DeepSeek-Reasonix trends on HN — a community coding agent that runs on $12 instead of $61 by treating prefix cache as a hard invariant
esengine's Reasonix is a DeepSeek-native, MIT-licensed terminal coding agent designed entirely around prefix-cache stability. Real case study: 435M input tokens, 99.82% cache hit, ~$12 vs ~$61 without cache on V4 Flash. 7k★ on GitHub, npm install -g reasonix away.
A community-built terminal coding agent called Reasonix trended on Hacker News yesterday (387 upvotes within 10 hours of submission) with one architectural bet that’s easy to summarise: build the whole agent loop around DeepSeek’s prefix cache instead of around the model. The repo at esengine/DeepSeek-Reasonix is MIT-licensed, sits at ~7k stars and 372 forks, and ships a npm install -g reasonix install path.
What it actually is
A Node ≥ 22 terminal agent (TypeScript, distributed on npm) that talks to DeepSeek V4 Flash / V4 Pro directly. It’s not a wrapper around Claude Code or Aider — the codebase, the planner, the tool loop and the diff renderer are written from scratch with one stated invariant: every byte of the input prompt across turns must be cache-stable. The project’s own line: “Cache stability isn’t a feature you turn on; it’s an invariant the loop is designed around.”
Install and first run:
npm install -g reasonix
reasonix code my-projectFeature surface (from the README’s capability grid):
- Cell-diff renderer (SEARCH/REPLACE edit proposals; nothing touches disk until
/apply) - MCP integration (so it can also call external MCP servers)
- Plan mode +
/todoworkflow - Per-workspace shell allowlist with exact-prefix matching
- Web dashboard, persistent sessions, hooks/skills/memory
- Semantic search, auto-checkpoints, transcript replay,
/effortknob
The number that drove the HN thread
The README pins a real case study from 2026-05-01: a single day’s session that drove 435M input tokens at 99.82% cache hit, costing ~$12 total versus the ~$61 the same workload would have cost without prefix caching on V4 Flash. That math lines up with DeepSeek’s published pricing — V4 Flash cache hits run at $0.0028 / 1M input tokens vs $0.14 / 1M cache miss (a 98% delta), and DeepSeek dropped cache-hit price to 1/10 of launch on 2026-04-26.
The point is not that Reasonix invented prefix caching — it’s available to anyone hitting DeepSeek’s API. The point is that most agent frameworks invalidate the cache constantly (timestamps in system prompts, per-turn tool descriptions, RAG snippets shuffled into the middle of the prefix). Reasonix’s design choice is to enforce byte-stable prefixes as a contract across every layer, so the 99% cache-hit number isn’t aspirational — it’s the steady state.
What this means if you’re building with DeepSeek
1. The economics of “leave the agent running” finally make sense. With Claude Code on Opus or Cursor’s Composer 2.5 Standard, a long pair-programming session is real money — $5-15/hour is normal. With Reasonix on DeepSeek V4 Pro at the permanent 75% discount, and the cache hit rate the design enforces, that same long session is roughly an order of magnitude cheaper. The “stop the agent because it’s burning credits” reflex stops being a thing.
2. It’s intentionally not multi-provider. Most agents try to be model-agnostic. Reasonix is the opposite — it’s coupled to one backend on purpose, because the cache-stability invariant only holds against a single provider’s tokenizer and cache implementation. If you want to swap to Anthropic or OpenAI you’d need a different agent. For builders who’ve already standardised on DeepSeek for the cost reason, that coupling is the feature.
3. The local-vs-API agent stack just gained a third lane. Until now the choice was: (a) Claude Code / Cursor on premium frontier APIs, (b) llama-server with --tools all on your own hardware. Reasonix is (c) — a frontier-quality reasoning model via API, but with cost behaviour that’s closer to local. For most production work the privacy/latency profile is still “you’re sending code to a server in China,” which is its own decision, but the cost profile changes the calculus.
Honest caveats
- Community project, not officially endorsed. DeepSeek hasn’t published anything about Reasonix. If DeepSeek changes its cache semantics, Reasonix’s whole thesis bends.
- Single backend. No graceful fallback to a second provider when DeepSeek’s API has an outage — and DeepSeek has had public outages.
- Sandbox is opt-in. The shell allowlist is exact-prefix matching, not a real sandbox. Don’t point it at a directory you wouldn’t
rm -rfyourself. reasonix(no DeepSeek- prefix) on npm is a different package. The install command isnpm install -g reasonixper the README; double-check the source repo if you’re picky about supply chain.
How it fits the rest of the 2026 agent landscape
If you’re already standardised on Cursor’s IDE-in-buffer model, see Cursor Composer 2.5 — different value prop, different price floor. If you’re testing the llama.cpp built-in tools route, Reasonix is the “I want frontier reasoning + API economics” lane in the same conversation. And for context on why DeepSeek V4 Pro’s pricing is the underlying enabler here, see the V4 Pro deep dive — Reasonix is, in a real sense, the first agent that’s been engineered around that price floor.