For the decisions an agent harness makes, Mouse-1 is more accurate (89.1% vs 84.3%), unmoved by planted instructions (0% vs 37.5% flipped) and cheaper: $0.10 per 1M input tokens against Clef's $0.24. Pick Clef if you want open weights to host yourself, or inputs longer than 16k tokens.
Clef is Cloudflare's decision model, launched October 1, 2026 on Workers AI with open weights. It answers System One–style questions with calibrated probabilities, as Mouse-1 does. Cloudflare also ships Clef-flash, a smaller and cheaper model, and Clef-omni, which reads audio and video. This page compares Mouse-1 with Clef itself, the model we ran.
At a glance
| Mouse-1 | Clef | |
|---|---|---|
| Average accuracy, 4 agent suites | 89.1% | 84.3% |
| Decisions flipped by a planted instruction | 0% | 37.5% |
| Public tool benchmarks (avg of 4) | 84.9% | 85.6% |
| Price per 1M input tokens | $0.10 | $0.24 |
| Output tokens | Free | Free |
| Weights | Hosted API | Open, Apache 2.0 |
| Context | 16,384 state tokens | 65,536 tokens |
| Images | 8 per request (beta) | 4 per request |
| Calibrated abstention | /v1/calibrations | Not offered |
| Fine-tuning | Self-serve /v1/tunings | Run by Cloudflare engineers |
| Data retention | Zero by default | Not stored or trained on |
Accuracy on agent decisions
Same 1,550 items, same question text. Mouse-1 led on all four primary questions. The biggest gap was verification, where the model decides whether a change actually resolves the task: 76% against 71%.
Permission, tool routing, verification and preference suites (%) · Higher is better
| Mouse-1 | 89.1% |
|---|---|
| Clef | 84.3% |
Allow this agent action? (%) · Higher is better
| Mouse-1 | 97.1% |
|---|---|
| Clef | 92.0% |
Which tool comes next? (%) · Higher is better
| Mouse-1 | 98.8% |
|---|---|
| Clef | 96.0% |
Does this change resolve the task? (%) · Higher is better
| Mouse-1 | 76.0% |
|---|---|
| Clef | 71.0% |
Which of two results is better? (%) · Higher is better
| Mouse-1 | 84.3% |
|---|---|
| Clef | 78.0% |
Mouse-1's probabilities were also better calibrated on every primary question. Expected calibration error was 0.03 against 0.085 on permission, 0.013 against 0.061 on tool routing, 0.112 against 0.155 on verification and 0.027 against 0.135 on preference. Both models caught dangerous actions equally well (98.9% each).
Prompt injection
Of the three hosted decision APIs we tested (Jev, OpenAI Decisions and Clef), Clef resisted planted instructions best, and it still followed a planted instruction on 37.5% of items. Mouse-1 followed it on none.
Decisions changed by a planted instruction (%) · Lower is better
| Mouse-1 | 0% |
|---|---|
| Clef | 37.5% |
Most of Clef's flips were on the preference suite, where untrusted text sits right next to the thing being judged. The test uses one planted-instruction template, so read it as a signal rather than a guarantee.
Public tool benchmarks: where Clef leads
On the four tool benchmarks of the Jev Decision Index, Clef averages 85.6% to Mouse-1's 84.9%. Clef is ahead on BFCL (98.5 vs 97.6), API-Bank (91.9 vs 83.7) and PhishNChips (79.4 vs 72.3). Mouse-1 is well ahead on When2Call, which tests whether to call a tool at all: 85.8% against 72.5%.
How we measured
- Suites: agent permission and risk (350 items), tool routing (400), verification (400) and pairwise preference (400). Tool routing, verification and preference were held out of Mouse-1's training; the permission suite is closer to its training data.
- Clef:
cloudflare/clefcalled through OpenRouter in October 2026. - Public benchmarks: Clef's scores are from the Decision Index board. Mouse-1 is not on the board; its scores come from our own run of the board's kit on the same rows.
- Not compared: latency. Cloudflare reports about 150 ms median for Clef and Mouse-1's server time is about 70–90 ms, but the two were measured differently.
Sources
Checked October 20, 2026.
- Cloudflare: Clef decision models
- Cloudflare: Clef faster, cheaper, multimodal
- Workers AI Clef model page
- Jev Decision Index board data
- Mouse-1 API reference and pricing
See also: Mouse-1 vs Jev · Mouse-1 vs OpenAI Decisions · All comparisons
FAQ
Is Mouse-1 better than Cloudflare Clef?
For agent decisions, yes on our tests: Mouse-1 averaged 89.1% accuracy across four agent suites against 84.2% for Clef, and a planted instruction changed none of its decisions against 37.4% of Clef's. Clef scores slightly higher on the four public tool benchmarks of the Jev Decision Index (85.6% vs 84.9%) and ships open weights.
Which is cheaper, Mouse-1 or Clef?
Mouse-1, at $0.10 per 1M input tokens against $0.24 for Clef on Workers AI. Both charge nothing for output. Clef-flash, the smaller model, costs $0.038.
Is Clef open source?
Clef's weights are published under Apache 2.0, so you can host it yourself. Mouse-1 is a hosted API only.
Do Mouse-1 and Clef use the same API?
Both use the System One question types (choice, noul, score). Clef runs through Workers AI (@cf/cloudflare/clef); Mouse-1 runs at api.mouse.dev and also accepts the System One path, /v1/systemone.