Some AI model API prices went up in 2026, and some went down. DeepSeek raised the peak price of its V4 Pro model about 3x for input and 4.6x for output in August 2026. Zhipu launched its newer models at higher prices, and Google set launch prices that double. OpenAI, Anthropic and Alibaba's Qwen priced their newest models below the ones they replaced. For an ecommerce brand of $5M to $200M, the change in the monthly bill is usually small. The risks that matter more are prices that move at short notice and where customer data is processed. This page compares the published prices and is compiled by Brand Media Works, an AI consulting and marketing data engineering firm for ecommerce brands.
Method note
Every price comes from the provider's own price page, read on 24 or 25 October 2026 and recorded in the Brand Media Works competitor register. Earlier prices come from copies of the same pages saved by the Internet Archive, dated in each table. Prices are list prices in US dollars per 1 million tokens. Input prices are for uncached input unless a row says otherwise, and they exclude batch, priority and promotional tiers unless noted. A newer model is not the same product as the one it replaced, so a change between models is a change in what the newest model costs, not a price rise on an existing contract. Multiples and monthly costs are our arithmetic. Press reports of wider price rises across Chinese providers could not be checked against an original source, so they are not used.
What happened to DeepSeek's API prices?
DeepSeek announced on 13 August 2026 that it was updating API prices and adding peak and off-peak rates. The new prices took effect at 16:00 UTC on 16 August. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, and off-peak prices are half the peak rate.
| Model and price | 14 July 2026 | 1 September 2026 (peak / off-peak) | 24 October 2026 (peak / off-peak) |
|---|---|---|---|
| V4 Pro, input | $0.435 | $1.32 / $0.66 | $1.32 / $0.66 |
| V4 Pro, output | $0.87 | $3.96 / $1.98 | $3.96 / $1.98 |
| V4 Pro, cached input | $0.003625 | $0.044 / $0.022 | $0.044 / $0.022 |
| Flash, input | $0.14 (V4 Flash) | $0.44 / $0.22 (V4 Flash) | $0.30 / $0.15 (V4.1 Flash) |
| Flash, output | $0.28 (V4 Flash) | $1.32 / $0.66 (V4 Flash) | $1.20 / $0.60 (V4.1 Flash) |
At peak, V4 Pro now costs 3.0 times its July price for input, 4.6 times for output and 12 times for cached input. Off-peak, the rise is 52% for input and 128% for output (our arithmetic). The July page shows a single flat price with no discount noted. DeepSeek cut Flash prices on 10 September, when it replaced V4 Flash with V4.1 Flash, but Flash output still costs more than in July at any hour. The same 10 September post said V4 Pro requests would be served by V4.1 Flash at Flash rates from 14 September. DeepSeek's price page on 24 October still listed V4 Pro at the prices above, so this page follows the price page. The price page also carries a standing warning:
Product prices may vary and DeepSeek reserves the right to adjust them.
Did other Chinese AI providers raise prices?
Not in the same way. The pattern differs by provider.
| Provider | Older or earlier model | Newest model | What changed |
|---|---|---|---|
| Zhipu (GLM) | GLM-4.7, newest model on 5 January 2026: $0.60 / $2.20, unchanged in October | GLM-5.3: $1.40 / $4.40 | New flagship priced at 2.3 times the input and 2 times the output price of January's newest model (our arithmetic) |
| Alibaba (Qwen), international | qwen3.7-max: $2.50 / $7.50 | qwen3.8-max: $2.00 / $6.00 | Newest model 20% cheaper; in Alibaba's mainland China region both cost $1.65 / $4.951 |
| MiniMax | No earlier copy read | MiniMax-M3: list $0.60 / $2.40, shown as $0.30 / $1.20 under a "Permanent 50% off" | A discount, not a rise |
| Moonshot (Kimi), on Amazon Bedrock | Kimi K2 Thinking, US regions: $0.60 / $2.50 | Kimi 3, global cross-region: $3.00 / $15.00 | Newest model at 5 times the input and 6 times the output price of the older one (our arithmetic); Moonshot's own price table could not be read |
So the claim that Chinese AI models have become much more expensive holds for DeepSeek's existing model and for the newest Zhipu and Kimi models. It does not hold for Qwen or MiniMax on their current price pages.
What happened to US model prices?
Mostly the opposite, with one exception.
| Provider | Earlier or older model | Newest model | What changed |
|---|---|---|---|
| OpenAI | gpt-5.4, flagship on 15 April 2026: $2.50 / $15.00; gpt-5.4-pro: $30 / $180 | gpt-6.1-sol: $2.00 / $10.00; gpt-6-astra: $10 / $50 | Lower at both levels |
| Anthropic | Claude Opus 5: $5 / $25; Claude Opus 4.1: $15 / $75 | Claude Opus 5.5: $4 / $20 | Lower |
| Gemini 3 Flash Preview, newest Flash on 14 April 2026: $0.50 / $3.00 | Gemini 3.8 Flash: $0.75 / $3.75 to 31 December 2026, then $1.50 / $7.50; Gemini 4 Argon: $2 / $10 to 30 November 2026, then $4 / $20 | Launch prices that double on a set date |
Price per token is not the whole story. Anthropic notes that its models from Claude 4.7 onward use a tokenizer that produces about 30% more tokens for the same text, and providers count tokens differently. The fair comparison is the cost of a real workload, measured on the brand's own data.
What does this mean in dollars for an ecommerce brand?
Less than the percentages suggest. Take a support team that uses AI to draft replies to 20,000 tickets a month, with about 3,000 tokens of order history and policy in and 500 tokens out per ticket. That is 60 million input and 10 million output tokens a month.
| Model and price | Monthly cost |
|---|---|
| DeepSeek V4 Pro, July 2026 | $34.80 |
| DeepSeek V4 Pro, now, peak / off-peak | $118.80 / $59.40 |
| DeepSeek V4.1 Flash, now, peak | $30.00 |
| DeepSeek v3.2 on Amazon Bedrock, US regions | $55.70 |
| Gemini 3.8 Flash, now / from January 2027 | $82.50 / $165.00 |
| Zhipu GLM-5.3 | $128.00 |
| Qwen qwen3.8-max, international | $180.00 |
| OpenAI gpt-6.1-sol | $220.00 |
The largest rise on the list, DeepSeek V4 Pro from July to today's peak price, adds $84 a month, or about $1,008 a year (our arithmetic). That is small next to the cost of building, checking and running the workflow. What a private LLM costs to run, beyond the model sets out those other costs. For most brands the API price matters less than three other questions: how quickly the price can change, whether the model can be swapped, and where customer data is sent.
What should a brand do about AI price changes?
- Price the workload, not the token. Run a month of real tickets, product records or reports through two or three models and compare the bills, not the rate cards.
- Budget for a change. DeepSeek's prices moved about 3x on three days' notice, and Google's launch prices have end dates. A budget that survives a doubling of the model line is safer than one built on today's rate.
- Keep the model swappable. Build the workflow so the model is a setting, not a rewrite, and keep the test set used to check quality, so a switch can be checked in a day.
- Move batch work to cheaper hours or tiers. DeepSeek's off-peak rate is half its peak rate, and OpenAI's and Amazon Bedrock's batch tiers are half price. Overnight reporting and catalog jobs rarely need peak speed.
- Check where the data goes. DeepSeek's privacy policy, which covers its apps, websites and related services, says it collects, processes and stores personal data in the People's Republic of China. Its models are published as open weights, so the same model can run in a brand's own cloud account instead. Amazon Bedrock lists DeepSeek v3.2 in US regions at $0.62 / $1.85 per 1M tokens.
Amazon Bedrock vs a self-hosted LLM compares the two ways to run an open model privately, and how to deploy an LLM so customer data never leaves your account covers the setup. Private LLM describes how Brand Media Works deploys models inside a brand's own cloud, with the model kept swappable.
Frequently asked questions
Did DeepSeek raise its API prices?
Yes. From 16 August 2026, DeepSeek V4 Pro costs $1.32 per 1M input tokens and $3.96 per 1M output tokens at peak hours, against a flat $0.435 and $0.87 in July. Off-peak prices are half the peak rate. Flash prices were cut back on 10 September but remain above July's level for output.
Are AI API prices going up or down in 2026?
Both. DeepSeek raised prices on an existing model, Zhipu priced its newer models higher, Kimi's newest model lists higher on Amazon Bedrock, and Google's new Gemini prices double on set dates. OpenAI, Anthropic and Qwen priced their newest models below the ones they replaced.
Is it safe to send customer data to DeepSeek's API?
That is a decision for each brand's legal and security team. DeepSeek's privacy policy, which covers its apps, websites and related services, says it collects, processes and stores personal data in the People's Republic of China. Because DeepSeek publishes its model weights, a brand that wants the model without sending data to DeepSeek can run it in its own cloud account.
Does an API price rise matter for a small brand?
Usually not much in dollars. On a typical support workload the biggest 2026 rise adds about $84 a month. It matters more as a sign that prices can move at short notice, which is a reason to keep the model swappable.