LLM API Pricing Climbs as Alibaba Lifts DeepSeek Rates
Alibaba quietly raised the list price of a popular budget model, according to a price tracker run by Stork AI. Its report dated October 20, 2026 shows DeepSeek V4 Flash 0731 output pricing moving from $0.40 to $1.32 per million tokens. Input pricing more than doubled as well.
If you build products on top of AI models, this matters this week. LLM API pricing has mostly fallen over the past two years, and many budgets still assume that trend will continue. One quiet change can erase the margin on a feature you already sold.
What Stork’s Price Tracker Recorded
Stork says its catalog checks first-party list prices every day. In the October 20 report, the tracker flagged 17 separate price movements across providers. The DeepSeek change drew the headline, but it was not alone.
| Model | Reported change |
|---|---|
| DeepSeek V4 Flash 0731, input | $0.20 to $0.44 per million tokens |
| DeepSeek V4 Flash 0731, output | $0.40 to $1.32 per million tokens |
| Mistral Ministral 3B | Up 150% |
| Mistral Ministral 8B | Up 50% |
| Alibaba Qwen3.6 35B-A3B, output | Up 52% |
| Zhipu GLM-5V-Turbo | Down 85% |
Taken together, the movements run in both directions, which makes budgeting harder than a simple upward trend would. The only large cut in the list is Zhipu’s, so most founders should expect the overall bill to rise.
There is an important caveat. The numbers come from Stork’s own tracker, and the report names no platform, no effective date, and no explanation from Alibaba. Treat them as reported observations until you confirm them on your own invoice or the vendor’s pricing page.
Why a Cheap Model Can Get Expensive Fast
Low per-token prices hide how quickly volume adds up. Picture an illustrative workload of 10 million input tokens and 2 million output tokens each month. At the old DeepSeek rates that costs about $2.80, and at the new rates it costs about $7.04, which is roughly 2.5 times as much.
Those are small dollars, but the ratio is what counts. A customer-facing feature that handles millions of requests turns the same ratio into real money. Meanwhile, your own customer prices probably did not change overnight, so the squeeze lands on your margin.
The reverse happened not long ago. Our coverage of AI model pricing showed one vendor cutting costs by about 75 percent. Stork’s author reads the latest moves as labs testing their pricing power, which is an opinion, but it is a useful warning to plan for both directions.
A Five-Minute Budget Stress Test
Start with your last monthly invoice for each model. Multiply your input and output token counts by double the current rate, then check whether the feature still earns a profit. If it does not, you found the risk before a vendor found it for you.
- Track tokens per customer, not only total spend.
- Set a hard monthly cap and an alert at 70 percent of it.
- Record which features depend on a single model.
Next, compare the result with your pricing page. If a feature now costs more to run than a customer pays for it, you have three options: raise the price, trim usage, or switch models. Each choice has a cost, so make it deliberately rather than after a surprise invoice.
In addition, ask whether a smaller model could handle part of the work. Teams that already practice cloud cost management will recognize the routine: tag the spend, review it weekly, and give it an owner.
Spreading Vendor Risk Without Slowing Down
The sensible answer is not to flee a vendor after one increase. Instead, keep your prompts and evaluations portable so a switch takes days rather than months. For example, save a test set of real customer questions and score any new model against it. Founders who read our note on Gemini API pricing will remember that intro rates can expire too.
Also read the vendor’s own rate card rather than relying on a third party. DeepSeek publishes official pricing documentation, and checking it monthly takes minutes. Zhipu’s reported 85 percent cut shows that prices can fall as fast as they rise, so re-shop on a schedule.
Questions Founders Ask About LLM API Pricing
Will AI model prices keep rising?
Nobody knows. Stork predicts more labs will raise prices by the end of the fourth quarter, but that is a forecast and not a fact. Plan for both outcomes.
How do I protect my margins?
Cap your spend, track cost per customer, and keep a second model tested and ready. Then revisit your own pricing if usage costs stay higher.
Should I lock in a contract?
Ask your vendor about committed-use discounts or fixed terms if your volume is steady. Fixed pricing trades flexibility for certainty, which is useful when a feature is central to your product.
Here is how to think about it: treat every model price as a variable, not a constant. The founders who model that variable now will not be surprised by the next invoice.
