Updated
Mistral Deprecates Medium 3.5 and Large 3: What Moving to Large 4 Costs
Short answer. Mistral has deprecated Mistral Medium 3.5 and Mistral Large 3. Both model pages, read on November 6, 2026, say “Deprecation date 10/31/2026” and “Replacement Mistral Large 4”, and neither model is on Mistral’s pricing page any more. Large 4 costs $0.70 in and $2.10 out: cheaper than Medium 3.5 ($1.50, $7.50), 1.40 times dearer than Large 3 ($0.50, $1.50).
In plain words. Two of Mistral’s models are on their way out. Mistral points both to Large 4, now generally available. From Medium 3.5, Large 4 costs less than half. From Large 3, it costs 40 percent more. The pages you see in a browser print no switch-off date.
Update, November 16, 2026: Mistral now dates Mistral Large 4’s general availability. Its model page, read on November 16, 2026, reads “November 5, 2026 GA Mistral Ecosystem Open Weight License v 26.10”, and its Markdown version “Release date: 2026-11-05”. Where this post says Mistral does not say when the stage changed, Mistral now gives November 5, 2026. The price is unchanged, $0.70 / $0.07 / $2.10.
What the model pages say
Mistral announces a deprecation on the model’s own page, not on its pricing page. Read on November 6, 2026, Mistral Medium 3.5’s page says “This model is deprecated. Use Mistral Large 4 for new integrations.”, “Deprecation date 10/31/2026” and “Replacement Mistral Large 4”; its header line reads “April 28, 2026 Blog Deprecated Modified MIT v 26.04”. Mistral Large 3’s page says the same three lines, with the header “December 2, 2025 Blog Deprecated Apache 2.0 v 25.12”. Both pages were read with the same words on November 3, 2026, the day AI Signal recorded the two deprecations; they are listed on the deprecations page.
The two models have left Mistral’s pricing page. Its table “Model | Input | Cached input | Output |”, read on November 6, 2026, prints “Mistral Large 4 ↗ | $0.7 | $0.07 | $2.1 |” and no row for Medium 3.5 or Large 3; the weekly check of November 2 was the first to find neither, and the last reading that priced them, on October 26, 2026, printed $1.50 / $0.15 / $7.50 for Medium 3.5 and $0.50 / $0.05 / $1.50 for Large 3. Those are the figures used below.
The prices, and what one request costs
| Per million tokens | Mistral Medium 3.5 | Mistral Large 3 | Mistral Large 4 |
|---|---|---|---|
| Input | $1.50 | $0.50 | $0.70 |
| Cached input | $0.15 | $0.05 | $0.07 |
| Output | $7.50 | $1.50 | $2.10 |
| Request of 10,000 in + 1,000 out | $0.0225 | $0.0065 | $0.0091 |
| Context window | 256K tokens | 256K tokens | 1M tokens |
| Status on the model page, Nov 6, 2026 | Deprecated | Deprecated | GA |
The sample request, with no cache: on Medium 3.5, 10,000 × $1.50 ÷ 1,000,000 = $0.015 plus 1,000 × $7.50 ÷ 1,000,000 = $0.0075, so $0.0225. On Large 3, $0.005 plus $0.0015, so $0.0065. On Large 4, $0.007 plus $0.0021, so $0.0091.
The asymmetry: cheaper from one, dearer from the other
Mistral names one replacement for two models at different prices, so the move costs the opposite thing depending on where you start.
| Moving to Large 4 from | Input | Cached input | Output | Sample request |
|---|---|---|---|---|
| Medium 3.5 | 1.50 ÷ 0.70 = 2.14×, cheaper | 0.15 ÷ 0.07 = 2.14×, cheaper | 7.50 ÷ 2.10 = 3.57×, cheaper | 0.0225 ÷ 0.0091 = 2.47×, cheaper |
| Large 3 | 0.70 ÷ 0.50 = 1.40×, dearer | 0.07 ÷ 0.05 = 1.40×, dearer | 2.10 ÷ 1.50 = 1.40×, dearer | 0.0091 ÷ 0.0065 = 1.40×, dearer |
From Medium 3.5, every figure falls, output most of all. From Large 3, every figure rises by the same 1.40, so the sample request rises by 1.40 too, whatever the mix of input and output. Large 4 also reads a longer prompt: the Markdown version of its page says “Context length: 1M” against the 256K of the two models it replaces.
Large 4: now generally available, and a price Mistral does not date
Large 4’s status changed this week. On October 30, 2026 Mistral Large 4’s page read “October 6, 2026 Public Preview Mistral Ecosystem Open Weight License v 26.10” and its Markdown version “Status: Public Preview”; AI Signal’s check of November 2 still found that line. On November 6, 2026 the page reads “October 6, 2026 GA Mistral Ecosystem Open Weight License v 26.10” and the Markdown version “Status: GA”. Mistral does not say when it changed: the change was first seen by this site on November 6, 2026, and the changelog has no entry about it. Mistral’s lifecycle page allows either stage for a replacement: “Deprecation is announced as soon as a replacement model is available. The replacement can be in General Availability or Public Preview.”
The price is the one Large 4 has had since October 30, when it fell from $1.36 / $0.14 / $4.18 to $0.70 / $0.07 / $2.10 (the post of that day). Mistral has not said whether that price is temporary: its changelog, read on November 6, 2026, still prints under October 6 “Launch pricing: 50% off for 2 weeks.” and has no later entry about Mistral Large 4.
When the two models stop answering
The pages as shown in a browser print no retirement date for Medium 3.5 or Large 3. The Markdown version of each page prints “Retirement date: 2027-04-30”, a line the visible page does not; AI Signal records what the visible page says and does not take that line as a published date.
What Mistral’s lifecycle page says about the period in between: a deprecated model is “Retirement announced. Still accessible until the retirement date.”, and “During the deprecation period, the model remains accessible. Once retired, requests to its identifiers fail with a 404 error.” The notice it promises depends on the stage: “General Availability | 6 months |”, “Public Preview | 1 month |”, “Third-party models | 1 month |”. Both deprecated models were GA.
The alias is worth a look. Large 3’s Markdown page still lists “API
names: mistral-large-2512, mistral-large-latest” on November 6,
2026; Large 4’s lists mistral-large-4 and mistral-large-4-0. A
request to mistral-large-latest therefore still goes to the deprecated
Large 3, at $0.50 and $1.50.
GLM 5.2 on Mistral, retired on October 31
The same deprecation page carried a third-party model to its end. Mistral’s changelog says: “Z.ai GLM 5.2 ( zai-glm-5-2 ) is deprecated and retires on October 31, 2026. Use Z.ai GLM 5.3 ( zai-glm-5-3 ) instead, at the same price.” The GLM 5.2 page, read on November 6, says “This model is deprecated. Use Z.ai GLM 5.3 for new integrations.”, “Deprecation date 9/29/2026” and “Replacement Z.ai GLM 5.3”. GLM 5.3 is on the pricing page at “$1.4 | $0.14 | $4.4”, the same figures GLM 5.2 had. The models Mistral serves for other makers are listed apart on the Mistral API pricing page.
How we checked
- Model pages: docs.mistral.ai/models/mistral-medium-3-5-26-04, mistral-large-3-25-12, mistral-large-4-0 and zai-glm-5-2, read by script on November 6, 2026, each with its Markdown version (the same address with “.md”). The visible pages give the deprecation and the replacement; only the Markdown versions print a retirement date, which is reported as such and not recorded.
- Prices: docs.mistral.ai/inference/pricing, read on November 6, 2026, for Large 4 and GLM 5.3; for Medium 3.5 and Large 3, AI Signal’s last reading of the same table with their rows, October 26, 2026.
- Large 4’s stage: the model page and its Markdown version on October 30 (Public Preview) and November 6, 2026 (GA); the weekly check of November 2 still found the Public Preview line.
- Changelog and lifecycle: docs.mistral.ai/getting-started/changelog and docs.mistral.ai/inference/model-lifecycle, read on November 6, 2026.
- Arithmetic: the sample request and every ratio are computed here from the prices quoted, and rounded only in the text.
Frequently asked questions
Is Mistral Medium 3.5 being retired?
Deprecated, not retired: its model page, read on November 6, 2026, prints 'Deprecation date 10/31/2026' and 'Replacement Mistral Large 4', and no retirement date. Mistral's lifecycle page says a deprecated model is 'Still accessible until the retirement date'.
What replaces Mistral Large 3?
Mistral Large 4, according to Large 3's model page read on November 6, 2026 ('Replacement Mistral Large 4'). Large 4 costs $0.70 per million input tokens, $0.07 for cached input and $2.10 for output on Mistral's pricing page, 1.40 times Large 3's $0.50, $0.05 and $1.50.
Is Mistral Large 4 cheaper than Medium 3.5?
Yes, on Mistral's pricing page as read on November 6, 2026: $0.70 against $1.50 per million input tokens (2.14 times cheaper), $0.07 against $0.15 for cached input (2.14 times) and $2.10 against $7.50 for output (3.57 times). A request of 10,000 input and 1,000 output tokens costs $0.0091 on Large 4 and $0.0225 on Medium 3.5.
When will Mistral Large 3 stop working?
Mistral's page for Large 3 prints no retirement date as shown in a browser; only its Markdown version prints 'Retirement date: 2027-04-30', which AI Signal does not take as a published date. Mistral's lifecycle page says 'During the deprecation period, the model remains accessible' and that, once retired, requests fail with a 404 error.
Drafted with AI assistance from the cited source and reviewed by Álvaro Lucero Rufino, who is responsible for what stays published here. Automated publishing was discontinued in September 2026; posts that no longer met that standard were removed rather than left up. This is original explanatory writing, not a reproduction — read the source above for the full story.