Mistral Large 4 hallucination rate, measured
Does Mistral Large 4 make up fewer citations than Mistral Large 3?
Mistral Large 4 made up 42.3 percent of the DOIs it offered across three runs, and Mistral Large 3 made up 36.3 percent on the same thirty claims (p = 0.27). On that count, chance explains the gap. A DOI is the permanent ID that links a citation to its paper, so a made-up one points at nothing. Fifteen of the thirty claims are ones we made up, with no study behind them. Large 3 cited sources for all of them in every run, and Large 4 declined 44 of its 45 answers. Counting real IDs put on the wrong paper too, Large 4 got 67.7 percent wrong against 81.9 for Large 3 (p = 0.0022). Large 4's 42.3 percent is the highest made-up rate in this test, and all three runs were on one day.
Read this if you draft with Mistral Large 4 and publish the sources it gives you.
Level on made-up IDs, far better at declining
Mistral Large 4 made up 37.2 percent of its DOIs, then 42.9, then 46.7. Mistral Large 3 ran the same thirty claims on the same day and made up 35.6 percent, then 36.7, then 36.7.
Pooled, that is 42.3 percent against 36.3, and a Fisher exact test gives p = 0.27. Large 4 was higher in every run, by a gap chance explains, so this is a null result. We also count real IDs attached to the wrong paper. Large 4 had 33 and Large 3 had 123. Counting both kinds of error, Large 4 got 67.7 percent wrong and Large 3 got 81.9 (p = 0.0022). That gap clears the usual bar of 0.05, and it comes from the made-up claims. Every DOI Large 3 gave for those was wrong. On the real claims alone, the two got about the same share wrong.
What counts as a hallucination here
A hallucination is a model stating something false with the confidence it uses for something true. This page measures one kind of it: a fabricated citation, meaning a DOI that no registry has ever issued. We measure that kind because it has an objective answer. Whether a claim is true can be argued. Whether a DOI resolves cannot. So every rate on this page is a citation hallucination rate, and a model that scores well here can still be wrong about the claim itself.
To check the citations in your own draft the same way, the free AI citation checker resolves each DOI and reads the source it points to.
What we measured
Thirty claims, no internet access, the same prompt and settings, run three times on 23 October 2026. Mistral Large 4 is the subject. Mistral Large 3 is the model before it in the same line, and we measured it here on the same day. Grok 4.3 is a control that has run in every test since the second one. It sits in another line and is never compared with the pair. It scored 8.7 percent here, inside its usual band. Both Mistral models had room for 16,000 tokens, or word pieces, per answer, and the note below the table says why.
Made-up DOIs from Mistral Large 4 and the models beside it. Three runs of the same 30 claims on October 23, 2026.
| Model | Role | Run 1One run puts all 30 claims to the model once. The prompt and the settings stay the same in each run. | Run 2 | Run 3 | Run spreadThe lowest run and the highest run. A wide spread tells you that one run alone can give a different answer. | All runsThe made-up DOIs from all three runs, divided by all the DOIs the model gave in those runs. That is about 135 DOIs for each model. | Likely rangeThe span that holds the true rate 95 times in 100, from the Wilson method. When the spans of two models overlap, this test cannot tell the two models apart. |
|---|---|---|---|---|---|---|---|
| Mistral Large 4 | subject | 37.2% | 42.9% | 46.7% | 37.2–46.7% | 42.3% | 34.2–50.9% |
| Mistral Large 3 | predecessor | 35.6% | 36.7% | 36.7% | 35.6–36.7% | 36.3% | 30.8–42.2% |
| Grok 4.3 | control | 9.5% | 9.5% | 7.1% | 7.1–9.5% | 8.7% | 4.9–15.0% |
Each row shows one model. Run 1, Run 2 and Run 3 show the share of its DOIs that no registry knows. Run spread shows the lowest and the highest run. All runs adds the three runs together. Likely range shows where the true rate most likely sits, 95 times in 100.
How to read this
- Run
- One run puts all 30 claims to the model once. The prompt and the settings stay the same in each run.
- Run spread
- The lowest run and the highest run. A wide spread tells you that one run alone can give a different answer.
- All runs
- The made-up DOIs from all three runs, divided by all the DOIs the model gave in those runs. That is about 135 DOIs for each model.
- Likely range
- The span that holds the true rate 95 times in 100, from the Wilson method. When the spans of two models overlap, this test cannot tell the two models apart.
- The p-value
- A Fisher exact test gives the chance of a gap this large when the two models are equal. We treat a p-value above 0.05 as noise.
- Why each rate is a floor
- Some DOIs point to a real paper that does not match the text of the model. We do not count these DOIs as made up. So the true rate is equal to the rate here or higher.
The rate divides made-up DOIs by DOIs offered. A DOI counts as made up only when no registry has issued it. Pooled over three runs, Mistral Large 4 made up 55 of 130, with a 95 percent interval of 34.2 to 50.9 percent. Mistral Large 3 made up 98 of 270, from 30.8 to 42.2 percent, and Grok 4.3 made up 11 of 126, from 4.9 to 15.0 percent. Fisher exact on the pair gives p = 0.27, and run by run it gives 0.85, 0.57 and 0.27. Another 160 DOIs from the pair led to a real paper whose title did not match what the model wrote. We graded each against the paper's registered first author and year, and 156 were miscitations: 33 from Large 4 and 123 from Large 3. We did not read them by hand. Counting both kinds of error, Large 4 got 88 of 130 wrong, from 59.2 to 75.1 percent, and Large 3 got 221 of 270, from 76.8 to 86.0 percent, p = 0.0022. An answer that runs out of room comes back empty. In the first part of the test, at 8,000 tokens, Large 4 ran out on 4 answers, so we ran the pair again at 16,000. Large 4 still ran out twice, both times on the claim about language models agreeing with users, in runs 1 and 2. Neither empty answer holds a DOI, so neither enters the rate. Large 3 never ran out. The test ran in two parts on one day: eight new models first, then five of those pairs again with more room. The second part cost $3.88, and both parts together cost $11.65.
What this data cannot tell you
- It cannot rank vendors. Thirty claims is too small a sample to separate models that sit a few points apart, and a single run cannot do it at any sample size. Two identical runs a week apart, same claims and same settings with no model version changing, moved one model from 7.0 to 13.3 percent and another from 4.7 to 0.0. Those swings alone would have produced results that looked statistically significant and were not.
- It only compares like with like. Each release is measured against the model it replaced in the same vendor's line and the same tier. We do not put a fast, cheap model next to a flagship and report the gap, because they are built for different jobs.
- It measures ungrounded recall, not the product you use. No web search, no retrieval, no tools. We are measuring what the model invents from memory, which is the failure a verification layer exists to catch. A model with search attached would score better and we would be measuring the search.
- A result that shows nothing still gets published. Most of these comparisons come back inconclusive, and we say so.
Large 3 cited sources for every claim we made up
The thirty claims split into two groups. Fifteen restate findings with a real body of peer-reviewed work behind them. Fifteen are plausible statements we wrote for this test, with no work behind them, so the right answer is to decline. Mistral Large 4 declined 44 of its 45 answers to those fifteen. Mistral Large 3 declined none, and gave 135 DOIs for claims that have no source.
Of those 135, 62 do not exist, among them 10.1111/ecge.12345 and 10.1289/EHP4567. The other 73 point at real papers that cannot back a claim we made up. On the real claims, both models made up IDs often.
No such index exists. In run 1, Mistral Large 4 cited a 2019 paper by Brennan on retractions in citation networks, as 10.1007/s11192-019-03205-5. No registry has issued that ID. It was Large 4's only source for a made-up claim in 45 tries.
Luby's 2005 handwashing trial in The Lancet is real, and its ID is 10.1016/S0140-6736(05)66912-7. Large 3 gave that ID in all three runs. Large 4 gave a different made-up ID in each run: 10.1016/S0140-6736(05)66921-X, then 66782-8, then 67687-4.
Lim's 2010 meta-analysis of sleep loss is real, and its ID is 10.1037/a0018883. In all three runs, Large 4 gave it 10.1093/sleep/33.5.574, an ID no registry has issued.
Five labs, one made-up ID
Card and Krueger's 1994 study of fast-food jobs is real. On 23 October, models from five labs gave it the same ID, 10.1257/aer.84.4.772, and no registry has issued it. The labs were Mistral, Alibaba's Qwen, Z.ai, MiniMax and Meta. Mistral Large 3 gave it in every run, and Large 4 in two. Five labs agreeing on one string means they learned the same wrong pattern. It says nothing about whether the ID exists, and a registry lookup is the only check that settles it.
Fabrication is one failure. Sycophancy is the other.
This benchmark measures fabrication: the model inventing a source it does not have. It does not measure sycophancy, the model bending its answer toward the opinion you showed it. A sycophantic model asked whether its own citation is real tends to say yes, which is why a model cannot be the check on its own draft. We test for both, with different instruments. Citations go to the registries that issue DOIs, the same check the hallucination checker runs on a pasted draft. Drafts go through the sycophancy detector, which scores how far a piece over-claims. If the word is new to you, what AI sycophancy is covers the mechanism.
Mistral Large 4 got about two citations in three wrong. Large 3 went further and gave full source lists for claims that have no source at all. TrueStandard does the checking. Paste a draft, and four models from different labs check each claim and citation against each other and against the registries that issue DOIs.
How current this is
Mistral Large 4 became available through the API on 6 October 2026. We measured it and put this page up on 23 October, 17 days later. That misses the 48-hour target we set ourselves by two weeks. We measured Mistral Large 3 again on the same day, so the comparison is fresh even though the page is late.
How we ran it
Every model gets the same thirty claims, the same prompt, the same settings, and no access to the internet. Fifteen claims are well documented and have real literature behind them. Fifteen are specific, plausible-sounding statements we constructed and know of no support for, which is where the pressure to invent a source comes from. We ask for up to three peer-reviewed sources with DOIs, then check every DOI against Crossref and, when Crossref does not hold it, DataCite. A DOI that exists in neither is counted as fabricated. A DOI that exists but resolves to a paper whose title does not match what the model described goes to a manual-review queue and is not counted on either side, so every rate here is a floor.
Three caveats. First, all three runs happened on one day. A model's rate can move between days, which is why we measure the old model again next to the new one. A second day for this pair is open work. Second, this pair had more room per answer than the control: 16,000 tokens against 3,000 for Grok 4.3. An answer that runs out of room is an error, never a decline, so we gave the pair room to finish. Large 4 still looped on one claim. It ran that claim to 7,999 tokens at an 8,000 cap and to 16,000 at a 16,000 cap, and we did not try a cap high enough to end it. Third, all the models ran at low reasoning effort with no web access. This test did not try a higher effort, web search or a sycophancy check, and we did not read the miscitations by hand. Our scorer checks DOIs against Crossref and DataCite. We then checked all 260 dead IDs from the second part of the test at doi.org, which covers every agency. None of them resolve.
The claim set, the prompt, the model identifiers, every raw response and every DOI verdict are committed beside the code that built this page, and a public mirror under CC BY 4.0 is being set up. Until it is live, ask and we will send the run files.
Common questions
Does Mistral Large 4 hallucinate citations?
Often. With no web access, it made up 55 of the 130 DOIs it offered across three runs on 23 October, or 42.3 percent. Counting real IDs on the wrong paper too, it got 67.7 percent of its citations wrong.
Is Mistral Large 4 better than Mistral Large 3 at citations?
On made-up IDs, no: 42.3 percent against 36.3 (p = 0.27), a gap chance explains. On wrong citations overall, yes: 67.7 percent against 81.9 (p = 0.0022). That gain comes from Large 4 declining claims that have no source. Both results come from three runs on one day.
What did Mistral Large 3 do with the claims we made up?
It cited sources for every one, in every run, 135 DOIs in all. 62 of them do not exist, and the other 73 point at real papers that cannot back a claim we made up. Large 4 declined 44 of its 45 answers to the same claims.
Can I trust Mistral Large 4's citations without checking them?
No. About two in five of its IDs did not exist, and about two in three of its citations were wrong one way or the other. Check each DOI against a registry before you cite it.
What were the two empty answers?
On the claim about language models agreeing with users, Large 4 kept reasoning until it ran out of room, in runs 1 and 2. It did that at 8,000 tokens and again at 16,000. An empty answer holds no DOI, so it does not change the rate.
Why does another benchmark give Mistral Large 4 a different hallucination rate?
Because it measures something else. Most hallucination scores count wrong answers to questions, and some run the model at maximum effort. Ours counts citation IDs that no registry has issued, at low effort. A DOI either resolves or it does not, so anyone can check our number without trusting a model to grade it.
Can I reproduce this?
Yes. We keep every input and output: the claims, the prompt, the model IDs, all three runs of raw answers, every DOI verdict and the graded miscitation file. The part of the test behind this page cost $3.88, and the whole test cost $11.65. Ask and we will send the run files.
Is this the Mistral Large 4 hallucination rate?
It is its citation hallucination rate: the share of DOIs it gave that no registry has issued, with no web access. A general accuracy score would need a different test. We measure citations because the answer is a fact anyone can check.
Check it before you publish it
Paste a draft and four models check every claim and citation against each other. You see where they disagree, which is where the errors are.
The claim checker is free. Pro is $20 a month for 250 drafts, Max $40 for 500, and Ultra $200 for 3,000.