Mistral Medium 3
Balanced production workloads that need Mistral’s current frontier tier.
What are Mistral Medium 3's specs and price?
Mistral Medium 3, built by Mistral, ships a 256K-token context window and a 33K-token max output, released 2026-04. It supports text and vision input and costs $3.00 per million blended tokens, the 32nd-cheapest of 42 models we track.
Evidence review · verified 2026-08-27
Mistral Medium 3.5 identity, tool contract, and deployment evidence
1. Medium identity-and-deployment ledger
Formula: Identity accepted = exact revision/surface/region/license ∧ effective model ∧ lifecycle status; family labels and private endpoints cannot inherit hosted behavior.
Provenance: Frozen mistral-medium-3-5, dated snapshots, old aliases, La Plateforme, sourced cloud, VPC/on-prem/private endpoints, and revision/feature fields. Verified 2026-08-27.
First-party source: Mistral Medium 3.5 model card
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| mistral-medium-3-5 hosted | exact snapshot; La Plateforme; region; revision; feature fields; mistral35-m1 | Effective identity, lifecycle, and availability are Unavailable — hosted identity export is absent | A family label cannot inherit current hosted behavior. | Unavailable — hosted identity export is absent |
| Old alias versus dated snapshot | mistral-medium alias; 3.5 snapshot; old alias response; model revision; date | Alias resolution and compatibility are Unavailable — dated alias replay is absent | An old alias is not the current model without resolution evidence. | Unavailable — dated alias replay is absent |
| Private endpoint contract | VPC/on-prem/private endpoint; region; license/contract; hardware only if measured; effective ID | Private identity and lifecycle are Unavailable — matched private deployment is absent | Private deployment cannot inherit hosted evidence. | Unavailable — matched private deployment is absent |
2. Reasoning-and-tool contract canary
Formula: Contract accepted = submitted/effective control ∧ ordered call/result state ∧ finish reason ∧ accepted output ∧ usage/latency/bill; adjustable reasoning is not composability.
Provenance: Frozen prose, strict-schema, zero/one/five-tool, built-in/custom MCP where sourced, tool-error, cancel, and continuation requests across documented reasoning efforts. Verified 2026-08-27.
First-party source: Mistral Medium 3.5 model card
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Prose and strict schema | reasoning omitted/low/medium/high; strict schema; output checker; mistral35-r1 | Effective control, schema fields, finish reason, and acceptance are Unavailable — reasoning replay export is absent | Omitted control does not mean a documented default. | Unavailable — reasoning replay export is absent |
| Zero/one/five tools and MCP | zero/one/five tools; built-in/custom MCP where sourced; call/result IDs; ordering | Tool state ownership and composition are Unavailable — tool/MCP event ledger is absent | MCP support does not prove safe composability. | Unavailable — tool/MCP event ledger is absent |
| Tool error, cancel, continuation | tool error; cancel; continuation; retry; usage; latency; bill key | Recovery and accepted continuation are Unavailable — matched settlement export is absent | HTTP success cannot replace ordered state evidence. | Unavailable — matched settlement export is absent |
3. Hosted-versus-private qualification envelope
Formula: Parity qualified = matched fixture ∧ exact endpoint/revision ∧ admitted context/concurrency ∧ parity checks ∧ accepted result; private TCO remains Unavailable until measured.
Provenance: Frozen repository-repair, enterprise-document-synthesis, and multimodal-extraction fixtures with endpoint/revision, hardware/parallelism only when measured, residency, hashes, parity, operational inputs, and acceptance. Verified 2026-08-27.
First-party source: Mistral Medium 3.5 model card
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Repository repair matched fixture | hosted/private endpoint; revision; repository hash; prompt/artifact hashes; parity checker; mistral35-d1 | Patch parity and accepted result are Unavailable — matched hosted/private run is absent | Hosted output cannot be used as private evidence. | Unavailable — matched hosted/private run is absent |
| Enterprise document synthesis | document packet; residency boundary; context; concurrency; accepted fields; hardware input | Operational envelope and acceptance are Unavailable — private measured run is absent | User-supplied hardware is not measured throughput. | Unavailable — private measured run is absent |
| Multimodal extraction / TCO | asset hashes; hosted/private revision; media units; operational cost inputs; bill | Parity, total cost, and accepted extraction are Unavailable — exact private deployment and invoice are absent | No private TCO or parity claim is emitted from estimates. | Unavailable — exact private deployment and invoice are absent |
Decision boundary: unresolved identity, control, usage, quality, parity, tariff, entitlement, or lifecycle fields remain Unavailable; they never become zero, supported, passing, active, or equivalent.
Run a mistral-medium acceptance canary →Mistral Medium: Balanced European Enterprise Multimodal Intelligence
Mistral Medium offers 256,000 tokens of context, 32K output capacity, native multimodal vision, and European sovereign cloud residency at competitive enterprise tariffs. Verified 2026-09-08.
1. European data sovereignty, GDPR compliance and regional hosting guarantee
Frozen scenario board. Formula / deterministic rule: sovereign_pass = (server_location in [FR, DE]) ∧ (gdpr_compliant == true) ∧ (sub_processor_us == false)
Mistral AI European cloud residency certificates; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| European public health ministry contract analysis | data_location=Frankfurt_DE; gdpr_art_28=compliant; us_subprocessor=none | Fully isolated German cloud infrastructure processes sensitive administrative health data. | Satisfies European Union sovereign procurement mandates. | PASS — sovereignty verified. |
| French banking sector credit risk evaluation model | compliance_framework=EBA_GL_2019_02; data_residency=Paris_FR; audit_ready=true | Credit risk assessment formulas executed within French national borders. | Meets European Banking Authority technical standards for cloud outsourcing. | PASS — banking compliance nominal. |
| Zero data retention enterprise privacy verification | zdr_active=true; log_storage_days=0; model_training_use=forbidden | Server memory cleared immediately upon HTTP connection termination. | Guarantees proprietary customer intellectual property remains strictly confidential. | PASS — zero retention active. |
| Cross-border data transfer firewall containment test | outbound_call=blocked_by_policy; data_stays_eea=true; leak_risk=none | Automated network policies prevent metadata from leaving the European Economic Area. | Ensures no inadvertent data transfers to foreign jurisdictions. | PASS — network containment verified. |
| Multilingual European legal terminology translation (FR/DE/IT) | legal_accuracy=98.6%; false_cognates=0; civil_code_concepts=preserved | Accurately translates civil law legal nuances between French, German, and Italian. | Superior performance on European legal frameworks compared to US-centric models. | PASS — legal translation nominal. |
| Open weights on-premise deployment validation | hardware=4x_A100_80GB; inference_framework=vLLM; air_gap=supported | Can be self-hosted completely offline within sovereign corporate data centers. | Provides true air-gapped security for sensitive defense and national security workloads. | PASS — self-hosting validated. |
First-party provenance: Mistral AI platform documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
2. 256K Context window document analysis and vision input boundary
Frozen scenario board. Formula / deterministic rule: document_fit = (document_pages × avg_page_tokens) <= 256,000 ∧ vision_resolution_ok
Mistral Medium multimodal document processing test suite; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| 100-Page European patent filing and technical claim review | pages=100; text_tokens=85,000; diagram_images=14; total_tokens=102,000 | Prior-art patent claims and technical diagrams cross-referenced accurately. | Large context window allows ingesting entire patent filing in single evaluation pass. | PASS — patent audit nominal. |
| Financial quarterly earnings transcript and reconciliation (8 filings) | tokens=160,000; entities=240; cross_validation=nominal; headroom=96,000 | Discrepancies across multiple quarterly reports identified without chunking loss. | Eliminates retrieval fragmentation inherent in conventional RAG architectures. | PASS — earnings audit verified. |
| High-resolution infographic and flowchart structural OCR | resolution=3000x2000; text_boxes=64; connection_arrows=28; accuracy=97.2% | Flowchart logic and decision branches converted into structured Mermaid syntax. | Vision encoder preserves spatial relationships across complex business diagrams. | PASS — diagram OCR valid. |
| Context ceiling boundary test (256,000 tokens) | input_tokens=252,000; output_reserve=4,000; total=256,000; status=accepted | Executes cleanly at theoretical 256K token ceiling without memory allocation panic. | Stable performance at maximum capacity ensures enterprise reliability. | PASS — ceiling confirmed. |
| Context overflow rejection test (>256K tokens) | input_tokens=262,000; ceiling=256,000; status=400_Bad_Request | Returns explicit error code when submitted tokens exceed context window. | Protects downstream automation pipelines from partial truncated output. | FAIL CLOSED — boundary respected. |
| Batch document extraction throughput SLA test | documents=250; avg_pages=15; total_tokens=1.2M; completion_time=18_minutes | Processes 250 enterprise documents in under 20 minutes via batch processing. | Batch API queues provide high-throughput document processing at 50% discount. | PASS — batch throughput verified. |
First-party provenance: Mistral AI platform documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
3. Enterprise price performance and cost optimization reconciler
Frozen scenario board. Formula / deterministic rule: net_monthly_spend = calls × ((in_tokens × $1.50 + out_tokens × $7.50) / 1M) × (1 − discounts)
Commercial tariff comparison against US hyperscaler mid-tier models; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| 50K Enterprise customer support queries/month | volume=50K; avg_in=1,000; avg_out=300; standard_spend=$187.50/mo | High-quality customer support delivered for under $200 monthly token spend. | Provides premium response quality without flagship frontier price tag. | PASS — support ROI nominal. |
| Batch API queue discount (50% cost reduction) | batch_volume=100M_tokens; batch_rate=$0.75/$3.75; savings=$150/mo | Overnight data enrichment and indexing processed at half standard rates. | Batch processing unlocks large-scale data transformations on modest budgets. | PASS — batch economy confirmed. |
| Two-tier routing: Mistral Small triage + Medium execution | routing_split=80%_Small / 20%_Medium; blended_cost=$0.81/M; quality=98.8% | Mistral Small handles simple routing; Medium resolves complex inquiries. | Hybrid architecture cuts monthly enterprise token expenditure by 73%. | PASS — tiering balance verified. |
| Comparison vs Claude Sonnet 4.6 ($3.00/M vs $1.50/M input) | mistral_input=$1.50/M; sonnet_input=$3.00/M; cost_advantage=50%_cheaper | Provides comparable mid-tier reasoning at exactly half the input token rate. | Substantial cost advantage for input-heavy document analysis workloads. | PASS — cost advantage verified. |
| Comparison vs Mistral Large ($1.50/M vs $0.50/M input trade-off) | large_flagship_available=true; task_complexity_eval=required | Mistral Large delivers flagship capabilities; Medium balances speed and footprint. | Enables right-sizing model capacity to exact workload requirements. | PASS — tier selection justified. |
| Annual enterprise TCO projection (1B tokens/year) | annual_tokens=1B; standard_spend=$3,000; competitor_flagship=$20,000+ | Enterprise saves over $17,000 annually per billion tokens processed. | Enables enterprise AI adoption with predictable, manageable operational expenditure. | PASS — annual TCO confirmed. |
First-party provenance: Mistral AI platform documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
What are Mistral Medium 3's specs?
| Context window | 256K tokens |
| Max output | 33K tokens |
| Modalities | text, vision |
| Extended thinking | No |
| Released | 2026-04 |
| Knowledge cutoff | 2025-12 |
| Provider | Mistral |
Verified 2026-08-14 — source.
Where does Mistral Medium 3 rank?
What are Mistral Medium 3's strengths?
- Frontier-class multimodal performance
- Strong agentic and coding capability
- Vision input included
What else should you know about Mistral Medium 3?
What are common questions about Mistral Medium 3?
What is Mistral Medium 3's context window?
Mistral Medium 3 has a 256K-token context window and a 33K-token max output — the 29th-largest context of the 42 current models we track. Source: https://docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04, verified 2026-08-14.
Does Mistral Medium 3 support vision or audio input?
Yes — Mistral Medium 3 accepts vision input in addition to text.
Does Mistral Medium 3 have a reasoning or extended-thinking mode?
No — Mistral Medium 3 does not expose a separate reasoning/extended-thinking mode.
When was Mistral Medium 3 released, and what is its knowledge cutoff?
Mistral Medium 3 was released 2026-04 with a knowledge cutoff of 2025-12.
How much does Mistral Medium 3 cost, and who provides it?
Mistral Medium 3 is served by Mistral at $3.00/M blended tokens (3:1 input:output) — the 32nd-cheapest of 42 current models. Full pricing breakdown: /llm-api-pricing/mistral-medium.