DeepSeek Built Its Own Coding Agent — Reasonix Takes Aim at Claude Code and Codex

Last week DeepSeek fired the opening shot of a price war with a permanent 75% discount. This week the target isn’t the model — it’s the tooling. DeepSeek has launched Reasonix, its own coding agent, and walked straight into the territory Claude Code and Codex have quietly carved up. What makes this more than another Cursor clone is the architecture: Reasonix is the first serious example of a model and an agent designed as one product.

What Reasonix Actually Is

Reasonix is a native coding agent built directly on top of DeepSeek’s reasoning model. Until now, the model vendor and the agent vendor were almost always different companies. Anthropic’s Claude Code was the exception. Everyone else — Cursor, Cline, Aider, Continue — was a third party renting API access and wrapping it in clever prompts.

DeepSeek took the other path. Tool use, file edits, and multi-step reasoning were baked into the model during training, not bolted on as a system prompt afterward. Early testers on Hacker News describe the behavior as noticeably less “jailbroken” than a wrapper-style agent — fewer awkward retries, fewer moments where the model forgets it’s holding a hammer.

The Real Weapon Is Caching

The headline isn’t the price. It’s the caching architecture. Coding agents reread the same codebase dozens of times in a single session. File trees, function signatures, prior turns of the conversation — all of it gets re-shipped to the model with every request. Pay full freight on every token and a single engineer’s monthly API bill can clear four figures fast.

DeepSeek claims Reasonix is engineered so that over 90% of tokens hit the cache, and cached tokens are priced at roughly one-tenth of fresh ones. Stack that on top of the 75% permanent discount and the math gets uncomfortable for the incumbents: equivalent work runs at roughly one-tenth the effective cost of Claude Code.

How Claude Code and Codex Users Are Reacting

Benchmarks and side-by-sides are already showing up on Reddit and X. The early consensus is roughly this.

Claude Code is still faster and smarter on the hard stuff. Anthropic’s infrastructure spend wasn’t for nothing. One developer summed it up: “For non-trivial refactors, Claude 4.6 is still two steps ahead.”

For routine work, Reasonix is fine. Boilerplate, test scaffolding, small bug fixes — the bread and butter of daily coding — land somewhere around Codex quality. And it does that work at a tenth of the price.

Then there’s the open-source angle. The Reasonix CLI is open source, and the model weights are downloadable. Enterprises that banned cloud coding assistants for security reasons suddenly have an option that runs behind their firewall.

Where the Market Goes From Here

Claude Code and Codex have effectively had no competition. They were expensive because they could be. Is Reasonix strong enough to flip that? Honestly, not yet. Code quality on complex tasks, long-context stability, and IDE polish are all visibly behind.

But the shape of the market has shifted. DeepSeek’s implicit thesis is that only vertically integrated players — model plus agent plus infrastructure — survive the coming price compression. OpenAI is already there with Codex. Anthropic is already there with Claude Code. DeepSeek just added two more legs to the stool: open weights and aggressive pricing.

The harder question lands on the third-party agents. When the model vendors ship their own first-party agents, what exactly is the wrapper business selling?

The Takeaway

Nobody is ripping Claude Code out of their workflow tomorrow. But the assumption that coding agents are inherently a premium SKU is starting to crack. The tool you’ll be using a year from now probably isn’t the one you’re using today. The interesting question for your team isn’t which agent is best — it’s whether “AI coding spend” has already become its own line item next to AWS, or whether you’re still pretending the free tier will hold.