Reasonix — the DeepSeek-native coding agent that proves V4 Pro's cache math (99.82% hit, $12 instead of $61)

Last updated: May 25, 2026. Read time: 7 minutes. What you’ll learn: What Reasonix is (and isn’t — it’s not DeepSeek-official), why a “DeepSeek-only” coding agent matters, the 99.82% cache-hit benchmark and what it proves, install in 2 minutes, and the honest Claude Code / Cursor / Aider comparison.

Two days after DeepSeek made the V4 Pro 75% discount permanent (see our V4 Pro walkthrough), a third-party open-source coding agent called Reasonix hit Hacker News at 450↑. The pitch is simple: a terminal AI coding agent engineered around DeepSeek’s prefix cache so token costs stay near-zero across long agent sessions.

The author publishes a real benchmark — 435M input tokens in a single day, 99.82% cache hit, total cost ~$12 instead of ~$61 — which makes Reasonix the cleanest proof-of-concept for the cache math I wrote about in the V4 Pro tutorial. This article unpacks what Reasonix is, what it isn’t, and whether it’s worth picking up.

1. What Reasonix is (and isn’t)

Reasonix is NOT a DeepSeek-official product. It’s a community open-source project by GitHub user esengine, MIT-licensed, hosted at github.com/esengine/DeepSeek-Reasonix.

What it IS:

  • A terminal-first AI coding agent (CLI / TUI, not an IDE plugin)
  • DeepSeek-only on purpose — quoting the README: “Coupling to one backend is the feature, not a limitation.”
  • Built in TypeScript, distributed via npm (npm install -g reasonix)
  • MIT-licensed, fully open-source
  • Compatible with MCP (Model Context Protocol) as a first-class citizen

What it’s competing with:

  • Claude Code — Anthropic’s terminal agent, closed-source, premium pricing
  • Cursor Composer — IDE-based, subscription + per-use, multi-backend
  • Aider — Apache 2 multi-backend (OpenRouter), no prefix-cache engineering

The “competitive” framing in Reasonix’s own keywords: “Claude Code alternative, Cursor alternative, Aider alternative” — but it’s narrower than that. Reasonix is the right answer when you’ve committed to DeepSeek as your backend; for anyone bouncing between providers, Aider is still the more general tool.

2. The benchmark that matters

The Reasonix README publishes a one-day production benchmark from 2026-05-01 (transcripts and methodology in the benchmarks/ directory of the repo):

MetricValue
Input tokens435 million
Cache hit rate99.82%
Actual cost~$12
Cost if no cache~$61
Savings~80%

This is what V4 Pro’s $0.003625 / 1M cache-hit input pricing looks like in production. The number lines up with the cache math I did in the V4 Pro tutorial — and confirms it’s not theoretical.

Why this matters: for builders running long agent loops where the system prompt + project context + tool definitions stay stable across hundreds of calls, the cache hit rate is the actual variable that determines whether DeepSeek V4 Pro is “cheap enough to ship” or just “cheap on paper.” Reasonix’s 99.82% says the loop can be engineered to hit cache almost every call — if your agent harness is designed around prefix stability.

3. Install + first run (2 minutes)

Prerequisites

  • Node.js >= 22 (older Node won’t work — Reasonix uses recent ESM + worker features)
  • DeepSeek API key — get one at https://platform.deepseek.com/api_keys. The V4 Pro tier is the one you want; V4 Flash works but the cache benefit is smaller because per-token rates are already low.

Install

Two install options, same binary:

# Global install (recommended)
npm install -g reasonix

# Then in any project directory
reasonix code .

Or one-shot without install:

npx reasonix code

On first run, Reasonix prompts you to paste your DeepSeek API key. It’s stored in ~/.reasonix/config.json and reused across all future invocations.

Verify install

reasonix doctor

This runs a health check on:

  • Node version
  • API key validity
  • MCP server configuration

If doctor is green, you’re done.

Basic commands

CommandWhat it does
reasonix or reasonix code [dir]Main coding agent with full filesystem + shell access
reasonix chatChat mode — no filesystem / shell tools, just conversation
reasonix run "task"One-shot non-interactive execution
reasonix doctorHealth check
reasonix updateUpgrade to latest version

Inside an interactive session, slash commands include /todo (plan mode) and /skill new (custom skill definition).

4. The “DeepSeek-only” design choice

Quoting the README directly: “Cache stability isn’t a feature you turn on; it’s an invariant the loop is designed around. … DeepSeek-only on purpose.”

This is a meaningful design decision. Compare:

  • Aider is multi-backend (OpenRouter). Cache behavior varies wildly across providers, so it can’t optimize for any specific backend’s caching semantics. Result: cache hits are incidental, not engineered.
  • Claude Code uses Anthropic’s prompt-caching API, which has different mechanics from DeepSeek’s prefix cache (different TTL, different key shape, different pricing curve).
  • Cursor Composer abstracts away the backend, so the cache layer is whatever Cursor’s infrastructure picks for you.

Reasonix’s bet: if you commit to one backend, you can co-design the agent loop with that backend’s caching to hit 99%+ cache rates, and that’s worth more than the flexibility of multi-backend support. For DeepSeek users, the bet pays off (the 99.82% number is the evidence). For non-DeepSeek users, Reasonix is the wrong tool — pick Aider.

5. Honest Reasonix vs Claude Code / Cursor / Aider

From the README’s own comparison table (with my notes):

FeatureReasonixClaude CodeCursorAider
BackendDeepSeek onlyAnthropic onlyOpenAI + AnthropicAny (OpenRouter)
LicenseMIT (open source)closedclosedApache 2
Cost per tasklow (cache-driven)premiumsubscription + usagevaries by backend
Prefix cacheengineered around itAnthropic prompt cachen/aincidental
Persistent sessionsyespartialn/apartial
MCP / Skills / Hooksyesyesyespartial
IDE integrationterminal onlyterminal + IDE pluginsIDE-firstterminal only
Hardest-task quality(lower)higher (Opus 4.7)mid-highvaries

The honest read:

  • For routine coding agent work on DeepSeek: Reasonix wins on cost by a margin that matters (5-10× cheaper than Claude Code for output-heavy workloads)
  • For hardest reasoning tasks: Claude Code with Opus 4.7 is still the floor, not Reasonix with V4 Pro
  • For multi-provider freedom: Aider is the right tool, not Reasonix
  • For IDE workflows: Cursor Composer is the right tool, not Reasonix

6. When to use Reasonix (and when not to)

Reach for Reasonix when:

  • You’ve already committed to DeepSeek V4 Pro/Flash as your coding backend (see V4 Pro walkthrough for whether you should)
  • You live in the terminal, not an IDE
  • You’re running long agent loops where prefix stability matters (cache hit rate is the bottleneck)
  • You want MIT-licensed tooling you can audit / fork / extend
  • You’re cost-sensitive enough that 5-10× cheaper per task is decisive

Reach for something else when:

  • You’re not on DeepSeek (use Aider for multi-backend)
  • Your workflow is IDE-first (use Cursor Composer)
  • Your task needs frontier reasoning quality (use Claude Code with Opus 4.7)
  • You’re a non-technical user (Reasonix is CLI-only, no GUI)

7. Gotchas — what to expect

After scanning the README + a few hours of follow-up testing reports:

  1. Cache hit rates below 99% mean you’re doing it wrong. If your real-world rate is 80%, your system prompt is drifting between calls. Audit it for datetime.now() or per-request UUIDs.
  2. Cache only persists for tens of minutes. Burst your requests; don’t let the agent idle for an hour and expect cache hits on the next call.
  3. The “$12 instead of $61” benchmark is one specific workload. Your workload will look different. Budget on V4 Pro’s worst-case (no cache) pricing first, then enjoy the cache savings as a margin, not a baseline.
  4. DeepSeek API key requires a payment method. No free tier. If you’re just experimenting, top up $10 and you’ll get weeks of testing.
  5. Reasonix doesn’t include vision / multimodal. DeepSeek V4 Pro is text-only, so neither is Reasonix.
  6. Node 22 is hard requirement. Node 18 / 20 won’t work. Use nvm if you’re stuck on an older version for other projects.

8. The bigger picture

Reasonix is the first widely-shared piece of evidence that DeepSeek V4 Pro’s permanent 75% discount plus its prefix-cache pricing can be engineered into agent loops that cost almost nothing per step. Until now, the “DeepSeek is cheap” thesis was theoretical — V4 Pro list prices vs Claude Haiku list prices. Reasonix turns the thesis into a concrete number: 99.82% cache hit, $12 instead of $61, on a real day’s workload.

For builders deciding between “build on Claude Haiku 4.5 because it’s mature” vs “build on DeepSeek V4 Pro because it’s cheaper”: Reasonix is the existence proof that the second option ships. Whether you adopt Reasonix specifically is less important than recognizing what its benchmark says about the new cost floor for production agent work.

Sources