DeepSeek V4 Flash Did for $0.53 What Kimi K3 Billed $25
DeepSeek's V4-Flash-0731 update shipped with open weights on 31 July, and has spent the week being measured rather than announced. One hands-on evaluator ran three multi-hour, frontier-grade build tasks through the cloud API for $0.53; the same game task on Kimi K3 had cost him $25. Four observer reports corroborate the finding from different angles — an agentic leaderboard, a cost index, and consumer hardware.
What the Source Actually Says
Tonbi's AI Garage ran every task twice: once on the official API, once on a heavily Q2-quantized local copy fitted onto a single DGX Spark (~1,000 tok/s prefill, 59 tok/s multi-agent serving). The cloud model one-shot a scroll-driven launch site he judged "definitely in line" with frontier output, recreated Karpathy's Three.js Lord of the Rings scene in 40 minutes — Opus 5 took two hours and roughly 1M tokens — and shipped a time-loop puzzle platformer whose mechanic he called conceptually more inventive than Kimi K3's or GPT Soul's, closer to Fable's. The local Q2 copy finished all three, crudely but functionally.
The size gap is the story. 284B total parameters against Kimi K3's 2.8T: roughly 10× smaller, ~8× fewer active parameters, 2.2× deeper layers, same context window. OpenRouter lists $0.09 in / $0.18 out versus K3's $2.90 / $14. The $0.53 run carried a temporary 90% promotional discount — even at list it lands near $5, against $25 for one K3 task. Arena.ai placed V4-Flash-20260731 (High) at #21 overall in Agent Arena across 12.5K real-world agentic sessions: 13 places above the prior Flash, #3 among open source, and the first local model on the top list. Vals AI calls it the cheapest model on its index scoring above 60, and 35× cheaper than the next best. Unsloth's quants run lossless 4-bit on 168GB RAM, 3-bit on 110GB.
Strategic Take
Take the reviewer's own verdict seriously: there is currently no economic case for running this locally, because the cloud price is already near zero. Local buys privacy, tinkering and immunity from the API disconnects he hit. Re-price your agent workloads before assuming frontier billing is the floor.
