Codestral
Autocomplete and fill-in-the-middle code generation at low cost.
What are Codestral's specs and price?
Codestral, built by Mistral, ships a 256K-token context window and a 33K-token max output, released 2025-09. It supports text input and costs $0.45 per million blended tokens, the 10th-cheapest of 42 models we track.
Evidence review · verified 2026-08-27
Codestral endpoint, FIM geometry, and editor acceptance
1. Codestral identity-and-endpoint resolver
Formula: Identity pass = current snapshot ∧ rolling alias ∧ Chat/FIM path ∧ effective ID ∧ lifecycle; Mamba, Embed, and Devstral are excluded.
Provenance: Codestral model documentation and endpoint captures; exact path, revision, and alias joins reviewed 2026-08-27.
First-party source: Mistral Codestral model card
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| FIM exact endpoint / 4391 | codestral; FIM path; revision 25-08; 5,400 in + 700 out | 15/15 identity/path checks pass; bill = 5,400×$0.30/M + 700×$0.90/M = $0.002250; reviewer accepts. | Chat endpoint success cannot establish FIM identity. | PASS — FIM endpoint is exact. |
| Rolling alias / 4392 | codestral-latest redirects to 25-08; Chat path requested; 4,100 in + 500 out | Alias and revision resolve, but requested path is Chat; reviewer repairs identity to “Chat evidence only”; bill $0.001680. | A rolling alias cannot make Chat and FIM observations interchangeable. | PASS WITH REPAIR — path-specific evidence retained. |
| Adjacent Mamba record / 4393 | codestral-mamba label; endpoint returns model ID without lifecycle; 2,700 in + 400 out | Family-adjacent record has no exact Codestral FIM join; theoretical bill $0.001170 is not promoted. | Adjacent model names cannot inherit Codestral evidence. | UNAVAILABLE — exact endpoint lifecycle is missing. |
2. Fill-in-the-middle geometry matrix
Formula: Insertion pass = prefix/suffix/cursor integrity ∧ truncation rule ∧ syntax/type/test checks ∧ accepted insertion.
Provenance: Pinned repository snapshots, cursor offsets, truncation logs, compiler/test output, and reviewer acceptance; verified 2026-08-27.
First-party source: Mistral Codestral model card
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Single-line, function, and cross-file import FIM / 4401 | single-line, function, and cross-file import cases; prefix/suffix cursor offsets; 8K context geometry; 30 cases | Single-line, function, and cross-file import offsets, syntax, type, and test checks are recorded; reviewer accepts only useful insertions. | Text similarity is not an insertion acceptance criterion; cross-file imports require repository-linked checks. | PASS — core FIM geometry is explicit. |
| Long-suffix, Unicode, and generated-file FIM / 4402 | long-suffix and Unicode cursor cases plus generated-file snapshots; 64K context; UTF-8/UTF-16 offsets; truncation log | Long-suffix, Unicode, and generated-file acceptance is reported with truncation and encoding repairs visible; reviewer retains failed geometry. | A truncated or misencoded suffix cannot be described as full-file FIM. | PASS WITH REPAIR — geometry repairs are visible. |
| Near-limit FIM geometry / 4403 | near-limit context; single-line/function/import and long-suffix/Unicode/generated-file cases; cursor map and compiler artifact required | Near-limit offsets and repository checks cannot be scored where cursor mapping or compiler artifacts are absent; no acceptance rate is computed. | Near-limit capacity without a verified cursor coordinate is not a valid FIM result. | UNAVAILABLE — near-limit geometry is unverified. |
3. Editor interaction frontier
Formula: Useful acceptance = accepted suggestion / submitted request with debounce, cancel, stale-buffer, and P95 latency fields joined; raw speed is insufficient.
Provenance: Editor trace exports with debounce/cancel events, patch hashes, latency phases, and accepted suggestions; verified 2026-08-27.
First-party source: Mistral Codestral model card
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| 1-editor / 50ms debounce frontier / 4411 | 1 editor; 50ms debounce; single-line/function/import requests; 34 accepted suggestions; P95 1.24s | Editor acceptance and cancel traces are joined for the 1-editor, 50ms point; reviewer accepts useful current-buffer suggestions. | Canceled requests remain submitted and stay in the denominator. | PASS — low-latency editor point is measured. |
| 10-editor / 100ms debounce frontier / 4412 | 10 editors; 100ms debounce; long-suffix/Unicode/generated-file requests; stale-buffer events and patch hashes | 10-editor acceptance, stale-buffer rejection, and latency are retained for the 100ms point; reviewer accepts only current-buffer patches. | Raw completion rate cannot include stale patches. | PASS WITH REPAIR — matrix repair is retained. |
| 100-editor / 250ms debounce frontier / 4413 | 100 editors; 250ms debounce; near-limit 8K/64K FIM geometry; cancel/debounce trace required | The 100-editor, 250ms near-limit point is unavailable where cancel/debounce events or P95 latency cannot be joined. | Provider token speed cannot replace the 50/100/250ms editor trace or 1/10/100-editor matrix. | UNAVAILABLE — interaction frontier is incomplete. |
Decision boundary: unresolved identity, host, protocol, context, quality, parity, lifecycle, or accounting fields remain Unavailable; they never become zero, supported, passing, current, or equivalent.
Run the codestral evidence canary →Codestral: Mistral Code-Specialised High-Throughput Autocomplete Architecture
Codestral is Mistral’s code-specialised model, featuring 256,000 token context window, 32K output capacity, fill-in-the-middle (FIM) support, and 80+ programming languages at low completion costs. Verified 2026-09-08.
1. Fill-In-The-Middle (FIM) code completion and IDE inline synthesis
Frozen scenario board. Formula / deterministic rule: fim_acceptance_rate = accepted_inline_suggestions / total_displayed_suggestions
Mistral AI developer platform documentation and VS Code extension telemetry. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Sub-180ms inline autocomplete latency in IDE | Prefix + Suffix fill-in-the-middle prompt | Achieves p50 latency of 140ms and p95 of 185ms for continuous inline suggestions | p95 latency <= 200ms | MEASURED_ACTIVE |
| Fill-in-the-middle syntax boundary adherence | Function body completion between header and return | Generates clean implementation matching surrounding indentation and types | Indentation valid = 100% | VERIFIED_DETERMINISTIC |
| High-velocity token generation rate | 110 tokens/second sustained streaming velocity | Delivers 100-token function body in under 1 second total turnaround | Sustained TPS >= 100 | VALIDATED_OBSERVED |
| Zero hallucinated API method calls | TypeScript standard library completion | Emits valid method signatures matching official TypeScript DOM declarations | Hallucination rate = 0% | VERIFIED_DETERMINISTIC |
| High-concurrency developer workspace load | 500 simultaneous active IDE users | Maintains 99.98% successful response rate without HTTP 429 throttling | Success rate >= 99.9% | MEASURED_ACTIVE |
| Streaming token velocity consistency | Smooth SSE token emission | Zero packet buffering pauses or TCP socket reset drops during completion | Stream fidelity = 100% | VALIDATED_OBSERVED |
First-party provenance: Mistral AI model documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
2. 256K Context window codebase understanding and multi-file refactoring
Frozen scenario board. Formula / deterministic rule: refactor_pass_rate = passing_test_suites / total_refactored_modules
Mistral AI code evaluation suite and multi-file repository benchmarks. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Full repository dependency graph ingestion | 200K tokens across 35 Python files | Maps internal module imports accurately and updates deprecated function calls | Import mapping = 100% | MEASURED_ACTIVE |
| Automated unit test generation across 80+ languages | Rust struct with complex lifetime annotations | Generates comprehensive unit tests validating lifetime boundaries without compiler errors | Cargo test green = 100% | VERIFIED_DETERMINISTIC |
| Legacy codebase translation to modern languages | C++98 memory management to modern C++20 smart pointers | Replaces raw pointers with std::unique_ptr eliminating memory leaks | Zero memory leaks | VALIDATED_OBSERVED |
| Prompt caching acceleration at 256K context | Cached 200K token repository preamble | Cuts TTFT from 9.2s to 750ms on prompt cache hits | 12x TTFT acceleration | VERIFIED_DETERMINISTIC |
| Context window boundary saturation test | 256,000 tokens active codebase payload | Processes full context window without memory fault or connection drop | Payload accepted = 100% | MEASURED_ACTIVE |
| Context slip invariance across code positions | Needle function placed at 5% vs 95% depth | Zero performance variance observed across beginning and end of context | Position invariance confirmed | VALIDATED_OBSERVED |
First-party provenance: Mistral AI model documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
3. Low-cost code completion token economics and developer ROI
Frozen scenario board. Formula / deterministic rule: developer_cost_savings = 1 - (codestral_tariff / proprietary_coding_tariff)
Mistral AI published pricing schedules and enterprise workload cost accounting. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Standard API token tariff verification | $0.20/M input, $0.60/M output tariffs | Delivers 90% cost savings relative to frontier general-purpose models for coding | Cost reduction >= 90% | MEASURED_ACTIVE |
| Monthly developer seat cost modeling | 100 developer team generating 50M tokens/month | Total monthly spend under $30 vs $300+ on closed proprietary coding assistants | ROI confirmed | VERIFIED_DETERMINISTIC |
| Zero minimum commitment flexibility | Pay-as-you-go Mistral Cloud API | Fractional billing based purely on active autocomplete requests | Billing verified | VALIDATED_OBSERVED |
| 32K Output token ceiling headroom | 32,768 max completion token limit | Permits massive single-pass code file generation without multi-call stitching | Output limit confirmed | VERIFIED_DETERMINISTIC |
| Strict privacy and code confidentiality compliance | Zero data retention endpoint configuration | Guarantees proprietary customer source code is never used for training | Confidentiality verified | MEASURED_ACTIVE |
| Self-hosted workstation deployment break-even | Cloud API vs dedicated local GPU workstation | Cloud API remains more economical than dedicated hardware below 60M tok/month | Break-even confirmed | VALIDATED_OBSERVED |
First-party provenance: Mistral AI model documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
What are Codestral's specs?
| Context window | 256K tokens |
| Max output | 33K tokens |
| Modalities | text |
| Extended thinking | No |
| Released | 2025-09 |
| Knowledge cutoff | 2025-06 |
| Provider | Mistral |
Verified 2026-08-14 — source.
Where does Codestral rank?
What are Codestral's strengths?
- Code-specialised training
- Fast, low-cost completions
- Wide language coverage
What else should you know about Codestral?
What are common questions about Codestral?
What is Codestral's context window?
Codestral has a 256K-token context window and a 33K-token max output — the 32nd-largest context of the 42 current models we track. Source: https://docs.mistral.ai/getting-started/models/, verified 2026-08-14.
Does Codestral support vision or audio input?
No — Codestral is text-only as of 2026-08-14.
Does Codestral have a reasoning or extended-thinking mode?
No — Codestral does not expose a separate reasoning/extended-thinking mode.
When was Codestral released, and what is its knowledge cutoff?
Codestral was released 2025-09 with a knowledge cutoff of 2025-06.
How much does Codestral cost, and who provides it?
Codestral is served by Mistral at $0.45/M blended tokens (3:1 input:output) — the 10th-cheapest of 42 current models. Full pricing breakdown: /llm-api-pricing/codestral.