Pricing Watch | Mistral Large 4's Open-Ended Sale Becomes the Permanent Price: $0.70/$2.10

🌏 中文版

Summary

When Mistral put flagship model Large 4 into public preview on 2026-10-06, the model card quoted a list price of $1.36 input / $4.18 output per 1M tokens, but the actual billing on Studio and OpenRouter ran a flat 50% off — $0.68 input / $2.09 output — with no end date published anywhere (see 10/10 Pricing Watch). Three weeks later, on 2026-10-29, Mistral's open-weight release announcement settled the question with one line: "Starting today, Mistral Large 4 is priced at $0.70 input / M, $0.07 cached input / M, and $2.10 output / M tokens." No "limited time," no "promo" label — this is now the standing rate. Run the math and the new price sits only 2-3% above the original sale, which means Mistral chose to keep the 50%-off rate permanently rather than let the "when does this bill double" risk that developers were quietly carrying actually land. The timing isn't a coincidence: the price change and the open-weight release landed on the same day. Once a model's weights are downloadable, the hosted API price has to track the marginal cost of self-hosting — a vendor can no longer lure developers in at a sale rate and then quietly walk it back up.

Before/After

The new permanent price is nearly identical to the sale price: input is only $0.02 higher (+2.9%), output only $0.01 higher (+0.5%), and cached input didn't move at all. The official announcement and model card now show only the new permanent price — the original $1.36/$4.18 list price has been removed from the docs. Mistral still hasn't published a separate batch rate for Large 4; third-party tracking shows no batch figure listed as of this writing.

Cost Simulation

Scenario: An agent handling 10,000 customer-support conversations per day, averaging 1,500 input tokens + 500 output tokens per conversation — the same scenario used in the 10/10 article, for direct comparison.

The number that matters is $630.00 vs. $619.50: a team that was already piloting on the sale price back in October sees its bill go up by just $10.50/month (+1.7%) under the new permanent rate — and no longer has to worry about a promo ending. Against the original list price, the $609/month saved (49.1%) is now a durable discount, not a sale that could be clawed back at any time.

Impact for Developers and Enterprises

Who benefits most

Teams the 10/10 article told to treat the sale price as POC-only can now write Large 4 into their production cost model directly — the new permanent price is within 1.7% of the sale rate they were already budgeting against, so there's no need to re-run the numbers. With weights now open (Mistral Ecosystem Open-Weight license, a modified MIT license), enterprises with data-sovereignty requirements — the announcement cites European bank La Banque Postale as a case — also get a "test via API, self-host once committed" path instead of having to commit before the model was even open.

Competitive landscape

Putting Large 4's old and new prices back into the same competitive table (USD per 1M tokens):

Large 4 jumped from the most expensive option in this price band to the second cheapest, trailing only its own predecessor, Large 3. Third-party pricing tracker TokenCost re-ran per-task cost using Artificial Analysis's Intelligence Index v5.0: Large 4 (medium reasoning) and GLM 5.3 (low) land on nearly the same score (23.1), but Large 4 costs 10% less. Push both to their top reasoning tiers and GLM 5.3 scores 5.7 points higher but costs $5.01 a task versus Large 4's high-mode $2.06. Step outside open-weight models, though, and the gap is still real: GPT-6.1 Terra scores 16 points higher for an eighth of the price, and Claude Haiku 5.5 beats it by 3.6 points at a lower cost. The price cut narrowed the gap with proprietary models — it didn't close it.

Action items

• If you were piloting on the sale price back in October: no action needed. The new permanent price is only 1.7% above what you were already paying, and you can write that number straight into your budget
• If you're weighing whether to self-host an open-weight model: check TokenCost's per-task cost figures first ($2.06 at high reasoning vs. GLM 5.3's $5.01) — Large 4 is now the cheapest option at matching scores, but its output-token usage runs much higher than proprietary models like GPT-6.1 Terra, so your actual bill depends heavily on your task's average output length
• If raw per-token price is what you're optimizing for: GPT-6.1 Terra ($1 input / $5 output) and Claude Haiku 5.5 still lead on per-task cost; Large 4's edge is self-hostable weights and data residency, not the lowest sticker price

Today's Takeaway

The 10/10 article treated the undated 50%-off sale as a red flag — developers were effectively absorbing the risk of "how long does this price hold" on the vendor's behalf. Three weeks later, the outcome ran the other way: instead of letting the sale quietly expire and doubling the bill, Mistral turned the sale price (with only a 2-3% bump) into the permanent list price, and announced it the same day it open-weighted the model. That points to a signal worth tracking that matters more than any single promo deadline: once a model's weights are about to go public and readers can work out the marginal cost of self-hosting themselves, the official API price has no choice but to track that line — otherwise the post-sale list price just pushes customers toward self-hosting or a competitor, and the vendor has no reason left to keep the business.

References

• Mistral: Mistral Large 4 open-weight release (official announcement, includes new permanent price)
• Mistral docs: Mistral Large 4 model card
• TokenCost: Mistral Large 4 Price Cut, $0.70/$2.10, Open Weights (includes pre-sale list price, cost simulation)
• This site's earlier coverage, 10/10: Mistral Large 4 launches at 50% off, no end date
• AI Business Weekly: OpenAI API Pricing 2026 (current GPT-6.1 Terra/Sol/Luna rates)