Mistral Small 3 API pricing compared across access paths

Running Mistral Small 3 in production means choosing how you call it. This Mistral Small 3 API pricing comparison looks at three access paths—Mistral’s own platform, a cloud marketplace, and an OpenAI-compatible inference gateway—so you can see where the dollar goes beyond the headline token rate. All three expose the same 24B weights and 32k context, but the surrounding cost model and operational tax differ sharply.

Direct Mistral API

Mistral hosts Small 3 on its own platform at api.mistral.ai. You get the model as published, with no intermediary between your request and their inference fleet.

Capabilities

You call the standard Mistral chat completions endpoint. Function calling, structured output, and JSON mode are supported. System prompts work as expected. There is no cross-provider fallback—if Mistral’s region is degraded, your call fails.

Price/cost model

Public list price for Small 3 is $0.10 per 1M input tokens and $0.30 per 1M output tokens. You pay exactly that, metered per token, no minimum spend. The batch API cuts this 50% if you accept asynchronous 24h turnaround. In this Mistral Small 3 API pricing comparison, direct is the baseline every other path is measured against.

Latency/throughput

You connect straight to Mistral’s infrastructure. Cold starts are negligible; p50 latency for a 24B model is competitive. Throughput scales with your concurrency tier: free tier is tightly rate-limited, paid tier expands with quota requests.

Ergonomics

Native SDKs exist for Python and TypeScript. Auth is a single bearer key. Minimal surface area:

import requests
r = requests.post(
    "https://api.mistral.ai/v1/chat/completions",
    headers={"Authorization": f"Bearer {MISTRAL_KEY}"},
    json={
        "model": "mistral-small-latest",
        "messages": [{"role": "user", "content": "Summarize this log"}],
        "temperature": 0.2
    }
)

Ecosystem

Tightest integration with Mistral’s roadmap. You get early access to new revisions and first-party docs. No built-in multi-model routing; you wire that yourself if needed.

Limits

Per-key rate limits enforced server-side. Max context is 32k tokens. No dedicated instances unless you negotiate enterprise terms.

Cloud Marketplace (Azure AI / AWS Bedrock)

Hyperscalers list Mistral Small 3 in their model catalogs. You provision it inside your existing cloud account and call it over the cloud’s inference API.

Capabilities

Same base model, but wrapped in the cloud’s schema. On Azure AI, you use the model inference REST endpoint (OpenAI-compatible-ish). On Bedrock, you invoke via bedrock-runtime with a Mistral-specific body. Both support the core chat completion use case; advanced Mistral features like native function streaming may lag behind first-party.

Price/cost model

List token price mirrors Mistral’s ($0.10/$0.30), but you also absorb cloud egress, possible endpoint hosting cost if you choose provisioned throughput, and region premiums. Committed-use plans can lower effective rate. There is no separate gateway margin—you pay the cloud and the cloud pays Mistral.

Latency/throughput

Serverless offerings add negligible overhead versus direct. Provisioned throughput instances give predictable tail latency at fixed hourly cost. Cross-region calls add milliseconds and egress line items.

Ergonomics

You manage IAM, resource groups, and endpoint URLs. Code differs from Mistral native:

# Azure AI model inference
r = requests.post(
    "https://my-resource.services.ai.azure.com/models/mistral-small/invoke",
    headers={"api-key": AZURE_KEY},
    json={"messages": [{"role": "user", "content": "Summarize this log"}]}
)

Bedrock requires signed requests via boto3, which is heavier but familiar in AWS shops.

Ecosystem

Native VPC peering, existing cloud logging, and consolidated billing. If your data already lives in that cloud, this path minimizes exfiltration reviews.

Limits

Quotas are per-region and per-account. Context window remains 32k. Dedicated instances impose minimum hourly spend even when idle.

OpenAI-Compatible Inference Gateway

A gateway fronts multiple providers behind one OpenAI-style endpoint. n4n.ai is one such gateway: a single endpoint addresses 240+ models, including Mistral Small 3, with automatic fallback when a provider is rate-limited or degraded.

Capabilities

You send the standard chat/completions payload. The gateway routes to Mistral (or an equivalent deployment) and forwards provider cache-control hints so prefix caching works end-to-end. If Mistral’s API throws 429, it fails over to another region or provider without client changes. Client routing directives are honored—pin a provider with a header when you must.

from openai import OpenAI
client = OpenAI(base_url="https://api.n4n.ai/v1", api_key=N4N_KEY)
resp = client.chat.completions.create(
    model="mistral/mistral-small-3",
    messages=[{"role": "user", "content": "Summarize this log"}],
    extra_headers={"x-n4n-route": "mistral:us-east"}
)

Price/cost model

Gateways typically pass through provider list price and add a small per-token margin or charge flat list. With n4n.ai, per-token usage metering matches provider cost; you get one invoice across all models. The Mistral Small 3 API pricing comparison stays transparent because you pay provider rate, not a mystery markup.

Latency/throughput

The extra hop adds single-digit milliseconds. Because the gateway can route to the least-loaded region or provider, tail latency can beat direct during incidents. Throughput is bounded by backend quota, but fallback hides throttling from your app.

Ergonomics

One key, one schema. Swap base_url from OpenAI or Mistral and your existing SDK works. No per-cloud IAM to juggle.

Ecosystem

Access to 240+ models behind one contract. Honors client routing directives and forwards cache-control. Unified usage dashboards replace scattered provider consoles.

Limits

Gateway-wide rate limits apply on top of model limits. Context window stays 32k (model-bound). You add a dependency on gateway uptime, though automatic fallback mitigates provider-side risk.

Head-to-head Comparison

Which to choose

Solo builders and Mistral-only stacks. Use direct API. You avoid any middleware and pay exact list. The Mistral Small 3 API pricing comparison shows no cheaper token than $0.10/$0.30, so keep it simple.

Enterprise in Azure/AWS with compliance needs. Take the cloud marketplace. You get VPC isolation, existing SOC2 controls, and consolidated cloud billing. Accept the egress and possible premium as cost of governance.

Multi-model apps or risk-averse production. A gateway wins. You write once, call Mistral Small 3 and 200+ other models, and survive provider outages via automatic fallback. n4n.ai’s per-token metering keeps the comparison honest—you pay provider rate, not a mystery margin.

Cost-optimized batch jobs. Direct batch API at 50% off list is unbeatable if you can tolerate async. Gateways may not expose that discount yet; check before assuming parity.

Latency-critical single-region. Direct or provisioned cloud instance. The gateway hop is negligible but adds a dependency; only use it if you need fallback or multi-model access.

Pick the path that matches your operational surface, not the one with the cutest dashboard. Token price is identical across all three; the difference is the tax you pay for convenience, isolation, or resilience.