How Pi + DeepSeek V4 Flash Reduces LLM Input Costs to a Few Dollars

Why Nearly One Billion Tokens Cost So Little

When the DeepSeek V4 Flash model cached input at $0.14 per million tokens for cache‑misses and $0.0028 for cache‑hits, the price difference was 50×. Because the Pi agent repeatedly sends the same prompt prefixes, most of the 9.5 × 10⁸ cached tokens were billed at the low rate, while only about 3.5 × 10⁶ tokens were uncached.

At the time of the test, the cache‑hit rate was 99.93%, turning a potential $132 bill into $2.65. Today DeepSeek uses peak‑off‑peak pricing (0.05–0.10 CNY per million cached tokens, 1.5–3 CNY per million uncached), so the same usage would cost roughly ¥58–¥117, but uncached usage would exceed ¥1,400–¥2,800.

What Is Pi?

Pi is an open‑source Node.js agent framework written in TypeScript and distributed via npm. It provides four usage modes—TUI, SDK, Print, and RPC—allowing it to run as a terminal tool, be embedded in a Node.js program, or act as a background agent process.

The core consists of a unified model API, an agent runtime, a TUI, and an SDK, while extensions add tools, skills, prompt templates, and packages. Pi targets a “minimal terminal coding harness” that gives large language models file, command, tool, session, and UI capabilities.

Why Pi Achieves High Cache Hit Rates

DeepSeek’s context cache is enabled by default and automatically stores repeated request prefixes. In an agent workflow, the first round sends system prompts, project rules, tool definitions, and the user task; subsequent rounds reuse these prefixes and only append new model replies or tool results.

This pattern is similar to continuously extending a long document where the majority of tokens remain unchanged, allowing the cache to serve most of the input.

Actions that can break the cache include:

Injecting the current time into the system prompt each round

Changing tool definitions or their order

Dynamically adding or removing many tools

Re‑organizing historical messages in a middleware layer

Replacing the original session prefix after context compression

Pi’s documentation warns that dynamic tool activation or prompt reconstruction may invalidate the vendor’s cached prefixes. The reported 99% hit rate comes from a specific user, date, and workload; DeepSeek only guarantees best‑effort caching and may clear unused caches after hours or days.

Benchmark: Pi vs. Other Agent Harnesses

Composio evaluated four agent harnesses on 30 tasks using the same DeepSeek V4 Flash model. Pi completed 20 tasks, achieving the highest pass rate, a median cost of $0.012 per task, and a median runtime of 132 seconds.

When the comparison was expanded to eight harnesses, the results were:

Pi: 20 tasks

Oh My Pi: 17 tasks

Claude Code: 16 tasks

Codex: 16 tasks

Deep Agents: 16 tasks

Prime Agent: 15 tasks (6 runs unscored)

Hermes Agent: 15 tasks

OpenCode: 14 tasks

The evaluation shows that performance depends not only on the underlying model but also on system prompts, tool sets, error‑feedback mechanisms, session persistence, and context compression strategies.

DeepSeek later open‑sourced its own harness, also built with TypeScript and Node.js, but it was still in preview when the benchmark was conducted, so direct comparison data are unavailable.

Is This Combination a “Game‑Changer”?

For developers who want low‑cost, high‑throughput terminal tasks and are comfortable configuring models and extensions, Pi + DeepSeek V4 Flash offers an attractive pairing: Pi provides a transparent, modifiable Node.js agent base, while DeepSeek’s cheap cache‑hit pricing dramatically reduces input costs.

However, Pi does not include a sandbox; file I/O, command execution, and extensions run with the same privileges as the Pi process, so untrusted code must be isolated manually.

Ultimately, while the historic $2.65 bill demonstrates the cost‑saving potential, actual savings depend on workload characteristics, prompt stability, and cache‑friendly usage patterns.