Mistral API pricing is useful because the provider is not trying to sell one model for every job. The current Mistral catalog spans cheap utility models, code-specialist routes, general chat models, reasoning models, multimodal Pixtral models, and open-weight families that can also appear through other hosts.
In the current AI Pricing Guru tracker, Mistral Small 4 and Devstral Small 2 both cost $0.10 per million input tokens and $0.30 per million output tokens. Mistral Large 3 costs $0.50 input and $1.50 output. Codestral costs $0.30 input and $0.90 output. At the high end, Magistral Medium costs $2.00 input and $5.00 output, while Pixtral Large and Mixtral 8x22B are both tracked at $2.00 input and $6.00 output.
That price ladder makes Mistral a strong candidate for teams that want lower-cost text generation, coding models, European provider diversity, or an open-model strategy. It is not automatically the cheapest option for every workload. OpenAI GPT-5 nano and some Groq routes can beat Mistral on simple utility calls, while Claude or GPT premium models can still win when a higher success rate prevents retries.
For live rates, keep the Mistral pricing page, the full AI pricing table, and the token cost calculator open while modeling your own traffic. For direct alternatives, see Mistral vs OpenAI pricing, OpenAI pricing, Anthropic pricing, and Groq pricing.
Mistral API Pricing: Quick Reference
All prices below are USD per 1 million tokens from the current AI Pricing Guru tracker.
| Mistral model | Status | Input | Cached input | Output | Best fit |
|---|---|---|---|---|---|
| Ministral 3B | Active | $0.10 | n/a | $0.10 | Cheapest utility calls, routing, tagging |
| Mistral Small 4 | Active | $0.10 | n/a | $0.30 | Low-cost chat, extraction, summarization |
| Devstral Small 2 | Active | $0.10 | n/a | $0.30 | Cheap code triage and simple coding tasks |
| Ministral 8B | Active | $0.15 | n/a | $0.15 | Lightweight edge-style text workloads |
| Mistral NeMo | Active | $0.15 | n/a | $0.15 | Open-model portability and multilingual utility |
| Ministral 14B | Active | $0.20 | n/a | $0.20 | Stronger small-model utility tier |
| Codestral | Active | $0.30 | n/a | $0.90 | Code completion, refactors, IDE workflows |
| Devstral Medium 2 | Active | $0.40 | n/a | $2.00 | Harder coding-agent tasks |
| Mistral Large 3 | Active | $0.50 | n/a | $1.50 | General chat, support, analysis |
| Magistral Small | Active | $0.50 | n/a | $1.50 | Lower-cost reasoning route |
| Mistral Medium 3.5 | Active | $1.50 | n/a | $7.50 | Higher-quality general workloads |
| Magistral Medium | Active | $2.00 | n/a | $5.00 | Premium Mistral reasoning |
| Pixtral Large | Active | $2.00 | n/a | $6.00 | Multimodal image-aware workloads |
| Mixtral 8x22B | Active | $2.00 | n/a | $6.00 | Large open Mixtral route |
| Mixtral 8x7B | Active | $0.70 | n/a | $0.70 | Open Mixtral workloads |
The main takeaways:
- Mistral’s cheapest active tracked route is Ministral 3B at $0.10 input and $0.10 output.
- Mistral Small 4 and Devstral Small 2 are the practical low-cost defaults for text and code.
- Codestral is priced as a specialist coding model, not as the cheapest utility route.
- Mistral Large 3 and Magistral Small share the same tracked price, but target different workload shapes.
- No Mistral row in the current tracker includes cached-input pricing, so repeated context should be modeled at normal input rates unless your direct provider contract says otherwise.
Which Mistral Model Should You Use?
Use Ministral for cheap utility calls
Ministral 3B, Ministral 8B, and Ministral 14B are the first models to test when the task is simple, high-volume, and easy to validate. At $0.10 to $0.20 per million input tokens and the same rate for output, these routes are inexpensive enough for infrastructure-style calls.
Good fits include:
- intent routing
- tagging and classification
- title generation
- simple extraction
- query rewriting
- short internal summaries
- policy pre-checks
- routing before a stronger model
The risk is quality. A cheap utility model is only cheap if it produces accepted outputs. Use strict schemas, validators, confidence thresholds, and escalation rules. If the task is customer-facing or hard to check automatically, start with Mistral Small 4 or Mistral Large 3 instead.
Use Mistral Small 4 for low-cost production text
Mistral Small 4 costs $0.10 input and $0.30 output per million tokens. It is one of the best starting points for normal text workloads because it keeps output cost low while giving you a general-purpose Mistral route.
Use it for:
- support drafts
- help-center RAG answers
- document extraction
- summarization
- structured content generation
- customer-email drafts
- internal workflow automation
Mistral Small 4 is especially attractive when output volume matters. A model with cheap input but expensive output can become costly in chat, support, or writing workflows. Small 4 keeps both sides of the bill low enough to use as a default for low-risk production traffic.
Use Codestral and Devstral for developer workflows
Mistral has a clearer coding story than many non-frontier providers because Codestral and Devstral are purpose-built for developer work.
Codestral costs $0.30 input and $0.90 output per million tokens. It is a good candidate for code completion, repository-aware assistance, mechanical refactors, test generation, and IDE workflows where a code-specialist model can beat a generic cheap model.
Devstral Small 2 costs the same as Mistral Small 4 at $0.10 input and $0.30 output. That makes it a low-cost route for code triage, simple patch drafts, bug explanation, and lightweight agent steps. Devstral Medium 2 costs $0.40 input and $2.00 output, so it belongs in the escalation tier for harder changes.
For coding agents, compare on cost per accepted change. If Codestral or Devstral reduces failed patches, human cleanup, or premium-model escalation, it can beat a lower sticker-price model.
Use Mistral Large and Magistral when quality needs to rise
Mistral Large 3 costs $0.50 input and $1.50 output per million tokens. It is the stronger general Mistral route in the current tracker, and it is still far below premium frontier-model pricing from OpenAI and Anthropic.
Use Mistral Large 3 when Small 4 is not reliable enough for:
- nuanced support replies
- longer synthesis
- more complex extraction
- internal analysis
- policy-aware writing
- moderate agent planning
Magistral Small is tracked at the same $0.50 input and $1.50 output price, while Magistral Medium rises to $2.00 input and $5.00 output. Treat Magistral as the reasoning branch of the Mistral stack. It is worth testing when the workload needs more step-by-step judgment, but only if the better success rate justifies the higher output price.
Use Pixtral for multimodal work
Pixtral Large is tracked at $2.00 input and $6.00 output. It is not a bargain utility model. Its role is image-aware analysis, multimodal extraction, screenshot interpretation, and workflows where text-only models cannot see the evidence.
For buyers, the important comparison is not Pixtral versus Mistral Small 4. It is Pixtral versus OpenAI, Google, Anthropic, or another multimodal route on the exact image workload. If image understanding prevents human review or avoids a separate OCR and classification pipeline, the higher token rate can still make sense.
Example Monthly Costs
Assume a customer-support and internal automation product uses 100 million input tokens and 50 million output tokens per month.
| Mistral model | Monthly token cost |
|---|---|
| Ministral 3B | $15.00 |
| Ministral 8B | $22.50 |
| Mistral Small 4 | $25.00 |
| Devstral Small 2 | $25.00 |
| Ministral 14B | $30.00 |
| Codestral | $75.00 |
| Mistral Large 3 | $125.00 |
| Magistral Small | $125.00 |
| Devstral Medium 2 | $140.00 |
| Magistral Medium | $450.00 |
| Mistral Medium 3.5 | $525.00 |
For this workload, Mistral Small 4 and Devstral Small 2 are both $25 per month before retries and infrastructure costs. Mistral Large 3 is $125 per month, which is still modest for a stronger route. The expensive rows are only worth using when they materially improve acceptance rate.
Now assume a document-extraction workload with 1 billion input tokens and 100 million output tokens per month.
| Mistral model | Monthly token cost |
|---|---|
| Ministral 3B | $110.00 |
| Mistral Small 4 | $130.00 |
| Mistral NeMo | $165.00 |
| Codestral | $390.00 |
| Devstral Medium 2 | $600.00 |
| Mistral Large 3 | $650.00 |
| Mistral Medium 3.5 | $2,250.00 |
| Magistral Medium | $2,500.00 |
Input-heavy workloads reward low input prices. Ministral 3B, Mistral Small 4, and Mistral NeMo are the first candidates to benchmark for extraction, tagging, and preprocessing. Codestral and Devstral only make sense in this scenario if the extraction is code-heavy or developer-specific.
Hidden Mistral Costs to Watch
No tracked cached-input discount
The current AI Pricing Guru tracker does not list cached-input pricing for Mistral routes. That matters for RAG systems, coding agents, and workflows that repeat long system prompts, policies, or tool schemas.
If your app sends 20,000 repeated input tokens on every request, providers with prompt caching can look much cheaper than their base input price suggests. With Mistral, model repeated context at normal input rates unless your direct plan or hosted route gives you a discount.
Output price can dominate
Several Mistral models have output prices 3x to 5x higher than input. Mistral Medium 3.5 is $1.50 input and $7.50 output. Magistral Medium is $2.00 input and $5.00 output. Pixtral Large and Mixtral 8x22B are $2.00 input and $6.00 output.
That does not make them bad models. It means verbose answers are expensive. Use compact prompts, response budgets, schemas, and channel-specific answer lengths.
Open models still have operating costs
Mistral’s open-model story is strategically valuable, but self-hosting is not free. You need GPU capacity, serving software, observability, incident response, model updates, and security review. For low or unpredictable volume, the hosted API can be cheaper than running infrastructure yourself.
The best use of open optionality is leverage. Start with hosted API calls, identify stable high-volume workloads, then test alternate hosting or self-hosting only where the operational work is justified.
Quality failures erase savings
Cheap token rates are only useful when the model completes the task. Track first-pass success rate, retry rate, escalation rate, latency, JSON validity, and human review time. The buying metric is cost per accepted result, not cost per million tokens.
Best Mistral Model by Use Case
| Use case | Start with | Escalate to |
|---|---|---|
| Routing and tagging | Ministral 3B | Mistral Small 4 |
| Support drafts | Mistral Small 4 | Mistral Large 3 |
| High-volume extraction | Mistral Small 4 or Mistral NeMo | Mistral Large 3 |
| Code completion | Codestral | Devstral Medium 2 |
| Coding-agent triage | Devstral Small 2 | Codestral or Devstral Medium 2 |
| Reasoning-heavy analysis | Magistral Small | Magistral Medium |
| Image-aware extraction | Pixtral Large | OpenAI, Google, or Anthropic multimodal fallback |
| Open-model portability | Mistral NeMo or Mixtral | Alternate host or self-hosted deployment |
The most useful Mistral architecture is layered:
- Use Ministral or Mistral Small 4 for cheap, high-volume traffic.
- Use Codestral and Devstral for developer-specific work.
- Use Mistral Large 3 or Magistral when quality needs to rise.
- Use Pixtral only for workloads that actually need image understanding.
- Keep OpenAI, Anthropic, Google, DeepSeek, or Groq available when a Mistral route fails evals.
Mistral vs Other Providers
Compared with OpenAI pricing, Mistral usually wins against premium GPT routes on raw token price. OpenAI still has a stronger cached-input story, broader multimodal APIs, and a deeper premium ladder. GPT-5 nano can also beat Mistral on very simple input-heavy calls.
Compared with Anthropic pricing, Mistral is much cheaper on most listed token rates. Claude can still win when coding judgment, long-context reasoning, writing quality, or fewer retries matters more than sticker price.
Compared with Groq pricing, Mistral is a model provider with a stronger open-model identity, while Groq is a fast inference host across multiple model families. Benchmark Groq when latency is the product feature. Benchmark Mistral when its own model behavior, EU vendor story, or coding models are the draw.
Compared with DeepSeek pricing, Mistral is often more expensive on raw tokens but gives buyers a different vendor profile, code-specialist models, Pixtral multimodal routes, and open-model optionality.
FAQ
What is the cheapest Mistral API model?
Ministral 3B is the cheapest active Mistral model in the current AI Pricing Guru tracker at $0.10 per million input tokens and $0.10 per million output tokens.
How much does Mistral Small 4 cost?
Mistral Small 4 costs $0.10 per million input tokens and $0.30 per million output tokens in the current tracker.
How much does Codestral cost?
Codestral costs $0.30 per million input tokens and $0.90 per million output tokens. It is Mistral’s specialist route for code completion, refactoring, and developer workflows.
Does Mistral have prompt caching?
The current AI Pricing Guru tracker does not list cached-input pricing for Mistral models. If your workload depends on repeated prompts or long tool schemas, model those tokens at normal input rates unless your direct provider plan says otherwise.
Is Mistral cheaper than OpenAI and Claude?
Often, yes, especially compared with premium GPT and Claude routes. But Mistral is not always cheaper than every model. OpenAI GPT-5 nano and some Groq routes can be cheaper for simple utility calls. Compare the exact workload with the token cost calculator.
Bottom Line
Mistral API pricing is strongest when you route by workload. Ministral and Mistral Small 4 handle cheap utility and text traffic. Codestral and Devstral cover developer workflows. Mistral Large 3 and Magistral provide stronger escalation routes. Pixtral belongs in image-aware workflows.
Do not pick Mistral only because the table looks cheaper than frontier models. Test it on your own prompts, measure accepted outputs, and compare against OpenAI, Anthropic, Google, DeepSeek, and Groq. If Mistral passes those evals, it can lower your bill while giving you vendor diversity and a clearer open-model path.
Last updated: July 20, 2026, using AI Pricing Guru’s tracked pricing data.