Loading page
Mistral
Mistral Medium 3.5
Mistral Medium 3.5 is a mid-tier large language model from Mistral with a 262,144-token context window and measured output throughput of roughly 143 tokens per second.
Context 262K
Input from
$0.750 / 1M tokens
across 3 providers
API Pricing
Cheapest on OpenRouter — 45% below avg| Provider | Input / 1M | Output / 1M |
|---|---|---|
| $0.750 | $3.75 | |
| $1.50 | $7.50 | |
| $1.50 | $7.50 | |
| $1.74 | $8.69 |
Prices updated daily. Last check: Sep 4, 2026
Mistral Medium 3.5 pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Intelligence
30.4 / 100
Coding
46.9 / 100
Output Speed
122 t/s
Latency (TTFT)
652ms
Reasoning & Knowledge
- GPQA Diamond74.8%
- Humanity's Last Exam13.8%
Coding
- SciCode39.6%
Agentic & Tool Use
- Terminal-Bench Hard33.3%
- Terminal-Bench v2.150.6%
- τ²-bench94.2%
- τ-bench Banking15.1%
Instruction & Long Context
- IFBench68.8%
- Long-Context Reasoning65.3%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- Mistral
- Context Window
- 262K
- Modalities
- Text
Capabilities
- Tool Calling
- No
- Open Source
- No
Strengths & Limitations
Strengths
- 262,144-token (256K) context window, enough for long documents, large transcripts, or multi-file code context in one request
- Measured output speed of about 143 tokens per second (Artificial Analysis), suitable for streaming interactive responses
- Time to first token of roughly 671 ms in third-party testing, keeping perceived latency low for chat-style use
- Mid-tier positioning in Mistral's lineup, targeting a middle ground between the company's small and large models
- Available from multiple inference providers, so buyers can compare per-token rates and latency rather than being tied to a single host
- Iteration on Mistral Medium 3, giving teams already using that model a same-family upgrade path
Limitations
- Mid-tier model — Mistral's larger models are generally expected to score higher on the hardest reasoning and coding tasks
- We do not track task benchmark scores (e.g. coding or reasoning evals) for this model, so quality comparisons must come from vendor or third-party sources
- Multimodal support, tool-calling behavior, and licensing terms are not recorded in our metadata; verify against Mistral's documentation before designing around them
- Throughput and latency figures come from a single third-party measurement and will vary by provider, region, and request size
- Very long 256K-token prompts increase per-request cost and latency regardless of the headline throughput number
Key Features
•262,144-token context window
•Approximately 143 output tokens per second measured throughput
•~671 ms time to first token in third-party benchmarking
•Mid-tier member of the Mistral model family
•Successor iteration to Mistral Medium 3
•Served through multiple inference API providers
•Long-context document and codebase ingestion in a single request
About Mistral Medium 3.5
Mistral Medium 3.5 is a large language model developed by Mistral, the French AI company behind the Mistral and Mixtral model lines. As the name indicates, it sits in the "Medium" tier of Mistral's lineup — positioned between the company's smaller, cheaper models and its larger top-end offerings — and the 3.5 designation marks it as an iteration on the earlier Mistral Medium 3 release.
The model supports a context window of 262,144 tokens (256K), which is large enough to hold long documents, extended chat histories, or sizeable code repositories in a single request. Independent throughput measurements from Artificial Analysis put output generation at approximately 143 tokens per second with a time to first token of around 671 ms, which places it in the range of a responsive model suitable for interactive use rather than one optimized purely for batch throughput. Capabilities beyond these figures — such as vision input, tool calling behavior, or licensing terms — are not tracked in our database for this entry, so consult Mistral's own documentation before relying on them.
In practice, mid-tier models like Mistral Medium 3.5 are used where a small model is not accurate enough but a large frontier model is more capacity than the workload requires: document analysis, retrieval-augmented generation over long sources, chat assistants, drafting and summarization, and coding assistance. Because it is served through multiple API providers, per-token pricing and latency can differ meaningfully between hosts — see the pricing table on this page for current per-provider figures.
Common Use Cases
Mistral Medium 3.5 fits workloads that need solid general-purpose language ability with room for long inputs, without paying for a top-tier model. The 256K context window makes it a candidate for long-document summarization, contract and report analysis, retrieval-augmented generation where many retrieved chunks are passed in at once, and multi-turn assistants that accumulate long histories. Its measured ~143 tokens/second output rate and sub-second time to first token suit user-facing chat, drafting, and coding-assistant scenarios where streaming responsiveness matters. For very high-volume, simple classification or extraction work, a smaller Mistral model is usually the more economical choice; for the most demanding reasoning, math, or agentic tasks, a larger model is the more common pick.
Frequently Asked Questions
How much does Mistral Medium 3.5 cost to run?
Pricing depends on which inference provider you use and on the pricing model — separate input and output token rates, batch or cached-input discounts, and committed-capacity or self-hosted GPU options all price differently. Rates also change frequently. Check the pricing table on this page for the current per-provider figures rather than relying on a fixed number.
What is Mistral Medium 3.5 best used for?
It suits general-purpose text work where long inputs are involved: long-document summarization and analysis, retrieval-augmented generation over large source sets, chat assistants with extended histories, drafting, and coding assistance. Its ~143 tokens/second output rate and roughly 671 ms time to first token make it workable for interactive, streaming applications.
How large is the context window?
Mistral Medium 3.5 supports 262,144 tokens (256K) of context, which can accommodate book-length documents, long meeting transcripts, or substantial portions of a codebase in a single request.
How does it differ from Mistral Medium 3?
Mistral Medium 3.5 is a later iteration in the same Medium tier of Mistral's lineup. Our database tracks its 256K context window and throughput measurements; for a detailed capability-by-capability comparison against Mistral Medium 3, refer to Mistral's own release notes.
Should I choose Mistral Medium 3.5 or a larger Mistral model?
Medium-tier models are typically selected when a small model is not accurate enough but the workload does not justify a larger model's cost and latency. If your tasks involve hard multi-step reasoning or complex agentic workflows, evaluate a larger model; if they are high-volume and simple, a smaller Mistral model is usually cheaper. Running your own evaluation on representative prompts is the reliable way to decide.