Which subscription does the work?
Atris Bench gives paid AI tools the same practical work, then publishes the result and the transcript behind it.
Hard pack
Eight longer tasks with hidden tests and traps. Built to separate the top of the floor pack.
| Rank | Model and subscription | Reliable pass3 | Pass at one | Median time | Tasks won | Transcripts |
|---|---|---|---|---|---|---|
| 1 | Muse Spark 1.3opencode (contributor free) · $0/mo | 100% | 100% | 40s | 3 | 24 |
| 2 | GPT-5.6 Sol (Codex)ChatGPT Pro · $200/mo | 100% | 100% | 45s | 1 | 24 |
| 3 | Grok 4.6SuperGrok · $30/mo | 100% | 100% | 1m 12s | 0 | 24 |
| 4 | Gemini 3.8 Flash (High)Antigravity · $0/mo | 75% | 87.5% | 1m 45s | 0 | 24 |
| 5 | Claude Fable 5.1Claude Max · $200/mo | 62.5% | 100% | 32s | 4 | 24 |
| Model | Fix three hidden billing bugs | Reconcile a trapped quarter ledger | Apply superseding policy | Triage a trapped inbox | Schedule across four timezones | Rename an API without breaking callers | Own an outage escalation | Turn a local brief into a page |
|---|---|---|---|---|---|---|---|---|
| Muse Spark 1.3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 |
| GPT-5.6 Sol (Codex) | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 |
| Grok 4.6 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 |
| Gemini 3.8 Flash (High) | 2/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 2/3 | 3/3 |
| Claude Fable 5.1 | 3/3 | 3/3 | 3/3 | 2/3 | 3/3 | 3/3 | 2/3 | 2/3 |
Business pack
Eight practical tasks covering everyday business work.
| Rank | Model and subscription | Reliable pass3 | Pass at one | Median time | Tasks won | Transcripts |
|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.1Claude Max · $200/mo | 100% | 100% | 19s | 5 | 24 |
| 2 | Muse Spark 1.3opencode (contributor free) · $0/mo | 100% | 100% | 24s | 2 | 24 |
| 3 | GPT-5.6 Sol (Codex)ChatGPT Pro · $200/mo | 100% | 100% | 34s | 0 | 24 |
| 4 | Grok 4.6SuperGrok · $30/mo | 100% | 100% | 35s | 0 | 24 |
| 5 | Gemini 3.8 Flash (High)Antigravity · $0/mo | 75% | 87.5% | 52s | 0 | 24 |
| Model | Find the payout mismatches | Triage a small inbox | Fix two invoicing bugs | Find shared calendar slots | Remove dead JavaScript exports | Reply to a dental customer | Write the owner's morning brief | Answer contract questions |
|---|---|---|---|---|---|---|---|---|
| Claude Fable 5.1 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 |
| Muse Spark 1.3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 |
| GPT-5.6 Sol (Codex) | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 |
| Grok 4.6 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 |
| Gemini 3.8 Flash (High) | 3/3 | 3/3 | 3/3 | 3/3 | 2/3 | 2/3 | 3/3 | 3/3 |
How it works
The tasks are ordinary business work, including reconciling payouts, triaging inboxes, fixing small bugs, and reading contracts.
Each model runs through the agent command line tool included with the subscription a buyer would pay for.
Every model receives the same fresh folder and the same prompt, with tools available and no hints.
Code checks decide the result where possible. When judgment is needed, a model from a rival company judges the work.
Reliable pass measures whether every attempt passed. Pass at one shows how often the first attempt worked.
Every result links to a transcript. New task packs arrive weekly, and old packs remain available for trend checks.
What this does not measure
Atris Bench does not measure raw API quality, long-horizon work, or cost per token. These products sell flat-rate subscriptions, so the comparison shows completion time instead of token cost.