← All models

Mistral Large: token counter & pricing

Mistral · approximate, within ±3% of reference · pricing as of 2026-05-31.

Updated 2026-05-31 · By Clinton Patrick · Methodology

What is Mistral Large?

Mistral Large is the flagship model from French AI lab Mistral, frontier-tier capability, 128k context, available both as a hosted API and (for enterprise customers) for on-premises deployment. Strong on European-language workloads where US-trained models sometimes underperform.

How tokens are counted here

Mistral's BPE tokenizer isn't bundled in-browser yet, so estimates here come from a character-class-aware heuristic: input is bucketed into ASCII letters, digits, CJK characters, and whitespace, with a separate characters-per-token ratio per class. On typical English text that is accurate to about ±3% of the reference tokenizer, and results are marked ≈±3%. One Mistral-specific wrinkle: heavily accented French or German prose contains characters outside the plain-ASCII bucket, so counts can drift somewhat more there than on English. If you are costing a large multilingual workload, spot-check a sample against Mistral's own tokenizer (mistral-common on PyPI) before locking a budget.

When to use Mistral Large

• European / multilingual workloads. Mistral models tend to outperform US-trained competitors on French, German, Spanish, Italian.
• EU data residency requirements. Mistral's hosted API runs in EU data centers; weights available for on-prem to enterprise customers.
• You want a non-US-vendor option for procurement or compliance reasons.
• Function-calling reliability. Mistral's native tool-use is well-implemented.

When not to use it:

• Pure cost optimization on English workloads, open Llama or Gemini Flash typically win.
• Frontier reasoning benchmarks where Claude Opus / GPT-4o / Gemini Pro lead.

Pricing notes

At $2 input / $6 output per million, Mistral Large is positioned between GPT-4o and Claude Sonnet on price. Mid-tier pricing for mid-tier-to-frontier capability.

Mistral also offers cheaper models (Codestral, Mistral Small, Ministral) for specific use cases. They're not in this calculator yet, request via email if you want them added.

What Mistral Large costs in production

The workload where Mistral Large gets picked is rarely the cheapest one; it is the compliant one. Picture a European bank processing customer correspondence under data-residency constraints: 200M input tokens and 40M output tokens a month, all required to stay on EU infrastructure. On Mistral's hosted API that is 200M × $2 = $400 plus 40M × $6 = $240, about $640 a month.

GPT-4o mini would run roughly $54 ($30 + $24) on the same volume, but routes data through US infrastructure, which is the disqualifier in this scenario. Claude Sonnet 4.6 at $3/$15 would cost about $1,200 ($600 + $600) and doesn't solve residency by default either. When the compliance requirement is hard, Mistral's premium over budget US models is the cost of the requirement, not the model.

Migrating from GPT-4o

Coming from GPT-4o, the integration lift is modest: Mistral's chat API follows the familiar messages format, the model string is apiId mistral-large-latest, and native function calling is well-implemented. Two adjustments. Mistral Large is text-only, so any image inputs need rerouting to Pixtral or staying on a multimodal model. And token math shifts: GPT-4o's tiktoken counts were exact, while counts here are heuristic at ≈±3%, so add a small buffer to per-request budget caps. On price you move from $2.50/$10 to $2/$6, a 40% cut on output for comparable mid-tier capability.

Mistral Large vs the obvious alternative

GPT-4o is the model Mistral Large most often displaces: $2.50/$10 against $2/$6, making Mistral 20% cheaper on input and 40% cheaper on output, with EU hosting that GPT-4o doesn't offer. GPT-4o keeps the edge on multimodal input and ecosystem maturity. Cost both out with your real prompt sizes in the counter above; at these volumes the percentage gaps turn into concrete monthly numbers fast.

Common questions

Is Mistral Large open-source?

Mistral Large weights are not open-source under the most recent license, Mistral moved to a "research only" weights release for the Large tier. Smaller Mistral models (Codestral 22B, Mistral 7B, Mixtral 8x7B / 8x22B) remain Apache-2.0.

How does Mistral Large compare to Claude Sonnet?

Sonnet typically wins on instruction-following nuance and refusal calibration. Mistral Large wins on multilingual European tasks and on workloads where you need EU data residency. Cost is comparable.

Does Mistral support multimodal input?

Mistral has separate vision models (Pixtral). Mistral Large itself is text-only as of this writing, for image input, use Pixtral or another multimodal model.

Compare Mistral Large to other models

• GPT-4.1 (OpenAI, $2.00/$8.00)
• o3 (OpenAI, $2.00/$8.00)
• Gemini 3.1 Pro (Google, $2.00/$12.00)