The submission links to ARC Prize results for DeepSeek V4 Flash 0731, a late-July update to DeepSeek’s smaller, cheaper V4 Flash line. On the benchmark page it scores in the neighborhood of much pricier frontier offerings, which got attention because this is an open-weight model with unusually low serving costs. That immediately turned the conversation from benchmark bragging rights into workflow economics. People using coding harnesses like Pi, Oh My Pi, OpenCode, Codex, and Prime Agent said the model is now good enough for day-to-day programming, debugging, and document analysis, while being cheap enough to run many concurrent sessions without the usual quota anxiety. Several said that for ordinary software work the difference versus Claude or GPT-class models is smaller than the price gap suggests, so they save the expensive models for the hardest tasks and use DeepSeek as the default executor.
The price story turned out to be more specific than the headline chart implies. The standout advantage is not just low input and output token pricing. It is DeepSeek’s extremely cheap cached input reads, which dominate long coding sessions and agent loops where the same context is resent over and over. Multiple people posted numbers showing 90 to 98 percent cache hit rates, which makes DeepSeek far cheaper in practice than providers serving the same
open weights through
OpenRouter or other endpoints. That is also why the announced upcoming
API price increase did not cause panic so much as questions about what exactly will rise. Some expect the increase is mainly load shedding or the end of an intro discount. Others think DeepSeek’s caching system is the real moat and will remain hard to match even if raw token pricing converges.
A second theme was that this release feels materially better than the earlier preview, not just cheaper. Users running it locally on two
RTX Pro 6000 Blackwell GPUs reported enough throughput to make a roughly 300B-parameter
mixture-of-experts model feel interactive, which changes the local inference tradeoff. Instead of paying for larger hosted models, some said they now route most daily work to a local DeepSeek setup and escalate only when they need more intelligence or vision. That fed a broader conclusion running through the comments: the market is splitting. Frontier models still matter for long-horizon, high-stakes tasks, but there is a huge category of short-horizon, repetitive, high-volume work where “good enough and almost free” is the more important threshold.
The enthusiasm was not universal. A few people reported looping, random digressions, weak tool use, or token-heavy behavior, especially in certain harnesses. Others pushed back on ideas like automatic
CI fixes or blanket test generation, arguing that cheap inference can still create expensive review overhead if the model is inconsistent. The more grounded view that emerged was that V4 Flash 0731 is not a license to automate everything. It is a strong default building block. Use it where low marginal cost, high concurrency, and easy self-hosting matter more than absolute best-in-class reasoning, and keep stronger models in reserve for the tasks where small errors compound into real damage.
arcprize.org