Mistral API Pricing vs Self-Hosted LLMs: Cost and Privacy in 2026

Mistral Large 3 costs $0.50 per million input tokens and $1.50 per million output tokens on La Plateforme, a 75% cut from what Large 2 charged for the same job. That single fact changes the shape of this comparison versus the rest of this series. Our GPT-6 and Claude Opus 4.8 breakdowns found real crossover points where a self-hosted GPU cluster beats the API on pure cost, somewhere between 16M and 108M tokens a day depending on the model and hardware tier. Run the same math against Mistral's own API and the crossover barely exists: Mistral's rates are already low enough that a self-hosted cluster running Large 3 or Small 4 usually costs more per token, not less. The case for self-hosting Mistral has to rest on something other than "the API is expensive."

Mistral API Pricing in 2026: Large 3, Medium 3.5, Small 4, and the Full Model Line

Mistral's La Plateforme pricing, fetched from mistral.ai/pricing/api, covers eight models across three tiers: flagship, mid-size, and small/specialized. Every price below is per million tokens, listed as input/output.

Large 3, Medium 3.5, and Small 4 Rates (the Models with Self-Hosted Equivalents)

Medium 3.5 sits between Large 3 and Small 4 on price, and it's also the odd one out: unlike Large 3 and Small 4, Mistral hasn't released Medium 3.5's weights, so there's no self-hosting path for it at all right now. If your workload needs Medium-tier quality and you want to self-host, Large 3 is the nearest open option, at a lower per-token API rate than Medium 3.5 to boot.

Ministral, Codestral, Devstral, and Specialized Model Pricing

The smaller and task-specific models round out the line at correspondingly lower rates:

Ministral 3B, 8B, and 14B are dense models with a dedicated self-hosting guide, which makes them a straightforward lighter-weight alternative if the MoE overhead of Small 4 isn't worth it for your workload. Devstral 2's rate places it above Codestral on both input and output, and its output rate ($2.00/M) actually runs higher than Large 3's ($1.50/M) despite a lower input rate, reflecting its narrower agentic-coding focus rather than a straightforward discount. Magistral Medium and Magistral Small are Mistral's Premier-tier reasoning models, listed separately under a "Reasoning" category on the pricing page rather than alongside the general-purpose line. If you need Mistral's always-on chain-of-thought model rather than Small 4's configurable reasoning, Magistral is also open-weight and self-hostable, at the same API rate for the Small variant as Large 3 and a higher one for Medium.

Mistral Large 2 to Large 3: The 75% Price Cut Nobody Updated Their Comparison Post For

Mistral Large 2 was priced at $2.00 per million input tokens and $6.00 per million output tokens, confirmed via OpenRouter's provider pricing for Mistral Large. Large 3 launched in December 2025 and now runs $0.50/$1.50, a 75% cut on both input and output.

That cut is recent enough that it's still catching third-party trackers flat-footed. A Mistral pricing roundup on tokenmix.ai, last updated April 29, 2026, still lists "Large 3" at the old $2.00/$6.00 rate, four months after the actual price dropped. If you're pulling a comparison figure from a blog post or aggregator rather than Mistral's own pricing page, check the date. A four-month-old number here is off by 4x.

What You're Actually Paying For: Batch Discounts and Prompt Caching

Headline per-token rates are the ceiling, not what most production traffic actually pays. Two mechanisms bring the real bill down substantially, and neither shows up in the table above.

Batch API: Flat 50% Off for Async Workloads

Mistral's Batch API applies a flat 50% discount to both input and output tokens for asynchronous jobs. The pricing page's FAQ puts it plainly: "Batch processing gets a 50% discount" (mistral.ai/pricing). For offline classification, bulk document processing, or nightly enrichment jobs where a synchronous response isn't required, this is a straightforward win with no architectural cost. Large 3 batch pricing works out to $0.25/M input and $0.75/M output; Small 4 to $0.075/M input and $0.30/M output.

Prompt Caching: 90% Off Cached Input, in 64-Token Blocks, Opt-In via prompt_cache_key

Mistral bills cached input tokens at 10% of the standard input rate, a 90% discount, when a request shares an identical prompt prefix with a prior call and carries a stable prompt_cache_key (Puter developer docs). On Large 3, a cache hit drops input from $0.50/M to $0.05/M. On Small 4, it drops from $0.15/M to $0.015/M.

Two mechanics matter for whether you actually see that discount: caching operates in 64-token blocks, so prompts under 64 tokens never hit the cache regardless of repetition, and a hit isn't guaranteed even when the prefix matches. You keep the shared prefix identical across calls and pass a stable prompt_cache_key (a conversation, session, or workflow ID), but you still need to verify cache hit rates through your usage dashboard rather than assuming the discount applies.

Blended Cost per Million Tokens at 80/20 and 60/40 Input/Output Splits

Most production traffic sends more input than output. At an 80/20 split, uncached:

Blended Large 3 rate = (0.8 x $0.50) + (0.2 x $1.50) = $0.40 + $0.30 = $0.70/M tokens
Blended Small 4 rate = (0.8 x $0.15) + (0.2 x $0.60) = $0.12 + $0.12 = $0.24/M tokens

At a more output-heavy 60/40 split:

Blended Large 3 rate = (0.6 x $0.50) + (0.4 x $1.50) = $0.30 + $0.60 = $0.90/M tokens
Blended Small 4 rate = (0.6 x $0.15) + (0.4 x $0.60) = $0.09 + $0.24 = $0.33/M tokens

Stack batch and caching together and Large 3's effective input cost can fall as low as $0.025/M (10% cache-hit rate on top of the 50% batch discount), though output tokens only ever get the flat batch discount, so your real blended savings land below what that input-only figure implies.

Self-Hosting Mistral Large 3 or Small 4: GPU Cost at Volume

The short version: Mistral's own API is priced low enough that a self-hosted GPU cluster running Large 3 or Small 4 usually costs more per token, not less, at the on-demand and spot rates available today. The crossover volume where a single cluster's fixed cost undercuts the API works out mathematically to somewhere between roughly 350M and 1.6B tokens/day depending on model and hardware tier, well beyond what a single cluster can physically process before you'd need to add a second one at the same per-token cost. Self-hosting still makes sense for data control, not for cost.

Cluster Sizing Recap: 675B MoE vs 119B MoE, and the GPUs Each Actually Needs

Both open-weight Mistral models are Mixture of Experts architectures, which means VRAM sizing follows total parameters, not active parameters. The router can send any token to any expert, so every expert's weights have to sit resident in GPU memory regardless of how few activate on a given forward pass.

Full setup detail, including vLLM expert-parallel flags and quantization tradeoffs, is in the Mistral Large 3 deployment guide and the Mistral Small 4 deployment guide. The takeaway for this comparison: neither model runs on a single consumer or single-datacenter GPU. The cost comparison below is always against a multi-GPU cluster, never a single card.

Live GPU Cost per Token: Cluster $/Day vs Blended API $/M Tokens

Live Spheron GPU pricing per GPU, fetched 23 Jul 2026:

Multiplying by cluster size gives the fixed hourly cost of each configuration:

Large 3's own deployment guide clocks published throughput at roughly 2,500 tokens/sec on 8x H200 FP8 and roughly 3,000 tokens/sec on 4x B200 FP8. At those figures, self-hosted cost per million tokens comes out to:

Compare that to Large 3's own blended API rate of $0.70/M uncached, or $0.35/M on the Batch API. Even the cheapest self-hosted configuration in this table, 8x H200 on spot at $1.62/M, still costs more than double Mistral's uncached API rate, and more than 4.6x the batch rate. That gap doesn't close with volume, since throughput per dollar stays roughly flat once a cluster is at capacity; it only closes if you can negotiate GPU rates well below public spot pricing or push batch sizes meaningfully past what these figures assume.

Mistral hasn't published an official throughput figure for Small 4 the way it has for Large 3 ("throughput varies significantly by sequence length mix and batch composition," per the deployment guide), so a precise cost-per-token number needs your own benchmark. As a directional check, our GPU cost-per-token benchmarks measured a similarly sized dense model, Mistral Small 3.1 24B, at $0.47/M tokens on a 2x H100 cluster. Small 4's smaller active-parameter count should push real throughput higher than that dense-model baseline, but you'd need a large multiple of improvement to land anywhere near Small 4's own $0.24/M blended API rate. Benchmark before assuming the API is beatable on cost here too.

Daily Token Volume Crossover Table for Both Models

Cluster cost is fixed regardless of how many tokens you actually push through it, up to its throughput ceiling, so in theory the API becomes relatively more expensive as volume grows. The table below shows where that crossover falls mathematically for Large 3, using the $0.70/M blended API rate against each cluster's fixed daily cost:

Here's the catch: at Large 3's own published throughput, an 8x H200 cluster tops out around 216M tokens/day (2,500 tok/s sustained for 24 hours), and a 4x B200 cluster tops out around 259M tokens/day. Both throughput ceilings sit well under the mathematical crossover point for that same cluster. In practice, that means a single Large 3 cluster never actually reaches the volume where it beats Mistral's API on cost; you'd hit the cluster's physical throughput limit first and need a second cluster, which carries the same per-token economics as the first one. Small 4's clusters are smaller and cheaper ($87-$382/day depending on config and pricing mode), and its mathematical crossover against the $0.24/M blended rate lands between roughly 364M and 1,593M tokens/day, an even higher multiple of what a 2-4 GPU cluster is likely to sustain.

For comparison, our GPT-6 vs self-hosted analysis found a 16-22M tokens/day crossover, and Claude Opus 4.8 landed between 31M and 108M tokens/day. Mistral's own API pricing is so aggressive that the equivalent crossover moves out by more than an order of magnitude, past the point a single cluster can realistically reach.

Pricing fluctuates based on GPU availability. The prices above are based on 23 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

EU Data Residency: Why Some Teams Skip the API Regardless of Price

If cost alone rarely justifies self-hosting Mistral, the real argument is data control, and Mistral's own positioning leans directly into that.

Mistral's Paris HQ, GDPR-First Posture, and EU-Default Infrastructure

Mistral operates as a French entity running primarily from EU infrastructure, with EU data residency as the default posture for La Plateforme rather than an opt-in add-on, the reverse of how most US-headquartered API providers structure the same choice (meetily.ai). That's a genuinely different starting point than DeepSeek's, where API data lands on servers inside mainland China by default, as we covered in our DeepSeek API vs self-hosted breakdown. For a team whose only requirement is "don't send data outside the EU or to a jurisdiction with broad state-access laws," Mistral's API already clears that bar without any self-hosting at all.

Mistral CEO Arthur Mensch has made his own case for why enterprises might want more than an EU-resident API: he argues that closed AI providers gain "immense leverage" over customers by watching how they use a model and learning from their internal data as they integrate it into daily operations (The Next Web). The same reporting notes Mensch offered no evidence for that specific claim, and that the pitch conveniently favors Mistral's own control-plane products, Studio and Forge. Coming from the CEO of a company that also sells an API, it's still a real acknowledgment that even Mistral's own EU-resident hosting is a different proposition than running the open weights yourself.

Zero Data Retention Is Scoped Tighter Than It Sounds: Scale Plan Only, Stateless Calls Only

Mistral doesn't use paid API inputs and outputs to train its models by default. The exception is the free Experiment tier, where inputs and outputs enter training programs automatically and you have to manually opt out through the Admin Console's privacy settings (meetily.ai).

Outside of training, default retention on stateless La Plateforme endpoints, chat completions, embeddings, moderation, OCR, and audio, is the time needed to generate the response plus 30 rolling days for abuse monitoring, then deletion. The Agents API and Fine-Tuning API don't follow that window at all: they retain data until account termination or manual deletion.

Zero Data Retention (ZDR) removes that 30-day window entirely, but it's scoped narrower than the name suggests. It's available only on the Scale plan, and only for the same stateless call types listed above. Agents, batch jobs, Conversations, Libraries, and Le Chat itself are all excluded from ZDR coverage. If your architecture routes any meaningful traffic through the Agents API for its persistent memory, or through Batch for cost savings, that traffic isn't covered by ZDR even on the Scale plan.

When Self-Hosting In-Region Still Beats an EU-Headquartered API

Mistral being EU-headquartered doesn't mean every byte of API traffic stays inside the EU under every circumstance. Mistral's Data Processing Addendum permits transfers to non-EEA "Restricted Countries" under Standard Contractual Clauses, and requires any subprocessor handling that data to meet the same protection standard the DPA itself sets (legal.mistral.ai). SCCs are a standard, legally valid transfer mechanism, not a loophole, but for organizations whose contracts or sector rules prohibit any third-country transfer clause regardless of safeguard, that detail matters more than Mistral's Paris address.

Being EU-based hasn't insulated Mistral from privacy scrutiny either. A French lawyer filed a complaint with the CNIL in February 2025 alleging that free-tier Le Chat users lacked a real opt-out from having their data used in training, unlike paying subscribers (OECD.AI Incidents Monitor). Mistral added an email-based opt-out for free users shortly after, and the CNIL hasn't issued a public decision as of this writing. That episode is a reminder that "headquartered in the EU" and "beyond regulatory question" aren't the same thing.

Self-hosting in-region still wins over the API specifically when: your data governance policy bans any non-EEA transfer clause outright, even an SCC-backed one; you're running Agents or Batch workloads with data you need gone in under 30 days rather than retained until account termination; or you're operating under EU AI Act obligations that require documented control over model weights and behavior rather than a vendor's word for it. Our EU AI Act compliance guide for GPU cloud covers the broader governance requirements that sit on top of any single vendor's data terms. On Spheron, self-hosted instances come with SSH root access, no shared GPU tenancy, and no prompt or completion logging on the provider side, which is the baseline you want before treating self-hosting as a compliance shortcut rather than just relocating the same question.

Decision Framework: Mistral API vs Self-Hosting on GPU Cloud

For the rest of this series, see how the same crossover math plays out against a much pricier API in Claude Opus 4.8 vs self-hosted LLMs, against a comparably cheap one in DeepSeek API pricing vs self-hosted LLMs, and against xAI's flagship in Grok 4.5 API pricing vs self-hosted LLMs. For the FinOps view beyond a single breakeven table, AI Inference Cost Economics 2026 covers the broader lifecycle costs, and Batch LLM Inference on GPU Cloud is the self-hosted equivalent of Mistral's own Batch API discount.

If you've already benchmarked Large 3 or Small 4 against your own workload and the data-control case outweighs the per-token math, both models are ready to deploy on Spheron's multi-GPU instances.

H200 on Spheron → | B200 capacity on Spheron → | Check current GPU pricing →