AI API prices are still falling for most models, but several list prices rose in 2026, and some of those increases hit widely used models. DeepSeek V4.1 Flash, DeepSeek V4 Pro 0813 and Grok 4.7 all list above their first-seen prices. Per task, the effect is small for a chat answer and larger for a long document, so the list price is only part of the bill.
What "price increase" means in this data
A list price is the rate a provider charges per 1M tokens. It is not what a task costs. Cost also depends on input tokens, output tokens and reasoning tokens. Reasoning tokens are billed as output tokens, so a model with a low list price can still cost more per task if it thinks for a long time.
I track two kinds of change:
- First-seen vs latest. The price a model listed at when it first appeared in OpenRouter's weekly snapshots (88 snapshots since 2024-03-26), compared with the price on 2026-10-11.
- Recent changes. List-price moves on OpenRouter, logged by date.
The first shows where a model started. The second shows how often prices move.
Three models that went up
DeepSeek V4.1 Flash
It first appeared on 2026-09-14 at $0.15 input and $0.6 output per 1M tokens. Its latest list price is $0.3 input and $1.2 output, so both rates doubled in under a month.
The path was not straight. On 2026-10-05 the logged input price went from $0.30 to $0.003 while output rose from $1.20 to $2.40 (+100%). On 2026-10-11 input went back to $0.30 and output fell to $1.20 (-50%).
Per task, I measured this model. A chat answer cost 0.76¢ ($7.60 per 1,000 tasks) and a 20-page PDF summary cost 0.63¢ ($6.28 per 1,000). The chat run used 5,461 reasoning tokens against 859 output tokens, so reasoning drives that bill. Because the output rate doubled, the output-side cost of that run roughly doubled too.
DeepSeek V4 Pro 0813
It first appeared on 2026-08-14 at $0.435 input and $0.87 output. Its latest list price is $0.66 input and $1.98 output, so both rates are higher than at launch.
The log shows the swing. On 2026-10-05 it moved from $0.39 to $0.85 input and from $4.20 to $5.00 output (+19%). On 2026-10-11 it dropped to $0.66 input and $1.98 output (-60%). Net of launch, both rates are still up.
Per task, this model is estimated, not measured, so treat figures as approximate. A chat answer is 1.3¢ ($12.56 per 1,000) and a PDF summary is 1.3¢ ($12.50 per 1,000). Both sit above the medians across all 326 models: 0.24¢ for chat and 0.83¢ for the PDF.
Grok 4.7
It first appeared on 2026-09-21 at $1.6 input and $4.8 output. The latest list price is $2 input and $6 output. It has a 500,000-token context window and an LMArena Text score of 1443.
I measured it on both tasks. A chat answer cost 0.96¢ ($9.58 per 1,000). A 20-page PDF cost 3.4¢ ($34.26 per 1,000). The PDF job is larger because the model reads 13,954 input tokens, and the document dominates the cost.
| Model | First seen (in / out per 1M) | Latest (in / out per 1M) | Chat answer | 20-page PDF |
|---|---|---|---|---|
| DeepSeek V4.1 Flash | $0.15 / $0.6 (2026-09-14) | $0.3 / $1.2 | 0.76¢ (measured) | 0.63¢ (measured) |
| DeepSeek V4 Pro 0813 | $0.435 / $0.87 (2026-08-14) | $0.66 / $1.98 | 1.3¢ (estimated) | 1.3¢ (estimated) |
| Grok 4.7 | $1.6 / $4.8 (2026-09-21) | $2 / $6 | 0.96¢ (measured) | 3.4¢ (measured) |
Median across all 326 models: 0.24¢ per chat answer, 0.83¢ per PDF summary.
Other increases in the recent log
The 2026-10-11 log has more examples:
- DeepSeek V3.2 output rose from $0.42 to $0.80 per 1M (+90%).
- Qwen3 VL 8B Instruct output rose from $0.46 to $0.75 (+65%), and input from $0.12 to $0.25.
- Qwen3 235B A22B Thinking 2507 output rose from $2.30 to $3.50 (+52%).
- Kimi K3 output rose from $14.00 to $15.00 (+7%).
Some prices fell the same day. Claude Haiku Latest dropped from $1.00 to $0.10 input and from $5.00 to $0.50 output (-90%). DeepSeek Pro Latest output fell from $4.20 to $0.66 (-84%).
Some flagships also sit above where they started. Claude 3.5 Sonnet listed at $3 / $15 at first sight and at $6 / $30 on 2026-03-30. DeepSeek V4 Flash went from $0.14 / $0.28 to $0.03 / $1.28, with input down and output up.
Why the per-task view matters
A list-price rise only changes your bill if you use that model. The task shape matters more:
- Short chat answers are dominated by output and reasoning tokens. For DeepSeek V4.1 Flash, 5,461 reasoning tokens far outweigh the visible answer.
- Long documents are dominated by input tokens. Grok 4.7's PDF run costs 3.4¢ against 0.96¢ for chat.
- Model choice beats small price moves. On the chat task, the cheapest models sit at 0.06¢ to 0.09¢ per task, while the most expensive reaches 6.3¢ (Claude Fable 5.1, estimated).
Check your own token mix before deciding a price rise is a problem. Use the calculator and read the methodology for how I count reasoning tokens and what "estimated" means.
What to do about it
- Recheck list prices before a large batch. My log shows one model changing twice in a week.
- Compare per-task cost, not per-token price.
- Watch long-input jobs most closely, since input-rate increases hit them hardest.
- Keep a fallback model so you can move some tasks quickly.
See the pages for DeepSeek V4.1 Flash, DeepSeek V4 Pro 0813 and Grok 4.7, plus the tasks for chatbot answers and PDF summaries. Follow news for changes as they happen.
FAQ
Are AI API prices going up or down?
Both, depending on the model and the day. Many flagship prices fell over time, but DeepSeek V4.1 Flash, DeepSeek V4 Pro 0813 and Grok 4.7 now list higher than at first sight.
Does a higher list price always mean a higher cost per task?
No. Cost per task also depends on input, output and reasoning tokens. A model with a higher rate can be cheaper per task if it uses fewer tokens.
Why does one model's price change so much in a week?
List prices on OpenRouter move often, and some changes reverse within days. DeepSeek V4.1 Flash output rose 100% on 2026-10-05 and fell 50% on 2026-10-11. Check the current price before committing to a large workload.