Mistral Large 4 League Score correction: one wrong LiveBench row hid 41.7 points.
The Mistral Large 4 League Score was corrected on November 9, 2026 — from 11.1 to 52.8 — after our overnight cross-check against LiveBench found the database had been scoring the model's API-preview checkpoint (43.9) instead of the open-weights checkpoint (63.5) it has tracked since October 30. The model did not get faster or smarter overnight; the record was wrong, and this post shows the two rows, the maths, and every ranking the fix moves.
By AI Pro League · Published

The Mistral Large 4 League Score published on this site until today was wrong, and the error was ours. Our November 9, 2026 changelog records the fix in full: Mistral Large 4’s LiveBench raw score is corrected to 63.5 per LiveBench 2.0, replacing 43.9, and its League Score is recomputed to 52.8. The same cross-check matched every other tracked LiveBench 2.0 score (22 models) exactly, and no tracked model’s price had changed on its official pricing page — this was a single-row mix-up, not a field-wide revision. What follows is the before-and-after, computed from our November 8 snapshot and the corrected database.
Mistral Large 4 League Score: before and after the correction
| Measure | Before | After | Change |
|---|---|---|---|
| LiveBench raw score | 43.9 | 63.5 | +19.6 |
| League Score | 11.1 | 52.8 | +41.7 |
| Overall rank (of 45 scored) | #41 | #31 | +10 places |
| Open-weight rank | #18 | #10 | +8 places |
| Value Division rank (points per blended $) | #28 (10.6 pts/$) | #7 (50.3 pts/$) | +21 places |
Nothing about the model moved: the price is still $0.70 input / $2.10 output per million tokens (cached $0.070), so the standard 10M-input + 2M-output workload in our AI model pricing index still costs $11.20 — the same figure as in our Mistral Large 4 price cut analysis. Only the performance record changed, and with it every ranking that record feeds: the AI model rankings, the open-weight table and the Value Division.
Two LiveBench rows, two different checkpoints
LiveBench carries two Large 4 rows. The 43.9 recorded here on October 26, 2026 belongs to the original API-preview checkpoint. The “Mistral Large 4 1029 open” row at 63.5 is the open-weights checkpoint released on October 29, 2026 — the record this database has tracked since the model’s October 30 graduation to open weights, when its status moved from preview to current at the $0.70 / $2.10 price. Scoring the open checkpoint is also consistent with how every other open-weight model in the database is scored: the downloadable weights are the thing being ranked. Our LiveBench leaderboard now carries the corrected row, and our LiveBench explainer sets out why this benchmark carries 40% of the League Score — the joint-largest weight of the five components.
The normalisation maths: how 63.5 becomes 52.8
Large 4’s League Score rests on this single component, so its score is exactly its normalised LiveBench value — which makes the correction checkable by hand. Normalisation stretches the recorded LiveBench field — 31 models, raw scores from 38.7 to 85.7, median 69.8 — across a 0–100 scale:
| Row | Raw | Normalised |
|---|---|---|
| API-preview row (wrong record) | 43.9 | (43.9 − 38.7) ÷ 47.0 = 11.1 |
| Open-weights row (corrected) | 63.5 | (63.5 − 38.7) ÷ 47.0 = 52.8 |
The published scores — 11.1 then, 52.8 now — are those two normalised values. A 19.6-point raw gap became a 41.7-point League gap because normalisation stretches the field: the same stretching our MMLU-Pro explainer warns about on tight raw fields applies here in the model’s favour.
Where the corrected score sits in the League
| Rank | Model | League Score | Coverage |
|---|---|---|---|
| 27 | Mistral Medium 3.5 | 59.1 | 20% |
| 28 | Gemini 3.1 Pro Preview | 58.4 | 90% |
| 29 | Gemini 3.8 Flash | 57.8 | 55% |
| 30 | GPT-6 Luna | 55.1 | 40% |
| 31 | Mistral Large 4 (corrected) | 52.8 | 40% |
| 32 | Muse Spark 1.3 | 51.7 | 40% |
| 33 | Llama 4 Maverick | 47.7 | 15% |
| 34 | Qwen3-Coder-480B-A35B-Instruct | 41.3 | 20% |
| 35 | Gemma 4 26B A4B | 40.4 | 15% |
At #10 among scored open-weight models, Large 4 now sits in the open top 10 it was always priced to contest — see the full field in our best open-weight AI models ranking — and at #7 in the Value Division its 50.3 points per blended dollar make it one of the cheapest credible scores in our best value AI models analysis. Both tables were computed from the wrong record for ten days; both are recomputed from the corrected one now.
What the correction does not change
- The record is still thin. Coverage stays at 40% — one recorded component of five. A corrected thin record is still a thin record; our score coverage explainer is the standing caveat.
- No other model moved. The cross-check matched every other tracked LiveBench 2.0 score exactly, so no neighbouring score was recomputed — the places Large 4 gained are places the table owed it, taken from no one’s re-measurement.
- The price story stands. The October 30 cut and the open-weights release happened as recorded; only the benchmark row attached to them was the preview checkpoint’s.
Corrections like this are why the changelog is public and why every score on this site names its source and measured date (Nov 9, 2026 for this row). The wrong number was live for ten days of rankings; the fix, its reason and its full before-and-after are now on the same record.