Price an open-weight MoE model before you commit

You are deciding whether to route production traffic to Mistral Large 4 through an API or to self-host the open weights. The trade-off is arithmetic: total parameters set memory, active parameters set compute, and tokens per task set spend. Mistral AI announced open weights for Mistral Large 4. Mistral Large 4 (High) is priced at $0.70 per 1M input tokens and $2.10 per 1M output tokens. API cost equals (input tokens / 1,000,000) × $0.70 + (output tokens / 1,000,000) × $2.10, before retries, truncation, and review.

Active parameters set the serving bill

Mistral Large 4 has 1.1 trillion total parameters and 52 billion active parameters; the released checkpoint is a 1.05-trillion-parameter system with 1 million tokens of context. MoE (mixture-of-experts) inference pays for the active slice on each token, not the full weight file. Zhipu AI's GLM 5.2 does not activate its full 744B parameter set for every token during inference. That distinction changes GPU sizing, latency, and the price of a task.

Five lines separate a candidate from a cost model

Do the check in this order: model spec, workload, price, throughput, license. A one-page memo with five rows is enough. Each row needs a number, a source, and a pass/fail. If any row fails, stop. The remaining rows are a cost model, not a fix.

• Count active parameters. List active, total, and context on the model card. A serving plan that assumes dense total-parameter compute overstates latency and understates token throughput.
• Measure tokens per task on your workload. Take a median and p95 from your eval harness, not a vendor benchmark. Mistral Large 4 scores 48% on the Artificial Analysis Coding Agent Index, matching GLM-5.3, and used 8.5 million tokens per Coding Agent Index task, about 47% fewer than GLM-5.3's 15.9 million tokens per task. That efficiency changes the bill.
• Price input and output separately. Start from the published API rates for your chosen endpoint, then multiply by your measured tokens, retries, truncation, and human review.
• Check sustained throughput, not peak demo speed. Run a load test at your target concurrency. Mistral Large 4 (High) has an output speed of 161 tokens per second. A tighter p95 latency target may require more self-hosted GPUs than the API price suggests.
• Confirm license and deployment fit. On October 28, 2026, Mistral Large 4's Hugging Face countdown ETA changed from October 31 to October 29 while the weights and LICENSE file were still absent. The Mistral Ecosystem Open-Weight License calls for a distinct commercial license before commercial use if the licensee or an affiliate runs a model-as-a-service business. If that clause blocks your product, the model is not a candidate.

The published benchmark task cost does not capture the full cost of running the model in-house, adapting it, supervising operations, or reviewing outputs. The released Mistral Large 4 high-reasoning checkpoint costs $2.06 per Intelligence Index task, a 63% reduction from the preview's $5.54.

API cost is tokens times price. Self-hosting cost is amortized GPUs plus storage plus networking plus monitoring plus human review. The API price is no longer on the invoice, but the serving bill remains on your infrastructure line. Put both paths in the same table before you choose.

Benchmark scores are entry fees, license fit is a gate

On Terminal-Bench 4.0, a terminal coding-agent benchmark, Mistral Large 4's score rose from 28% to 41%. A score gain does not reduce the bill; higher token counts raise the bill. Put the token count next to the benchmark score, then compare the two models on the same task.

When your product exposes model outputs to customers, the model-as-a-service clause is the first line to price. For internal-only use, it may be a non-issue. The same question applies to derivatives, fine-tunes, and the serving layer you build around the weights. A clean pass is a short memo: the model, the workload, the token counts, the price per task, the throughput result, and the license conclusion. Rerun the five lines when the model or workload changes.