Our team of four devs blew through the Claude Code weekly budget last week. So on May 4th I topped up $20 on the DeepSeek API and we pointed our coding tools at DeepSeek V4 Pro instead. Twelve days later — today, May 16 — we've used $7. Across all four of us. 519 million tokens.
This post covers how DeepSeek V4 Pro stacks up against Claude Opus 4.7 and Sonnet 4.6 for actual day-to-day coding work, what it actually costs, and where the cheaper option falls short.
The numbers
$7.00 spent. 519,622,966 tokens. 2,389 API requests. Four developers. Twelve days.
For context, "tokens" are the units AI tools charge for — roughly four characters of text each. 519 million tokens is the kind of volume that would burn through a Claude Code Pro subscription's weekly limit in days, not weeks. We were hitting those limits constantly. On DeepSeek V4 Pro, we haven't hit anything.
Why it's so cheap
DeepSeek lists three prices per million tokens for V4 Pro:
- Input (cache hit): $0.003625
- Input (cache miss): $0.435
- Output: $0.87
Those are 75% off the regular rates, and the discount runs until May 31, 2026 — about two weeks from now. After that, the same usage costs roughly four times more. Even at the full price it's still meaningfully cheaper than the equivalent Claude API tier, but right now is the bargain window.
The "cache hit" price matters more than it looks. Coding tools re-read your codebase constantly — every follow-up question, every "now do the same thing for the other file." That repeated context comes back at the cheapest rate, not the cache miss rate. Most of our 519 million tokens were cache hits, which is why $7 stretched as far as it did.
Wiring it up with CC Switch
DeepSeek serves an Anthropic-format endpoint at api.deepseek.com/anthropic. That means anything built for the Anthropic API — including Claude Code — can talk to it without code changes. Same request shape, same tool-call format, same SDK.
We manage the actual swap with CC Switch, an open-source desktop app that holds all your provider configs in one place — Claude's official endpoint, DeepSeek, anything else — and flips Claude Code between them from the system tray. No config file editing, no environment variable juggling, no terminal restart needed. Claude Code is the one CLI in CC Switch's lineup that supports hot-switching, so the change takes effect immediately.
In practice: I have Opus 4.7 set as one provider and DeepSeek V4 Pro as another. When I'm about to do a heavy refactor I click Opus. For everything else I click DeepSeek. The team runs the same setup.
How the code quality compares
Better than Sonnet 4.6. Slightly worse than Opus 4.7. That's the headline.
If you've been getting your work done with Sonnet 4.6, DeepSeek V4 Pro is a strict upgrade at a fraction of the cost. If you've been spoiled by Opus 4.7, you'll notice a gap — but it's smaller than the price difference suggests.
Where it falls short of Opus 4.7
The gap shows up in one specific place: code generated by DeepSeek V4 Pro builds the feature you asked for, but doesn't think hard about everything the feature touches.
Opus 4.7 will look at your function and say "okay, but this calls into three other places — let me handle those edge cases too, and update the callers so nothing downstream breaks." DeepSeek V4 Pro builds the thing. It doesn't reach out into the dependency tree the same way, and scenario handling around the edges is thinner.
For greenfield code or self-contained features, you won't notice. For changes that touch a load-bearing function with lots of callers, you'll be cleaning up after it.
Verdict
$7 for 519 million tokens across four developers is the kind of math that pays for itself immediately. If you're a team hitting Claude Code's weekly limits, DeepSeek V4 Pro is worth wiring up as your default — especially before the 75% discount expires on May 31. With CC Switch in the tray, keeping Opus 4.7 around for the architecture-sensitive work — refactors, multi-file changes, "this touches a lot of stuff" tasks — is a one-click choice. Let DeepSeek handle everything else.
We'll keep running this side-by-side for the rest of the month and see if the math holds.