Read this before you compare the benchmark numbers
Mistral published Terminal-Bench 4: 28.3%. Reflection published Terminal Bench v2.1: 80.1. These are different versions of the same benchmark and the scores are not comparable. Anyone showing you those two figures side by side is either confused or selling something.
What each one actually is
| |
Mistral Large 4 |
Reflection Beam |
| Total parameters | 1T (launch post) / 1.05T (spec table) | 501B |
| Active per token | 49B (launch post) / 52B (spec table) | 23B |
| Context window | 1M tokens | 1M tokens (effective) |
| Input types | Text and images (1.6B vision encoder) | Text only |
| Licence | Not announced | Apache 2.0 (planned) |
| Weights | Due around 27 October | Later in October |
| Access today | Public API preview since 6 Oct | Waitlist, OpenAI-compatible beta |
| Published price | $0.68 in / $2.09 out per 1M | None |
Mistral Large 4
In public preview on the Mistral API since 6 October 2026 under the model name mistral-large-4. Mistral's launch post describes it as a 1 trillion parameter mixture-of-experts model with 49 billion active per token; the specification table published alongside it says 1.05 trillion and 52 billion. Both figures came from Mistral on the same day, so treat the exact numbers as unsettled.
Pricing, per million tokens, at the launch discount: $0.68 input, $0.07 cached input, $2.09 output. Those are 50% off list prices of $1.36, $0.14 and $4.18. The discount covers the first two weeks of preview, so it has either expired or is about to by the time you read this — check before you budget.
Where it is genuinely strong: security work. 82% on a reproduce-and-patch test, which Mistral says is the highest of any model it benchmarked, and 93% on Cybench. It also resists 93.3% of attacks on Lakera's B3 benchmark. If your workload is vulnerability analysis or patching, that is a specific, measured advantage rather than a general claim.
Where it is honest about falling short: Mistral published a Surge AI blind coding rating of 3.74 out of 5 against 4.22 for Claude Opus 5. Publishing a number where you lose is a reasonable signal of good faith about the rest of the card.
Reflection Beam
A 501 billion parameter sparse mixture-of-experts model with 23 billion active per token and a 1M-token effective context. Text only. Announced 5 October 2026.
The architectural argument is the point. Beam is the smallest of its peer group by total parameters, and Reflection claims it needs three to four times less inference compute than comparable open models on reasoning benchmarks. That is the whole pitch: not the best scores, the best scores per unit of compute.
Where it leads: 80.9 on SWE-bench Verified against 70.7 for Nemotron 3 Ultra.
Where it does not: on Terminal Bench v2.1, Beam scores 80.1, behind GLM-5.2 at 81.0, Kimi K3 at 88.3 and DeepSeek V4.1 Flash at 90.6. Reflection is not claiming the capability crown and neither should anyone quoting it.
The licence difference is the biggest practical gap. Beam promises Apache 2.0, which is about as permissive as licences get — commercial use, modification, redistribution, no copyleft. Mistral has not announced a licence at all. For anyone planning to ship a product on top of self-hosted weights, that is not a detail to resolve later. It is the first question.
Decision framework
- If you need image input → Mistral Large 4. Beam is text only, and no licence or efficiency advantage compensates for a modality it does not have.
- If you need a licence you can build a business on → Beam, on the Apache 2.0 promise. Revisit if Mistral announces something permissive.
- If you need to test something this week → Mistral Large 4. It has a public API; Beam has a waitlist.
- If your constraint is inference cost on your own hardware → Beam, on the 23B active parameters. Less than half Mistral's active count is a real serving-cost difference, not a marketing one.
- If you do security or vulnerability work → Mistral Large 4. The 82% reproduce-and-patch and 93% Cybench results are the most specific capability claim either model makes.
- If you want the strongest open model available regardless of these two → neither, today. Kimi K3 and DeepSeek V4.1 Flash both score higher on the one benchmark where Beam published a comparable figure.
What to actually do this month
Nothing irreversible. Neither model has shipped weights, Mistral has not announced a licence, and Beam has not announced a price. Every number above is a vendor-reported figure on a pre-release model.
The sensible move is to put 27 October in the diary for Mistral weights, join the Reflection waitlist because it costs nothing, and run both against your own workload when they land. Vendor benchmark cards have never predicted which model works on a specific codebase, and two models measured on different benchmark versions predict it even less.
FAQ
Can I compare Mistral Large 4 and Beam on Terminal-Bench?
No. Mistral reported Terminal-Bench 4 and Reflection reported Terminal Bench v2.1. Different versions produce different score distributions, so the figures are not comparable.
Which one is cheaper?
Unanswerable today. Mistral publishes $0.68 per million input tokens in preview; Reflection publishes no price at all. On self-hosted serving cost, Beam's 23B active parameters against Mistral's roughly 50B favours Beam, but that is an inference-economics argument rather than a price.
When are the weights out?
Mistral Large 4 around 27 October 2026. Beam is stated as later in October 2026 with no firm date.
Is Mistral Large 4 open source?
Not yet, and not necessarily. Open weights are promised, but Mistral has not announced which licence those weights will carry. Open weights and open source are not the same thing.
How big are these compared with Kimi K3?
Much smaller. Kimi K3 is 2.8 trillion parameters with 104 billion active. Beam is 501 billion with 23 billion active, and Mistral Large 4 is around 1 trillion with roughly 50 billion active.