DeepSeek V4 Flash Is So Cheap It’s Breaking My Brain
·
By Arjun Walia
I Almost Didn't Believe the Numbers
I saw a Reddit post last week where someone said they'd run DeepSeek V4 Flash with max thinking for four hours — subagent loops, planning, the works — and it cost them five cents. Five. Cents. I thought they were rounding down or running on cached context. So I tried it myself.
I gave it a weekend project: refactoring a Python app, adding a few features, and writing tests. Two full days of heavy use, probably 70+ requests, each with some context. At the end, my DeepSeek API bill showed $0.65. That's less than a vending machine coffee.
I'm still not sure I trust the accounting. But the numbers are right there. The market should be losing its mind over this.
The Pricing Doesn't Make Sense
Let's put it in perspective. A similar task on Claude Sonnet would have cost me around $10. On GLM 5.1, maybe $3. On GPT-4.1, forget it. DeepSeek V4 Flash is charging roughly 0.15 per million input tokens and 0.60 per million output. The caching is so aggressive that repeated prompts drop to nearly nothing. One commenter reported $3.50 for 163 million tokens over five days.
But there's a trap. The model has a 1 million token context window. If you let that grow unchecked — say, you keep loading more files without resetting — your per-request cost skyrockets. Another user said they hit $10 in a single day because they were refactoring a 20k-line codebase and never started a fresh session. The solution is brutal but simple: switch sessions for each task. At 40-80k context, you stay in the pennies. At 200k+, you're in dollars. Manage your context, or it manages your wallet.
It's Not Just Cheap — It's Good Enough
The big question: does Flash actually work? For most of my daily coding, yes. It handles plans, refactors, and test generation without drama. It's not as creative as GLM 5.1 on hard architectural problems, and it over-explains simple tasks sometimes. But for iteration-heavy workflows — fix this bug, add this endpoint, write these docs — it's faster than I am at deciding what to do next.
One commenter said they had to use GLM 5.1 as a fallback for only the hardest 10% of tasks. That matches my experience. Flash as the daily driver, GLM or Kimi K2.6 for the tricky stuff. And the cost difference means I'm comfortable spinning off five subagents instead of micromanaging one. That's a workflow change you can actually afford.
The Comparison Game (Everyone Has an Opinion)
In the thread, people were arguing rankings like they were debating draft picks. Kimi K2.6 is better at following instructions. GLM 5.1 catches bugs Flash misses. Mimo 2.5 Pro is smarter at deciding what to build. And Flash is the cheapest competent worker.
I've tested Kimi and GLM. They're stronger on specific tasks — GLM especially for reasoning through broken code. But for 90% of what I do (write, refactor, debug), Flash is fast enough and the price means I can let it run wild. I've seen it solve problems it had no business solving at that price point. One commenter said they used it for a sprint with Opus as manager — Flash did the work, Opus reviewed, cost $0.24 vs $9 for the equivalent Sonnet call. That's not a comparison anymore, that's a category shift.
Why Is Nobody Talking About This?
The original poster in the Reddit thread asked the same thing. My theory: people got burned by R1 hype. Everyone went crazy over R1, then cooled off when the next models came along and it wasn't a one‑click replacement for everything. Now a strong model drops and nobody wants to be the first to yell "game changer" again. But the cost numbers aren't hype — they're just math.
Another commenter nailed it: the market is moving from "best model wins" to "best developer workflow wins." When you can run a model for hours on pocket change, you stop optimizing for raw benchmark scores and start optimizing for iteration speed and context efficiency. That's a much healthier place to be.
Getting Started (and What to Watch For)
Easiest path: grab a DeepSeek API key directly. The discounts only apply there — OpenCode Go users don't get the reduced pricing, but the direct API is still stupid cheap. I'm using OpenCode with the DeepSeek provider and it works fine. A few people reported that Ollama Cloud had reliability issues (timeouts, server overloads) during peak hours. Haven't hit that myself, but worth checking current status.
You'll want to set Max Thinking and use Flash for the heavy lifting. Some people use Pro for planning then Flash for execution, which works if you need more reasoning upfront. I just use Flash for everything and escalate to GLM only when it gets stuck. Results vary by codebase size; if you're working on a monolith, be extra strict about context management.
Take It for a Spin
If you're still paying $10+ a day for Claude or GPT and not getting dramatically better results, try DeepSeek V4 Flash for a weekend. Give it a real task, not a demo. Watch the usage. See what it can do when you're not worried about the meter running.
I'm not saying it's perfect. It hallucinates more than GLM, and it overcomplicates trivial code sometimes. But for the price — and for the iteration freedom it unlocks — it might change how you think about coding with AI. I know it changed mine. Try it and see if your bill makes you laugh the way mine did.
