DeepSeek Just Raised API Prices by Up to 1,100%: What Changed and What to Do About It

DeepSeek Just Raised API Prices by Up to 1,100%: What Changed and What to Do About It

DeepSeek spent most of 2026 as the answer to "how do I run an agent workflow without watching the bill." That's over, at least for now. Starting August 16, 2026, DeepSeek raised API pricing across its V4 model family by margins ranging from 50% up to more than 1,100%, depending on the model, token type, and time of day. If you built anything assuming DeepSeek would stay the cheapest option indefinitely, it's time to check the math again.

What changed in DeepSeek's pricing

DeepSeek introduced peak and off-peak pricing on top of the increase itself. Peak hours run 01:00-04:00 and 06:00-10:00 UTC; everything else is off-peak. For V4-Flash, the new off-peak rate is $0.22 per million input tokens (cache miss) and $0.66 per million output tokens; peak rates double that, to $0.44 and $1.32. V4-Pro moved from $0.435 per million input tokens to $0.66 off-peak or $1.32 at peak, with output rising from $0.87 to $1.98 off-peak or $3.96 at peak. Off-peak pricing is still above the old flat rate in every listed category, so there's no version of the new schedule that costs the same as before.

DeepSeek's stated reason is resource allocation, not margin. The company says it wants developers to "schedule their tasks based on actual usage" rather than hammering the API at whatever hour is convenient. Reporting from InfoWorld and Reuters both frame it the same way: V4-Flash's launch pricing was so cheap that demand outran DeepSeek's own compute capacity within weeks.

The numbers that made this believable

Two engineers posted their actual DeepSeek dashboards on Hacker News after V4-Flash launched July 31, and the gap between "cheap" and "this is unsustainable" is obvious in hindsight:

At those rates, a workload that would cost hundreds of dollars a month on a frontier model was costing single-digit dollars on DeepSeek. That's the gap the new pricing closes, at least partially.

This is not just a DeepSeek story

Zoom out and the direction is bigger than one vendor. Inference now consumes a large majority of enterprise AI budgets, GPU rental prices spiked through 2026 (H100 contracts up roughly 40%, Blackwell up roughly 48%), and industry reporting points to 30-50% API price increases across multiple providers over the next 18 months as token demand keeps outgrowing available compute. None of that erases the longer trend: independent tracking still puts frontier token pricing about 88% below its March 2023 baseline. Prices are rising right now inside a multi-year trend that's still pointed down. Both things are true at once.

What to do about it

Re-check your cost assumptions this week. If you sized a workflow around July's V4-Flash pricing, rerun the math against the new peak and off-peak rates before an invoice surprises you. Shift non-urgent batch jobs to off-peak hours, since off-peak is still meaningfully cheaper than peak, even though it's pricier than the old flat rate. Stop assuming any single provider stays cheapest: the model that was the best deal in July isn't guaranteed to be the best deal in August. If cost-per-task matters more than brand loyalty, [routing and model-switching are worth the setup effort](https://questloops.com/blog/how-to-cut-your-ai-agent-s-api-bill-token-compression-smart-routing-and-when-a-gateway-pays-for-itself), whether that's manual or handled by a gateway.

The bigger lesson from this cycle isn't "DeepSeek got expensive." It's that inference pricing across the board is now unstable enough that hardcoding one model's price into your cost model is a mistake. Budget for volatility, not just for cost.

FAQ

**Did DeepSeek prices really go up over 1,000% for some tiers?** Yes, for specific token types and peak-hour requests. Reporting from InfoWorld confirmed increases "by more than 10x" on some V4 pricing categories, with the exact multiple depending on the model and cache status.

**Is DeepSeek still cheaper than OpenAI or Anthropic?** For most workloads, yes, even after this increase, DeepSeek remains priced well below flagship models from the major labs. It's no longer the outlier-cheap option it was in July.

**Should I switch providers entirely?** Not necessarily. The better move for most teams is routing simple requests to whichever model is currently cheapest for that task, rather than betting a whole workflow on one vendor's pricing staying flat.