DeepSeek V4 Flash：ARC-AGI达61.4%且成本仅为GPT-5.6的1/4

Subtitle: ARC-AGI-2 结果页显示 DeepSeek V4 Flash（max）在约 $0.04/任务 的成本下达到 61.4%，而 GPT-5.2（medium）以 $0.759 获得 26.7%——在对数尺度上这是 19 倍的成本-性能差距，实际效果比数字看起来更令人印象深刻。这条 HN 讨论共 454 条评论，与其说是在谈基准测试本身，不如说是在讨论「接近前沿的模型近乎免费」后哪些用例变得合理。以下是开发者们实际说了什么，以及这对你的构建方式意味着什么。

The ARC-AGI results page dropped DeepSeek V4 Flash 0731 into the top-left corner of the cost-vs-accuracy chart — comparable to GPT-5.6 Luna on score, roughly 1/4 the price on the raw numbers (and the x-axis is log-scaled, so the visual gap understates it). Hacker News gave it 756 points and 454 comments in two days. I read the whole thread so you don't have to: the benchmark is the hook, but the real signal is how fast "frontier-adjacent at commodity prices" is changing what developers build.

The chart, read honestly

The numbers people keep quoting:

• DeepSeek V4 Flash (max reasoning): 61.4% on ARC-AGI II at ~$0.04/task
• GPT-5.2 (medium): 26.7% at $0.759/task — the "remember when this was everyone's favorite" data point
• Kimi K3: comparable performance to a month ago, at a fraction of its price — a $/task line that moved 20x in a month

Two caveats the thread raised that are worth keeping:

Price is a moving target. DeepSeek emailed customers warning of a "significant" price increase. Counterpoint from the thread: it's an open-weights model served by 24+ providers on OpenRouter, so the market price is competitive — the official API hike mostly raises the ceiling, not the floor.

ARC-AGI-2 isn't ARC-AGI-3. Someone asked why the run wasn't against ARC-AGI-3; the answer is those results take longer to run. The benchmark is useful, but it's one slice.

Also note: max reasoning costs less than high on this model — the opposite of what you'd expect — which fed the thread's favorite joke about rework being the most expensive token.

The pattern HN actually cares about: new use cases

The most upvoted discussions weren't about the benchmark — they were about what becomes worth doing when a capable model is effectively free:

"Too cheap to meter" as a design principle. Multiple commenters described using DeepSeek V4 Flash for everything from CI test-failure triage (auto-generate a proposed fix on every failing PR) to re-ranking social feeds. The recurring line: "even if it's not SOTA, 99.99% of people don't need SOTA."

The local-run crowd. People are running it on laptops, on 2×DGX boxes, on 4×RTX Pro 6000s (~250 tok/s decode, ~2,400 aggregated tok/s across 48 concurrent requests), and even quantized GGUF Q8 on a 256GB DRAM server with no GPU. Simon Willison's comment — "a pretty great score for a model you can run on a (expensive) laptop" — captured the mood.

Harness-hopping is the new model-hopping. The thread is full of people comparing how it behaves in OpenCode vs. Prime Agent vs. Pi vs. Codex — one commenter noted it's "trained in the codex harness" and feels as good as OpenAI models at tool use, at extreme cheap with 1M context. The takeaway: with models this interchangeable on price, the harness is the differentiator, which is exactly why every agent platform launched in the last month is fighting for that layer.

The failure mode is real too. A meaningful minority reported infinite reasoning loops, tool-call stalls, and "talking to itself" — mostly in Pi, mitigated with stronger prompting or switching harnesses. The honest summary from the thread: both camps are right — it's a big step up at the price, and it still fails in frustrating ways on complex tasks.

The economics that matter most

Two data points summarize the shift:

ARC-AGI II: GPT-5.2 (medium) 26.7% at $0.759 vs. DSV4-Flash (max) 61.4% at $0.04. That's a 19x cost-performance gap against a model from roughly a year ago.

Cache economics are the hidden moat. DeepSeek's cache-read prices are ~1/10 of other providers ($0.0028/M tokens), and commenters report 99% cache-hit rates with cache-aware tooling — meaning real-world costs land far below list price for agentic workloads with repeated context.

The infrastructure takeaway: when intelligence costs this little, the scarce resource stops being tokens and becomes orchestration — routing, caching, loop-detection, and state management. The models are becoming interchangeable; the software around them is where the value concentrates.

What this means if you build with AI tools

Re-run your "too expensive to automate" list. Anything you shelved because per-task cost was prohibitive — bulk triage, draft generation, test repair, log analysis — is now rational at Flash-class pricing. The use-case frontier moved, not just the benchmark.

Budget for harness variance. Same model, wildly different behavior across Pi, OpenCode, Codex, and Prime Agent. Test your agent workflow in at least two harnesses before committing.

Watch for the price correction. The official API is announcing a significant hike; if your cost model assumes today's prices, build a tolerance for a 2-5x line item, or pin to OpenRouter's competitive market.

Cache-aware routing is the new performance tuning. With cache reads at a tenth of the rate, tools that maximize cache hits (like Reasonix, per the thread) can cut your bill by an order of magnitude on agent workloads.

The ARC-AGI chart is the hook, but the thread's real finding is the "floor rising" argument — the capability level available open and cheap is now high enough that the default assumption for builders should be "use a cheap good model for everything, escalate to frontier only when measured need appears." That inverts the 2023-2024 calculus and it's happening faster than the discourse has caught up.

The models will keep trading places on the chart. The durable shift — intelligence cheap enough to route, cache, and orchestrate like any other compute — is what actually changes how we build.

I review AI developer tools hands-on at toolsdepth.com — including full pricing breakdowns, HN-debate summaries, and harness comparisons for DeepSeek V4 Flash, Kimi K3, and the current agent-platform wave.