Recent complete observations
- Loading current observations…
Updates stopped
Codex Juice
Juice checks and updates have ended. Readings vary by account, so a single account cannot represent everyone’s experience. The former guide remains below for reference.
View the archived model & effort guide
This guide is no longer maintained. The selector uses the original rules; results and model advice may be out of date.
TASK → MODEL → EFFORT
Get a starting setup
Three task signals produce a transparent starting point. If verification fails, follow the fallback instead of guessing from Juice values.
Fast reference
These are starting points, not capability rankings. Verification difficulty and quality priority can move a task upward.
| Task | Start with | Effort |
|---|---|---|
| Quick and easy to check | Luna | Low |
| Routine and scoped | Terra | Medium |
| Complex repository work | Sol | High |
| Ambiguous or high-risk | Astra | XHigh |
30-SECOND GUIDE
Choose in 30 seconds
Start with task shape: use less effort when the path and check are obvious, and more when ambiguity or review risk is high. This guide uses OpenAI's quality-versus-speed direction plus external task results—not Juice values.
How the rule works
Model choice follows task breadth and verification risk; effort then controls how much reasoning to spend. The selector uses OpenAI’s model positioning plus external effort evidence, not Juice fingerprint size.
- Low Use for fast, low-risk work where the path and acceptance check are obvious.
- Medium Use for routine, clearly scoped tasks; it is a balanced starting point when the work is familiar.
- High The original guide used High as a starting point for substantial repository work. Check the result against your own task.
- XHigh Use for ambiguous, cross-module, or high-review-risk work when semantic completeness matters more than time and cost.
- Max The archived NixBench snapshot includes Max. It reflects one external task set and does not establish a universal recommendation.
- Ultra No comparable public evidence is available. Do not infer a recommendation from its Juice fingerprint.
Updates stopped
Runtime checks have ended
Juice values and changes are no longer checked here. The original collection method is retained below for reference.
unavailable
Current self-reported fingerprints
| effort | current | previous | delta | verification |
|---|---|---|---|---|
| Low | — | — | — | unavailable |
| Medium | — | — | — | unavailable |
| High | — | — | — | unavailable |
| XHigh | — | — | — | unavailable |
| Max | — | — | — | unavailable |
| Ultra | — | — | — | unavailable |
Loading current observations…
Confirmed change history
- Loading current observations…
Effort trend
Choose an effort to inspect confirmed values across retained sweeps.
Probe method
A restricted, ephemeral Codex CLI session observes one effort at a time. A new value is published only after a second observation from a fresh session matches it. Full sweeps are invalidated if the model identifier changes mid-run.
What this metric cannot prove
Juice is a non-official, self-reported runtime fingerprint—not actual reasoning-token use or a quality benchmark. A higher value, lower value, or change in either direction cannot be interpreted as stronger or weaker coding ability.
OpenAI provides only directional quality-versus-speed guidance for reasoning effort and says not every task benefits equally from more reasoning. Read OpenAI’s reasoning-effort guidance
Archived benchmark evidence
This retained NixBench snapshot covers GPT-5.6 Sol across five effort levels on 29 coding tasks. It is not a Codex Juice capability score. The data is no longer refreshed here; follow the sources for current results.
| effort | mean passed | 95% interval | seconds/task | trials | timeouts |
|---|---|---|---|---|---|
| Low | 23.6/29 | 22.2–25.0 | 27.4s | 5 | 0 |
| Medium | 24.0/29 | 22.5–25.5 | 38.5s | 5 | 0 |
| High | 23.6/29 | 22.9–24.3 | 47.4s | 5 | 0 |
| XHigh | 24.0/29 | 23.1–24.9 | 60.7s | 5 | 0 |
| Max | 23.2/29 | 20.1–26.3 | 81.4s | 5 | 3 |
Limits: NixBench uses its own harness, corpus and objective hidden shell evaluators; these results are not Codex Juice readings. Max has a public row here. Ultra is a product mode with no comparable public performance row found in this sweep.
Static snapshot · verified 2026-08-03
Open the live NixBench results Read the NixBench method Open the NixBench repository Compare the DeepSWE live board Read OpenAI’s GPT-5.6 context