Magistral Medium 1.2

“It made strategic sense in late 2025 as our reasoning answer to o-series and R1; by mid-2026 Medium 3.5's reasoning knob has weakened the case for a separate endpoint.”

At launch Magistral Medium 1.2 was the obvious dedicated-reasoning bet — strong math, visible traces, multilingual reasoning. The strategic picture shifted when Mistral's own Medium 3.5 added a reasoning_effort knob in one open-weight SKU. I routed the hardest math and PhD-class science to Magistral, the most reasoning-pure model in the lineup, until Mistral deprecated it on 2026-05-22 and retired it on 2026-07-31, naming Medium 3.5 as the replacement; for general agent loops I had already defaulted to Medium 3.5. Closed weights at the Medium tier (vs the open Magistral Small) is a strike. The auditable-trace angle kept it relevant for compliance-driven reasoning where you must show the work.

“A credible reasoning model whose distinct niche Mistral itself collapsed — its clearest remaining moat is auditable, in-language reasoning.”

Magistral's positioning as the European reasoning specialist was sound, but the strategic problem is internal cannibalisation: Medium 3.5 absorbed the reasoning use case into a cheaper-to-reason, open-weight generalist. Against DeepSeek R1 and Claude/GPT thinking it competes on multilingual traces and price-per-thought, not on top-end reasoning (AA Index 27 in May 2026, before the index was rescaled). The durable differentiators are the visible, in-user-language reasoning trace (valuable for regulated, multilingual auditability) and the open Magistral Small baseline. As a standalone product its market timing passed, and Mistral retired it on 2026-07-31 in favor of Medium 3.5.

“$2/$5 is reasonable per token, but long traces inflate output 3-5x — a single hard question can cost 20K tokens, so predictability is poor.”

The per-token rate ($2.00/$5.00) is cheaper than Claude Opus thinking, but reasoning traces inflate output counts so real bills run 3-5x what an equivalent non-reasoning call would cost — and a single hard math question can emit 20K+ tokens. The ~50% batch discount helps for async work. The honest financial read: poor predictability makes this the wrong tool for high-volume use and the right tool for selective hard-problem routing where the extra cost buys a correct answer. Cache discipline matters. For volume reasoning, Small 4's reasoning mode or Medium 3.5 is cheaper.

“The visible reasoning trace is debugger-friendly — I can see what it's thinking and tune prompts — and in-language traces are a real win for EU teams.”

For a builder the traces are the feature: I can watch the model reason and adjust prompts accordingly, and vision-input reasoning genuinely helps with chart/diagram debugging. The long output ceiling lets it fully work a problem. Multilingual traces are useful when collaborating with EU-language teammates. Negatives: latency is meaningfully higher than non-reasoning models, output bills add up fast, and it occasionally over-thinks simple questions. For most coding work I now reach for Medium 3.5 at high reasoning effort, which feels more pragmatic — but for pure reasoning observability, Magistral was the better instrument until its 2026-07-31 retirement.

“Noticeably smarter on hard questions, and the visible thinking trace is confidence-building — worth the 5-30 second wait when I want a careful answer.”

Magistral feels markedly smarter on hard questions — math, science, multi-step reasoning — than the chat-tier Mistrals. The visible thinking trace is novel and confidence-building for users who want to see the work. Latency is the felt cost: questions take 5-30 seconds before the final answer streams. For everyday chat it is overkill; for the moments I really want a careful answer, the wait is worth it. Multilingual reasoning (traces in my own language) is a unique strength I haven't found elsewhere.

“AIME 91.82% leads the press release; the May 2026 AA Index of 27 against a 36 median is the line they didn't print. Math-strong, broadly mid-tier.”

Magistral published real numbers, which is creditable, and the AIME math results are genuinely strong. But the framing cherry-picks: AIME and GPQA headline, while the broad reasoning aggregate (Artificial Analysis Index 27 in May 2026, flagged "below average" before AA rescaled the index) and the slow 39 tps go unmentioned. "Frontier-class reasoning" is generous — it is frontier on competition math and mid-tier on the wider reasoning surface, behind DeepSeek R1 and Claude/GPT thinking. Add the trace-inflated, unpredictable cost and Medium 3.5's encroachment, and the honest claim is "strong, auditable math-reasoning specialist," not a frontier reasoner. Hard math was the thing to route here until the 2026-07-31 retirement; don't expect across-the-board reasoning leadership.