API model string
devstralContext window
256K tokens
Max output not published
- Knowledge cutoff
- not published
- Released
- 2025
- Modalities
- text in, text out (code-specialist)
Last verified July 10, 2026
Rate limits by tier
Mistral La Plateforme organizes access into three tiers: a free Experiment tier for evaluation, a pay-as-you-go tier activated by adding a billing card, and a custom Enterprise tier negotiated with sales. Mistral no longer publishes exact RPM numbers publicly — your live org ceilings appear in Admin Console → Limits.
| Tier | Requirements | RPM | TPM | RPD | Concurrent | Notes |
|---|---|---|---|---|---|---|
| Free (Experiment) | No card required — sign up at mistral.ai and use the API for evaluation | not published (check Admin Console → Limits) | ~1B tokens/month cap | not published | — | Rate-limited and intended for eval only, not for production workloads. Codestral has a separate free IDE tier — Devstral is a different endpoint. |
| Pay-as-you-go (paid) | Enable billing in your La Plateforme account settings | ~300 RPM baseline (per third-party trackers — confirm in Admin Console) | not published | not published | — | No monthly minimums. Paid baseline is higher than free tier. Limits grow with usage history. Pin explicit model versions in production rather than using the -latest alias. |
| Enterprise (custom) | Contact Mistral sales | not published | not published | not published | — | Includes SAML SSO, audit logs, custom limits, and dedicated SLA. Mistralship startup program offers ~$30K in credits via application-only process. |
Swipe the table sideways to see every limit column.
- 1.Mistral no longer publishes exact free-tier RPM numbers; check Admin Console → Limits for your current org ceilings.
- 2.Standard HTTP 429 with Retry-After header; implement exponential backoff with jitter.
- 3.The -latest alias (e.g. devstral-latest) can silently pull a new model version with different pricing — pin explicit versioned strings (e.g. devstral-2025-05) in production.
Limits verified against the Mistral AI docs, July 10, 2026.
Token pricing
What you pay per million tokens (USD). Input and output are billed separately.
Input
$0.30
per 1M tokens
Output
$0.90
per 1M tokens
Batch discount
−50%
on batch jobs
- The $0.30/$0.90 per MTok figures are from the Codestral API rate card cited in third-party trackers; Devstral hosted pricing is described as 'between Mistral Small and Medium tiers' without an exact figure published on mistral.ai/pricing as of July 2026. Treat as an estimate — verify on mistral.ai/pricing before populating any budget calculation.
- Context caching is available and discounts repeated-prefix input; exact cached rate not published — verify on mistral.ai/pricing.
- Batch API provides 50% off input and output on supported models; confirm Devstral batch endpoint availability before use.
- Le Chat Pro subscription ($14.99/mo) does NOT include API credits — these are separate billing lines.
- Devstral 2 is a separate endpoint with its own rate card; verify on mistral.ai/pricing.
Side-project code assistant
~$2.40
per month
Assumptions
5M tokens in / 1M tokens out per month
5M × $0.30/MTok + 1M × $0.90/MTok = $1.50 + $0.90 = $2.40 (estimate using Codestral rate — verify Devstral exact price on mistral.ai/pricing)
Growing dev team agentic workflows
~$28.80
per month
Assumptions
60M tokens in / 12M tokens out per month
60M × $0.30/MTok + 12M × $0.90/MTok = $18.00 + $10.80 = $28.80 (estimate — no cached scenario published for Devstral)
High-volume code generation pipeline
~$192
per month
Assumptions
400M tokens in / 80M tokens out per month
400M × $0.30/MTok + 80M × $0.90/MTok = $120 + $72 = $192 (estimate — verify against current Devstral rate card)
Run your own numbers
Drag your real monthly token volumes and watch the bill update live — priced against rival models at the same usage.
Estimated devstral spend
$14/mo
Input: $9.00
Output: $4.50
30M in × $0.300 + 5M out × $0.900 = $14
Same volume, priced across models
- devstralThis model$14
- Mistral Medium 3.5$22
- Mistral Large 3$90
Rivals priced at their published input/output rates for the same monthly volumes. Prompt caching is model-specific, so it is applied to devstral only. Estimates for comparison; real bills vary with request shape and long-context surcharges.
devstral vs the alternatives
Devstral competes as a code-specialist model against Mistral's own general-purpose flagship and mid-tier offerings on the same La Plateforme infrastructure.
| Aspect | devstral | Mistral Large 3 | Mistral Medium 3.5 |
|---|---|---|---|
| Context window | 256K (per third-party trackers — verify) | 128K | 128K |
| Input $/MTok | ~$0.30 (verify) | $2.00 (verify) | $0.40–1.50 (verify) |
| Output $/MTok | ~$0.90 (verify) | $6.00 (verify) | $2.00–7.50 (verify) |
| Specialization | Agentic code edits | General flagship | General mid-tier |
| Open-weight option | Yes (Devstral Small, Apache 2.0) | No | No |
| Batch discount | 50% | 50% | 50% |
| Free eval tier | Yes (Experiment) | Yes (Experiment) | Yes (Experiment) |
| EU data residency | Yes (Paris) | Yes (Paris) | Yes (Paris) |
Swipe the table sideways to see every model.
Hitting a 429? The playbook
The exact errors you'll see
HTTP 429 Too Many Requestsrate_limit_errorRetry-After: <seconds>Why it happens & how to fix it
RPM exceeded on free Experiment tier
Upgrade to pay-as-you-go or space requests further apart. The free tier has a very low RPM ceiling that is not published; check Admin Console → Limits for your current org ceiling.
Burst spike on paid baseline (~300 RPM)
Implement exponential backoff with jitter. Mistral returns a Retry-After header in 429 responses — read it and wait the specified duration before retrying.
-latest alias routed to a new model version mid-deploy
Pin an explicit versioned model string (e.g. devstral-2025-05) in production. The -latest alias can silently switch to a new version with different rate cards or output behavior.
Monthly ~1B token cap exhausted on free tier
Switch to pay-as-you-go; there is no monthly token cap on paid tiers.
Codestral vs Devstral endpoint confusion
Codestral has a separate free IDE tier and is a different endpoint. Confirm your base URL is https://api.mistral.ai/v1 and your model string is 'devstral' (or pinned version), not a Codestral string.
Retry strategy
Honor the Retry-After header in the 429 response — wait that many seconds before retrying. Use exponential backoff: wait = min(2^attempt, 60) seconds with ±10% random jitter to prevent thundering herd. Mistral La Plateforme is OpenAI-compatible (swap base_url to https://api.mistral.ai/v1), so standard OpenAI SDK retry patterns apply directly.
1import OpenAI from 'openai';23const client = new OpenAI({4 apiKey: process.env.MISTRAL_API_KEY,5 baseURL: 'https://api.mistral.ai/v1',6});78async function callDevstralWithRetry(9 prompt: string,10 maxRetries = 511): Promise<string> {12 for (let attempt = 0; attempt < maxRetries; attempt++) {13 try {14 const response = await client.chat.completions.create({15 model: 'devstral-2025-05', // pin explicit version, not -latest16 messages: [{ role: 'user', content: prompt }],17 });18 return response.choices[0].message.content ?? '';19 } catch (err: unknown) {20 const error = err as { status?: number; headers?: Record<string, string>; message?: string };21 if (error.status === 429) {22 const retryAfter = parseFloat(error.headers?.['retry-after'] ?? '0');23 const jitter = 1 + (Math.random() * 0.2 - 0.1);24 const delay = retryAfter > 025 ? retryAfter * 100026 : Math.min(Math.pow(2, attempt) * 1000 * jitter, 60000);27 console.warn(`Rate limited. Retrying in ${(delay / 1000).toFixed(1)}s (attempt ${attempt + 1}/${maxRetries})`);28 await new Promise(resolve => setTimeout(resolve, delay));29 } else {30 throw err;31 }32 }33 }34 throw new Error('Max retries exceeded for Devstral API call');35}How to raise your limits
The ladder from the starter tier to enterprise — what each rung takes, and what it unlocks.
Experiment (free)
ImmediateSign up at mistral.ai — no card required. Access the Devstral endpoint immediately for testing.
Unlocks: ~1B tokens/month cap, low RPM (not published — check Admin Console → Limits), sufficient for evaluation and prototyping
Pay-as-you-go
Immediate upon card additionAdd a billing card in La Plateforme account settings. No minimum spend.
Unlocks: ~300 RPM baseline (per third-party trackers — verify in Admin Console), no monthly token cap, production-ready access
Admin Console → Limits (self-serve)
Immediate for self-serve increasesNavigate to Admin Console → Limits to view your current org ceilings. Some increases are available without contacting support.
Unlocks: Higher limits for your organization without a sales call
Enterprise / Mistralship
Sales negotiation or Mistralship review processContact Mistral sales for custom Enterprise terms. Apply to the Mistralship startup program at mistral.ai/mistralship for ~$30K in credits (application-only).
Unlocks: Custom RPM/TPM limits, SAML SSO, audit logs, dedicated SLA, startup credits
Cut your token spend
Context caching
Significant reduction on repeated-prefix input (exact discount not published — verify on mistral.ai/pricing)Structure prompts with a long, stable shared system-prompt prefix (e.g., the entire codebase context or a large file). Mistral caches repeated prefixes and discounts subsequent input tokens that hit the cache.
Batch API
-50% on input and output cost for non-latency-sensitive tasksSubmit batches of code review, analysis, or generation tasks asynchronously. Confirm Devstral batch endpoint availability in Admin Console before building the pipeline.
Open-weight fallback (Devstral Small)
-100% per-token cost on self-hosted inferenceDevstral Small is Apache 2.0 open weights. Run it on your own GPU infrastructure for zero per-token cost on high-volume, lower-complexity code tasks. Reserve the hosted endpoint for complex agentic edits.
Pin model versions
Prevents silent rate-card and behavior changesUse an explicit version string (e.g. devstral-2025-05) instead of devstral-latest in all production deployments. The -latest alias can silently pull a new version with different pricing or output characteristics.
EU data residency at no premium
GDPR compliance without additional cost or configurationMistral routes all La Plateforme traffic through Paris by default. No regional surcharge, no additional configuration needed — unlike US-only providers that charge for EU residency options.
Use dedicated Codestral tier for IDE autocomplete
Avoids consuming Devstral quota on lower-value tasksIf your use case is IDE autocompletion (not agentic edits), Codestral's dedicated free IDE tier is more appropriate and preserves your Devstral quota for complex agentic coding workflows.
Frequently asked questions
How do I increase Devstral API rate limits?
Start by adding a billing card to your La Plateforme account, which immediately upgrades you from the free Experiment tier to the pay-as-you-go baseline (~300 RPM per third-party trackers). Then check Admin Console → Limits for your current org ceilings — some self-serve increases are available without contacting support. For higher custom limits, contact Mistral sales or apply to the Mistralship startup program for ~$30K in credits.
Is the Devstral API free?
Yes — Mistral offers a free Experiment tier with no billing card required. It includes approximately 1 billion tokens per month but has a low RPM ceiling (not published publicly; check Admin Console → Limits). It is intended for evaluation, not production. Paid pay-as-you-go access unlocks higher limits with no monthly minimum.
What is Devstral's API pricing?
Per third-party trackers, Devstral is approximately $0.30 per million input tokens and $0.90 per million output tokens — verify on mistral.ai/pricing before quoting, as Mistral describes the price as 'between Small and Medium tiers' without a published figure. A 50% batch discount is available on supported models. Context caching discounts repeated-prefix input at a rate not yet published.
Devstral vs Codestral — what's the difference?
Codestral is Mistral's IDE autocompletion model with a dedicated free tier for tools like Continue.dev and VS Code plugins. Devstral is a separate agentic code-editing model designed for multi-step coding tasks and PR workflows. They have different endpoints, different pricing, and different rate quotas. Use Codestral's free IDE tier for autocomplete and reserve Devstral for agentic edits.
What does HTTP 429 mean on the Devstral API?
HTTP 429 Too Many Requests means you've exceeded your tier's RPM ceiling. The response includes a Retry-After header specifying how many seconds to wait. Implement exponential backoff (wait = min(2^attempt, 60) seconds ± 10% jitter) and honor that header. On the free tier, the RPM limit is very low; upgrade to pay-as-you-go for meaningful throughput.
Can I self-host Devstral?
Devstral Small is released as Apache 2.0 open weights, which permits commercial self-hosting at zero per-token cost. This is the open-weight variant of the model family. The full hosted Devstral endpoint on La Plateforme is a managed service. For high-volume, lower-complexity coding tasks, running Devstral Small on your own GPU infrastructure can eliminate per-token costs entirely.
How does Devstral compare to GPT or Claude for code tasks?
Devstral is purpose-built for agentic code editing — multi-step file modifications, PR generation, and structured edit workflows — with a 256K token context window (per third-party trackers; verify). At approximately $0.30/$0.90 per MTok, it is significantly cheaper than most flagship coding models. The EU Paris routing is a compliance advantage over US-only alternatives. RapidDev's engineering team can help you benchmark Devstral against your specific stack — reach out at rapidevelopers.com/contact.
Why does Mistral not publish exact RPM limits for Devstral?
Mistral moved away from publishing fixed public RPM tables and now surfaces per-organization limits in Admin Console → Limits. This means your actual ceiling can differ based on account history and usage patterns. The ~300 RPM figure for the paid tier comes from third-party trackers — confirm your specific ceiling in the Admin Console, as it is the authoritative source for your org.
We build AI apps that don't hit rate limits
- Retry, backoff & caching built in
- Multi-provider fallback routing
- Fixed price, you own the code
30-min call. No commitment.