Answer-First Lead
DeepSeek Reasonix, an open-source AI coding agent built exclusively for DeepSeek models, hit 8,000+ GitHub stars and 627 points on Hacker News after its May 2026 release. The twist? It’s not trying to out-feature Claude Code. It’s optimising for one thing Big Tech agents ignore — keeping DeepSeek’s prefix cache warm. One real user ran 435 million input tokens in a day and paid $12 instead of $61.
🔍 THE BOTTOM LINE
Reasonix proves the cheapest AI coding agent isn’t the one with the lowest per-token price — it’s the one that makes the cache work.
What Is DeepSeek Reasonix?
Reasonix is an open-source terminal-based AI coding agent that only works with DeepSeek models. It’s built in TypeScript with an Ink TUI, currently at v0.30.5. Unlike Claude Code or GitHub Copilot, which treat caching as a nice-to-have, Reasonix engineers its entire agent loop around prefix-cache stability — a DeepSeek-specific mechanic where previously processed prompt prefixes can be reused without re-computation if the byte sequence stays identical.
That sounds niche. It isn’t. Here’s why:
| Feature | Reasonix | Claude Code | GitHub Copilot CLI |
|---|---|---|---|
| Model support | DeepSeek only | Claude only | Multiple |
| Cache-first design | Yes (core invariant) | No | No |
| Cache hit rate (real user) | 99.82% | Not published | Not published |
| Cost (435M tokens/day) | ~$12 | ~$61 equivalent | N/A |
| Open source | Yes (MIT) | No | No |
| Stars | 8,070+ | N/A | N/A |
The Cache-First Architecture — Why It Matters
Most AI coding agents treat the model API as a stateless pipe: send tokens, get tokens, pay for tokens. Every request starts from scratch.
DeepSeek’s API has a prefix-cache feature: if your prompt starts with the same byte sequence as a recent request, the cached computation is reused. You pay less and get responses faster.
The problem? Most agents invalidate the cache every turn. They shuffle system prompts, append conversation history inconsistently, or inject timestamps. Each of these changes a single byte in the prefix, and the whole cache misses.
Reasonix doesn’t just “support” caching. It’s designed around four mechanisms that keep the prefix byte-stable across long sessions:
- Deterministic system prompt ordering — never shuffled, never appended to mid-loop
- Conversation windowing — old messages are pruned from the bottom, not inserted in the middle
- Tool-call repair — broken tool calls are fixed in-place rather than re-sent with different formatting
- Cost meter in the TUI — you can see your cache hit rate in real time
The result: one user on May 1, 2026, processed 435 million input tokens with a 99.82% cache hit rate. The cost was roughly $12. The same workload without caching on DeepSeek’s v4-flash would have been around $61. That’s an 80% saving from one architectural choice.
Why This Hit 627 Points on Hacker News
The HN crowd doesn’t get excited by “yet another AI coding agent.” Reasonix resonated because it solved a problem developers actually feel — cost on long sessions.
Claude Code is great, but a full-day coding session with Claude Opus can run $50-100+. DeepSeek V4 is already cheap per-token, but Reasonix makes it absurdly cheap by keeping the cache hot. The 99.82% cache hit stat is the kind of number that makes a developer stop scrolling.
The project also hit a nerve by being DeepSeek-only — not as a limitation, but as a design decision. Generic agents compromise cache stability to support multiple providers. Reasonix chose depth over breadth, and the community responded.
As one HN commenter put it: “Every other agent treats the cache as a bonus. This one treats it as the whole point.”
Three Pillars, Not Just One
Reasonix’s architecture has three pillars:
Pillar 1 — Cache-first loop: The agent loop is ordered to keep the prompt prefix byte-identical across turns. This is the core innovation.
Pillar 2 — Tool-call repair: When the model generates a malformed tool call (which happens often with DeepSeek’s open-weight models), Reasonix doesn’t throw it away and re-prompt — it fixes the call in-place, preserving the cache.
Pillar 3 — Cost control: Built-in spend limits, per-session cost tracking, and a real-time TUI meter showing tokens, cache hit rate, and spend.
These three pillars work together. Cache stability reduces cost. Tool-call repair preserves cache stability. Cost control makes long sessions viable. It’s a coherent system, not a feature checklist.
What About the DeepSeek-Only Limitation?
This is the obvious objection. If you’re a Claude shop or a GPT shop, Reasonix won’t work for you. It requires a DeepSeek API key and a Node ≥ 22 runtime.
But that’s the point. The prefix-cache mechanic is DeepSeek-specific. Supporting multiple providers would mean compromising on the one thing that makes Reasonix fast and cheap. It’s the same reason a Formula 1 engine isn’t designed to run on diesel — optimisation requires commitment.
If you’re already using DeepSeek V4 or V4-flash, Reasonix is a no-brainer. If you’re committed to Claude or GPT, you’ll wait for someone to build an equivalent cache-first agent for your model — or build it yourself.
NZ Lens
New Zealand developers paying in NZD feel API costs more acutely than US developers. A $50/day Claude Code habit is roughly $82 NZD. Reasonix on DeepSeek V4-flash at $12/day (about $20 NZD) is a dramatic difference for solo devs and small teams.
The open-source nature also matters: NZ companies with data sovereignty concerns can run DeepSeek models on local infrastructure and use Reasonix as the interface, keeping code and prompts in-country.
❓ Frequently Asked Questions
Q: Can I use Reasonix with Claude or GPT models? No. Reasonix is DeepSeek-only by design. The prefix-cache architecture only works because it’s tuned to DeepSeek’s specific API behaviour. A multi-model version would sacrifice the cache stability that makes it effective.
Q: How do I get started?
Install Node ≥ 22, run npm install -g reasonix, then reasonix code my-project. Paste your DeepSeek API key on first run. Or use npx reasonix code for a one-shot without global install.
Q: Is it production-ready? It’s at v0.30.5 and moving fast. The core loop is stable, the cache architecture is sound, and real users are running it for full workdays. But it’s still pre-1.0 — expect rough edges, especially on the desktop GUI client which ships as a prerelease.
🔍 THE BOTTOM LINE
Reasonix isn’t the most feature-rich AI coding agent. It doesn’t have Claude Code’s polish or Copilot’s distribution. But it does something none of them do — it makes long AI coding sessions genuinely cheap by engineering the entire loop around cache stability. In a market where everyone’s competing on model quality, Reasonix is competing on cost architecture. That’s a different war, and right now, it’s winning.
