Breaking

Arena Adds Mistral 3.5 to New Agent Mode for Real-World Task Showdowns

June 5, 2026 at 14:30 EDT

AI model comparison platform Arena.ai has added Mistral AI's latest model, "Mistral Medium 3.5," to its new "Agent Mode" feature. The effort pits frontier models against each other not on chat quality, but on their ability to complete real-world tasks using tools.

Apr 2026 · Arena × Mistral AI

Agents go to work: Arena adds "Agent Mode" and Mistral Medium 3.5

Arena's new Agent Mode puts models on real, multi-step tasks — spanning days, hundreds of turns, thousands of tool calls — and scores them on practicality. Mistral's 128B flagship folds chat, reasoning and code into one set of open weights.

77.6%

SWE-Bench Verified
(software engineering)

91.4%

τ³-Telecom
(multi-turn tool use)

256k

context window
128B dense weights

GPUs for local
self-hosting

Agentic benchmark scores

Mistral Medium 3.5 — both bars on a 0–100% scale

SWE-Bench Verified77.6%

τ³-Telecom91.4%

Said to surpass Devstral 2 and Qwen3.5 397B A17B on these agentic tasks.

How Agent Mode scores "practicality"

Web search→ Bash sandbox→ File ops→ Image gen

Sessions are scored across many axes:

Success / failure User praise / complaint Steerability Bash recovery Tool hallucination

Model at a glance

LicenseOpen (mod. MIT)

Input$1.50 / 1M tok

Output$7.50 / 1M tok

Reasoningper-request effort

Developers say

Coding and agentic skills are "genuinely strong" with "highly reliable tool calling" — driving full-stack builds, financial models and research-to-slides.

But critics note

"Too expensive for the performance" and "enterprise-leaning." Leaderboard data is still fresh — hallucination and recovery over long runs remain open challenges.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.