Mercury 3 vs Mistral Large 4: Two Unfinished Models and the Promises Attached to Each

Neither of these models is finished, and the ways in which they are unfinished are opposite. Mercury 3 was released by Inception Labs on October 19, 2026 as a shipping product at $0.20 per million input tokens and $0.70 per million output, and it has no independent evaluation of any kind — no Artificial Analysis score, no third-party harness, nothing. Mistral Large 4 was announced by Mistral on October 6, 2026 as a public preview at $1.36 and $4.18 per million, and it has an independent score of 38 on the Artificial Analysis Intelligence Index and a measured $1.13 per index task, but its weights — which Mistral promised would drop "by the end of the month" — had not been published on Hugging Face when this was written, and the preview's own reinforcement-learning run was still going at announcement. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

So the comparison is between a model whose numbers are all vendor-reported and a model whose numbers are not all final. They also occupy different product categories: Mercury 3 is a text-only diffusion language model with a million-token context and a throughput claim of 1,068 tokens per second, while Mistral Large 4 is a natively multimodal 1-trillion-parameter mixture-of-experts with 52 billion active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs in French datacenters and positioned explicitly around European AI sovereignty. If your workload is text and your constraint is cost, that difference is decisive before any benchmark is consulted. If it is not, the two models are barely competing.

Mercury 3, and the specific thing that is missing

Inception's model is the third generation of its parallel-decoding bet. Instead of one token per forward pass, Mercury denoises a block of positions and commits them together, which is where 1,000-plus tokens per second comes from. Mercury 3 is, by Inception's account, more than 40 times the total parameter count of Mercury 1, with a context window grown from 260,000 to 1,000,000 tokens and a 131,072-token output ceiling, and it runs at 1,068 tokens per second on unspecified NVIDIA hardware.

What is missing is any measurement Inception did not perform. The announcement's one structured quality result is a live-site agent test: 50 Mind2Web navigation tasks against real websites inside Inception's own production harness, where Mercury 3 completed 50% — 25 tasks — which Inception says is "significantly more than GPT-6.1-Sol" at the same token volume and step count. That is more interesting than a benchmark table because the sites were live, and less trustworthy for the same reason a vendor demo always is: Inception chose the harness, the targets and the scoring. There is no Artificial Analysis page for Mercury 3 as of October 20, 2026, so there is no neutral throughput figure, no neutral cost per task, and no neutral recall curve for that million-token window.

The consequence is concrete rather than philosophical. A million-token context with an unmeasured recall curve is a capacity claim: the tokens fit, and nobody has established how much of them the model actually uses. Mercury 2.5 scored 12 on the Artificial Analysis Intelligence Index, which tells you where the previous generation sat and nothing about this one.

Mistral Large 4, and the specific thing that is unfinished

Mistral's model is a different proposition entirely and the announcement is unusually specific about it. It is a 1-trillion-parameter natively multimodal model with 52 billion active parameters — sparse mixture-of-experts, so most of the trillion is not doing work on any given token. It handles documents, charts, natural images, technical drawings and gigapixel satellite imagery, and it was trained on data spanning more than 160 languages including every official EU language.

• Coding — 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4, for a combined Coding Agent Index of 49.8%. All Mistral-reported.

• Agents and cyber — 59.9% on AutomationBench, 1,393 Elo on AA-Briefcase, 82% on a real-vulnerability reproduction-and-patch test that Mistral says leads all models, 93% on Cybench, and 93.3% attack resistance on Lakera's B3.

• Vision — 42% on Dense 200 visual grounding, narrowly ahead of GPT-6 Astra at 41%.

• Human evaluation — 3.74 out of 5 on Surge AI's blind coding assessment, second of five models behind Claude Opus 5 at 4.22. A third party ran the study; Mistral commissioned and published it.

• Independent aggregate — 38 on the Artificial Analysis Intelligence Index, ranked 66th of 217, at $1.13 per index task, 67.5 output tokens per second and a 1.65-second time to first token.

• Status — preview, weights promised by end of October 2026, RL run still in progress at announcement.

The classification is worth being careful about. Mistral's own documentation lists the v26.10 generation as "Open" and open-weight; the Artificial Analysis page, read on October 20, still says "proprietary" and "the model weights are not publicly available". Both are correct at different times: the licence is the intent, the weights are not yet on disk. Anyone building a deployment on the assumption that Mistral Large 4 will be downloadable in November should note that the promise has a date and not yet a repository.

The two rate cards, and the number that makes the gap smaller than it looks

On paper Mercury 3 is dramatically cheaper: $0.70 per million output against $4.18, and $0.20 input against $1.36. That is a 6-fold gap at list, or 12-fold during Mercury's launch window at $0.10 and $0.35, with Mistral's own promotional pricing at $0.68 and $2.09 narrowing it to roughly 6-fold.

The per-answer number is where it compresses further, and it is the one number in this article that was measured by someone other than the vendor selling the model. Artificial Analysis puts Mistral Large 4 Preview at $1.13 per Intelligence Index task. Mercury 3 has no equivalent, so the only way to construct one is to assume a token volume — and the earlier Mercury comparisons showed why that assumption matters: a model that spends 31,000 output tokens where another spends 193,000 can be 20 times more expensive per token and still cheaper per answer. At $0.70 per million, Mercury 3 would need around 1.6 million output tokens to cost $1.13 for a single task. It will not spend that. On any realistic volume it lands far below $1.13, which is the honest core of its case: the per-token advantage survives the per-answer correction here in a way it did not against GPT-6 Sol.

• Context — Mercury 3: 1,000,000 tokens per Inception's documentation. Mistral Large 4: 1,000,000 per Mistral's model listing, against 524,288 on Artificial Analysis's tested configuration.

• Multimodality — Mercury 3: text in, text out, no images at any price. Mistral Large 4: documents, charts, photographs, technical drawings, satellite imagery.

• Speed — Mercury 3: 1,068 tokens per second, vendor-measured. Mistral Large 4: 67.5 tokens per second per Artificial Analysis, with a 1.65-second time to first token that is genuinely fast for a model this size.

• Weights — Mercury 3: none published, and Inception's Hugging Face organisation holds no model repositories. Mistral Large 4: promised by end of October, not yet published.

• Governance — Mercury 3: no published statement on training data or EU deployment. Mistral Large 4: trained in European datacenters, Europe-operated deployment, all official EU languages.

The long-context line is the one to read twice. Mistral's documentation lists a million tokens; Artificial Analysis tested at 524,288. That is not a contradiction — providers test at whatever power of two their harness supports — but it is a reminder that a context figure is a maximum, not a tested capability, and Mistral Large 4's long-context recall has not been published at the full window either.

Where the European argument actually bites

The sovereignty pitch is not decoration on this model; it is the reason several buyers will choose it over something cheaper and better. Trained on Blackwell GPUs in French datacenters, operated from Europe, multilingual across all official EU languages — for a public-sector buyer or a regulated European enterprise, that combination can outweigh twelve points of intelligence index, because the alternative may not be procurable at all.

Mercury 3 offers none of it. Inception has published nothing about where the model runs, what data it was trained on, or what retention terms apply, beyond noting that enterprise customers can buy dedicated capacity, autoscaling, compliance controls and configurable retention. Those are the right things to sell to an enterprise and they are not the same as a published data-processing position.

There is one more wrinkle in the European story that is worth flagging rather than resolving: the open-weights promise. If Mistral publishes the weights at the end of October, a European buyer gets something no American lab currently offers at this scale — a frontier-adjacent multimodal model that can run inside their own perimeter, on their own hardware, with no vendor in the request path. That is a materially different product from a preview API, and it is the reason the missing repository matters more than the missing benchmark points.

Both of these models are off our catalogue, which changes the recommendation

Mercury 3 is not on the OrcaRouter catalogue, and neither is Mistral Large 4 — our model API returns nothing for either, across every slug variant we tried. That is unusual enough to say plainly, because it means this comparison cannot end with "put both on one key". Anyone calling either model is going to the vendor directly, or to one of the several third-party platforms that host it.

What a routing layer still does in that situation is the part that does not require hosting the model: it holds the models we do route behind a single OpenAI-compatible endpoint, and it lets you express a rule about which traffic goes where. If Mistral's weights land at the end of October, the interesting architecture is not "use Mistral Large 4 for everything" — it is a self-hosted Mistral path for the documents that cannot leave your perimeter, a routed frontier model for the work that can, and a rule that decides between them. The same rule handles Mercury 3 if Inception ships it somewhere we can reach, and it handles the case where either vendor's promotional pricing ends, which for Mercury 3 has no announced date.

Who should pick which, and what would change the answer

Pick Mercury 3 if the work is text, the volume is high, and the answers are cheap to verify. The rate card is six to twelve times cheaper than Mistral Large 4 depending on which prices are live, the context window is the same headline size, and the absence of a neutral score is a risk you can measure in your own error rate within a day. What you give up is every form of input that is not text, and any ability to audit the model's behaviour before deploying it.

Pick Mistral Large 4 if the input includes images, drawings or scans; if procurement requires European training and operation; if a 1.65-second time to first token matters more than raw throughput; or if the open-weights promise is the actual reason you are interested, in which case wait for the repository rather than the announcement. What you give up is a factor of six on price and any claim to being the fastest thing available.

Two dates would settle most of the open questions in this piece. The first is the end of October, when Mistral either publishes the weights or does not. The second is whenever Artificial Analysis runs Mercury 3, which would replace every vendor number in the first section with a measured one. Neither date has been announced. Until they arrive, the defensible summary is that Mercury 3 is the cheaper text engine and Mistral Large 4 is the more complete product, and that both are being sold on claims that the buyer currently has to take on faith.