DeepSeek's Price Hike Signals the End of Cheap AI APIs

DeepSeek plans to raise API prices across the board, its steepest reversal since undercutting rivals with $0.14-per-million-token pricing. The company said on August 6 that "a relatively large increase is expected" and told customers to "plan your usage accordingly." The move ends a year in which AI API prices only went one direction: down.

DeepSeek Ends the Price-Cutting Era

DeepSeek had driven "near-frontier intelligence" down to floor prices with its V4-Flash model at $0.14 per million input tokens. It had also already announced a peak/off-peak pricing mechanism, not yet in effect: during two daily windows, 9:00-12:00 and 14:00-18:00 Beijing time, input token prices double from $0.14/M to $0.28/M, and output prices double from $0.28/M to $0.56/M. Cached input stays far cheaper than a cache miss — $0.0028/M versus $0.14/M, a 50x gap — which means how a team structures its prompt prefix now matters more than which model it picks. OpenAI and Anthropic still price by model and token count regardless of time of day; DeepSeek is the first frontier provider to break from that.

Claude Code Users Already Feel the Squeeze

The price pressure isn't limited to DeepSeek. A cost-management guide aimed at engineering leads warns that unmanaged Claude Code usage can drain more than €1,500 a month from a 10-person team. The gap between Anthropic's Sonnet and Opus pricing runs roughly 5-to-1, according to the guide, making model selection the single biggest cost lever available to a team lead. At typical usage — 40 calls a day per developer, 4,000 input and 800 output tokens per call — a 10-person team lands at 180-300 EUR a month on Sonnet pricing, or 900-1,500 EUR a month if the same usage runs on Opus.

Self-Hosting Llama 3.3 70B: Is It Worth It?

One widely shared deployment guide claims a self-hosted Llama 3.3 70B setup — running on a $12-a-month DigitalOcean GPU droplet with vLLM and dynamic LoRA adapter routing — served 50-plus concurrent requests and processed 500,000 tokens of test inference for under $1. The author claims the per-token cost works out to roughly $0.11 per million input tokens, versus $15 per million on Claude Opus, a 135x gap by their own math. That figure comes from a single self-reported test, not an independent benchmark, so treat it as a directional signal for high-volume, low-sensitivity workloads rather than a drop-in replacement for Claude Code on daily coding tasks.

Option	Input token price	Source
Claude Opus API	$15/M	deployment guide
DeepSeek V4-Flash (off-peak)	$0.14/M	DeepSeek pricing notice
DeepSeek V4-Flash (peak window)	$0.28/M (announced)	DeepSeek pricing notice
Self-hosted Llama 3.3 70B (claimed)	~$0.11/M	author's self-reported test

What to Watch

• DeepSeek hasn't published the exact hike percentage or effective date — engineering teams billing against its API should watch for that announcement rather than budgeting on assumptions.
• Whether Anthropic or OpenAI follow DeepSeek into time-of-day pricing, or hold to flat per-token rates, will shape how teams design prompt-caching strategy over the next few months.
• Model-gateway patterns that route requests by price tier — treating "which model" as a config value instead of a hardcoded constant — are the practical hedge against further swings in either direction.