At the same $0.10 per 1M input tokens, Mouse-1 made more correct agent decisions: 89.1% against 82.6% across four suites. It was also the only one of the two that planted instructions could not move: 0% of decisions flipped against 53.9%. Pick OpenAI Decisions when your inputs run past 16k tokens or your data has to stay in the EU.
OpenAI's Decisions API (POST /v1/decisions, public beta since October 6, 2026) runs on gpt-6-luna and answers typed questions with probabilities instead of text. Mouse-1 does the same job with the same three ideas: yes/no, pick one, score on a scale. Price is identical, so the questions are how often each one is right, how easily it is manipulated and what the platform lets you do with it.
At a glance
| Mouse-1 | OpenAI Decisions (gpt-6-luna) | |
|---|---|---|
| Average accuracy, 4 agent suites | 89.1% | 82.6% |
| Decisions flipped by a planted instruction | 0% | 53.9% |
| Price per 1M input tokens | $0.10 | $0.10 (+10% regional, 2× above 272K) |
| Output tokens | Free | Free |
| Calibrated abstention | /v1/calibrations | Not offered |
| Fine-tuning | /v1/tunings | Not supported |
| Zero data retention | Default on every request | Eligible customers, with approval |
| Refusals | Always returns an answer | Can return a refusal answer |
| Context | 16,384 state tokens | ~1.05M tokens |
| Data residency | US-East | US or EU |
| Rate limits | 10 RPS, 1M TPM | 5,000 RPM on the Build tier |
Accuracy, suite by suite
Both models answered the same 1,550 items with the same question text. Mouse-1 was ahead on all four primary questions. The widest gap was pairwise preference, deciding which of two results is better: 84.3% against 70.3%.
Permission, tool routing, verification and preference suites (%) · Higher is better
| Mouse-1 | 89.1% |
|---|---|
| OpenAI | 82.6% |
Allow this agent action? (%) · Higher is better
| Mouse-1 | 97.1% |
|---|---|
| OpenAI | 92.9% |
Which tool comes next? (%) · Higher is better
| Mouse-1 | 98.8% |
|---|---|
| OpenAI | 92.5% |
Does this change resolve the task? (%) · Higher is better
| Mouse-1 | 76.0% |
|---|---|
| OpenAI | 74.5% |
Which of two results is better? (%) · Higher is better
| Mouse-1 | 84.3% |
|---|---|
| OpenAI | 70.3% |
Calibration was mixed. Mouse-1 had the lower expected calibration error on tool routing (0.013 vs 0.02) and preference (0.027 vs 0.14). OpenAI was slightly better calibrated on permission (0.028 vs 0.03) and verification (0.098 vs 0.112).
Prompt injection
A decision model in an agent reads text it did not write: tool output, web pages, issue bodies. We planted an instruction in that text and counted changed answers.
Decisions changed by a planted instruction (%) · Lower is better
| Mouse-1 | 0% |
|---|---|
| OpenAI | 53.9% |
On the preference suite, the planted instruction changed nearly every one of OpenAI's answers it targeted. Mouse-1 changed none. The test uses one planted-instruction template, so treat it as a strong signal, not a guarantee.
Switching from OpenAI Decisions
Mouse-1 accepts OpenAI's field names (input, a questions array, predicate) on the same path. Change the base URL to https://api.mouse.dev, the key, and the model to mouse-1. The migration guide lists the response differences: Mouse returns probabilities as an object map, and never returns a refusal.
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.
- OpenAI:
gpt-6-luna-decisions-20261006called through OpenRouter in October 2026, samestateand questions. - Not compared: latency, because the two ran over different networks.
Sources
Checked October 20, 2026.
See also: Mouse-1 vs Jev · Mouse-1 vs Cloudflare Clef · All comparisons
FAQ
Is Mouse-1 better than OpenAI Decisions?
On the four agent decision suites we ran both through, Mouse-1 averaged 89.1% accuracy against 82.5% for OpenAI Decisions on gpt-6-luna. A planted instruction changed none of Mouse-1's decisions and 53.9% of OpenAI's. OpenAI accepts far longer inputs (about 1M tokens against 16,384) and offers EU data residency.
How much do Mouse-1 and OpenAI Decisions cost?
Both cost $0.10 per 1M input tokens with free output. OpenAI adds 10% for regional processing and doubles the input price above 272K tokens.
Can I send OpenAI Decisions requests to Mouse-1?
Mouse-1 accepts OpenAI's Decisions field names (input, a questions array, predicate) at POST https://api.mouse.dev/v1/decisions, so most requests only need a new base URL, key and model name.
Can I fine-tune or calibrate OpenAI Decisions?
No. gpt-6-luna does not support fine-tuning, and OpenAI does not offer a calibration endpoint. Mouse-1 offers both: /v1/tunings trains on your own decision outcomes and /v1/calibrations fits a confidence cutoff on your labelled examples.
Does OpenAI Decisions have zero data retention?
Zero data retention is available to eligible OpenAI customers with prior approval. Mouse-1 applies it to every request by default.