The Vibe Father
Model review
Every few months a model shows up whose price looks like a typo. DeepSeek V4 Pro, released April 27, 2026 by the open-weight lab DeepSeek, costs $0.435 per million input tokens and $0.87 per million output. That's not "cheap for the quality" — that's cheaper than some labs charge for models two generations behind. The question we've spent the last two months answering: is the quality real, or is this the usual budget-model bargain where the discount comes out of your review time instead of your wallet?
Short version: the quality is real, with one asterisk and one missing number. Long version follows.
The numbers
Our Vibe Coding Index weights SWE-bench Verified 40%, Terminal-Bench 30%, LiveCodeBench 30%. V4 Pro has published two of the three:
| Model | SWE-bench Verified | Terminal-Bench | LiveCodeBench | $ in / $ out per M |
|---|---|---|---|---|
 Claude Sonnet 5 | 85.2 | — | 82.4 | 3.00 / 15.00 |
 Gemini 3.5 Flash | 79.3 | 76.2 | 87.6 | 1.50 / 9.00 |
 DeepSeek V4 Pro | 77.6 | not yet published | 87.5 | 0.435 / 0.87 |
 Qwen3.7 Max | 77.3 | — | 87.1 | 2.50 / 7.50 |
Sit with that row for a second. 77.6 on SWE-bench Verified is real-repo capability within 1.7 points of Gemini 3.5 Flash and ahead of GPT-5.3 Codex's 74.8. The 87.5 LiveCodeBench is 0.1 behind Flash and effectively tied with the entire frontier pack. And it costs less than a third of Flash's input price. The asterisk: Terminal-Bench is not yet published, so we genuinely don't know how it handles long agentic shell sessions — the skill where budget models most often crack. The other honest caveat is speed: 64 tokens per second is the slowest of any model we'd recommend, roughly a third of Qwen3.7 Max's pace. You'll feel it on long generations. The 1M context window, at least, is full-sized.
The price math
Here's where it stops being a benchmark story and becomes a budget story. A heavy month — 50 million input tokens, 10 million output — costs 50 × $0.435 + 10 × $0.87 = $30.45. The identical month on Gemini 3.5 Flash is $165. On Claude Sonnet 5 it's $300. On Claude Opus 4.8 it's $500. You could run V4 Pro sixteen times harder than Opus 4.8 for the same bill, on a model that trails it by 11 points on SWE-bench, not 40. Our cheapest-models roundup ranks the whole budget tier, and V4 Pro's price-per-benchmark-point is the best number in it. This is also why it anchors our open-weight guide: the weights are open, so the price floor is enforced by competition, not by one vendor's mood.
Its best seat on the team
V4 Pro's seat is budget builder behind a strict reviewer. The economics make a specific workflow rational: let V4 Pro do the volume — first drafts, routine tickets, test scaffolding, the wide middle of your backlog — and put a stronger model in the reviewer seat to catch what a 77.6-SWE model misses that a 95.0 model wouldn't. Because V4 Pro's tokens cost so little, even a generous review budget on top still lands your total far below running a flagship end-to-end. The one rule we'd call non-negotiable until Terminal-Bench numbers exist: keep a verification gate on its agentic work. Unpublished isn't failing, but it isn't passing either, and long unattended shell sessions are exactly where we'd want the evidence first.
The open-weight status deserves one more paragraph, because it changes the risk calculus in a way pricing tables don't capture. When weights are open, the hosted price you pay is set by competition among providers rather than by a single vendor's margin targets — which means the $0.435/$0.87 floor is defensible, not promotional. It also means you have an exit: if a provider degrades quality, throttles you, or repricess, the same model is available elsewhere, or on your own hardware if the workload justifies it. For teams that have been burned by overnight API repricing — and by mid-2026, that's most teams — that optionality is worth real money on top of the sticker savings.
Two things it's not. It's not your fast-feedback pair-programmer — at 64 tok/s the interactive loop drags, and Qwen3.7 Max or the models in our fastest-models guide own that seat. And it's not your architect for the hairiest cross-cutting refactors, where the flagships in our overall rankings justify their premiums. Between those poles sits most of the actual work, and that's V4 Pro country.
Verdict
DeepSeek V4 Pro is the strongest price disruptor on our board: frontier-adjacent scores on two of three benchmarks at roughly one-fifth the cost of its closest quality peer. Paired with a good reviewer and a verification gate, it turns a $500 model bill into a $50 one without turning your codebase into a crime scene. Who should skip it: anyone whose workflow is interactive and speed-sensitive — the 64 tok/s throughput will annoy you daily; teams that lean heavily on unattended agentic terminal work, at least until Terminal-Bench numbers arrive; and anyone whose employer pays the bill and whose time is the scarce resource, for whom a flagship's fewer retries is the better trade. Everyone else should be auditioning it this week.