Breaking

DeepSeek Slashes Cached Input Prices Tenfold, Reshaping the Economics of AI Agents

July 12, 2026 at 18:54 EDT

Foundation Models
AI Agents
Infra & Chips

DeepSeek has cut the price of cached input tokens across all of its API models to one-tenth of their launch cost, a move that dramatically lowers the running cost of multi-step AI agents and could redraw the competitive map among frontier model providers. The change, effective from April 26, 2026 (UTC), applies to every DeepSeek API model and arrives alongside a permanent 75% discount on the newly released DeepSeek-V4-Pro.

April 26, 2026 · DeepSeek API

DeepSeek Slashes Cached Token Prices to One-Tenth, Igniting an Agent-Era Price War

Cached input tokens across every DeepSeek API model now cost a fraction of their launch price — making repetitive, multi-step AI agents dramatically cheaper to run and pressuring frontier rivals to follow.

1/10th

Cached input token cost vs. launch price

75%

Permanent discount on DeepSeek-V4-Pro

1.6T

Total params (49B active) MoE, 1M context

V4-Pro input cost, per million tokens

Column height is proportional to price — cache hits are ~120× cheaper than fresh input.

$0.87

Output

$0.435

Cache-miss
input

$0.0036

Cache-hit
input

Cache discount depth — how far each provider goes

Effective input savings on repeated context. More stacked blocks = deeper discount.

~99%

DeepSeek V4-Pro

~75%

xAI Grok 4.5

~10%

GPT-5.5 / Opus-class

Why cache pricing matters

In agents, RAG, and long system prompts, the same context is sent again and again — pricing those repeats near-zero can make agentic input costs "nearly free."

On high cache-hit workloads, effective input drops to around $0.05 / million tokens — via a disk-based KV cache with automatic prefix matching, no code changes required.

Developer upside

Agentic input costs described as "nearly free"; V4-Pro pitched as a cost-effective stand-in for Opus-class models on reasoning and coding.

The catch

Savings depend on exact prefix matching — a small change to a prompt's start knocks the request out of cache and back to full price.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.