.7 for hard bugs | Skip GPT-5.5 in OpenClaw; use in Codex-CLI |
| High-Risk One-Shot Tasks (Migrations, Prod Fixes) | Opus 4.7 | — | Costs justified by single-failure costs |
| Cost-Sensitive Bulk Work | DeepSeek V4 Flash (Test) | Sonnet 4.6 | Validate stop discipline before scaling |
| Multimodal Tasks (Vision, Audio) | Sonnet 4.6 with Vision | Opus 4.7 | Most consistent behavior in OpenClaw |
The lever we pull most often: escalate from Sonnet to Opus on retry, never the other way around. If Sonnet fails twice, Opus usually solves it on the first try. If Opus fails, a retry with Opus rarely helps — the task probably needs different framing.
FAQ
Q: Can I switch models mid-task in OpenClaw? Yes — OpenClaw supports model switching via CLI flag or per-job config. We use this in our Cron fallback cascade: If Sonnet times out twice, the next run automatically tries a different provider. Set this up before you need it.
Q: Is "Gemini 3.1 Pro" actually a released model? As of April 26, 2026, we find no official Google announcement for "Gemini 3.1 Pro". David Ondrej's video may refer to an internal name or a silent rollout. Treat any "Gemini 3.x" claim as unconfirmed until Google's blog confirms it.
Q: Should I always use the strongest model "just to be safe"? No. Opus 4.7 for a job that Sonnet 4.6 does well costs five to ten times as much for the same result, and the longer reasoning loops can introduce their own error patterns. Match the model to the workload — escalate on retry, not by default.
Q: How often should I retest which model is the best fit? At least monthly, weekly if you run productive OpenClaw deployments. Model behavior shifts after each minor release, and framework compatibility changes faster than benchmark values suggest.
Conclusion
For most teams running OpenClaw in 2026, the right default is Sonnet 4.6 with Opus 4.7 as an escalation path. GPT-5.5 belongs in the Codex-CLI, not in OpenClaw. DeepSeek V4 is worth evaluating for cost-sensitive workloads, but is not yet production-ready in this framework. Gemini variants remain inconsistent.
Benchmark values will continue to shift. What counts in OpenClaw is Tool-Call discipline, Stop discipline, and Context Economy — and on these three axes, the Anthropic models are still ahead.
If you need help setting up the right model cascade for your OpenClaw deployment — defaults, fallbacks, escalation rules — schedule a discovery call with Context Studios. We've done this for enough customers to skip the trial-and-error phase.