DeepSeek V4 Flash 0731 achieves 89% ARC‑AGI‑1 and 61.4% ARC‑AGI‑2 at $0.02‑$0.04 per task

Overview

DeepSeek V4 Flash 0731 delivers top‑tier performance on the ARC‑AGI benchmark suite—89.0% on ARC‑AGI‑1 (Semi‑Private) and 61.4% on ARC‑AGI‑2 (Semi‑Private)—at a per‑task cost of $0.02 and $0.04 respectively. These numbers place the model among the highest‑scoring systems while keeping inference expenses an order of magnitude lower than many proprietary alternatives.

Benchmark Results

The official ARC‑AGI leaderboard reports three reasoning variants for DeepSeek V4 Flash 0731:

Variant	ARC‑AGI‑1	ARC‑AGI‑2	ARC‑AGI‑3
Max	89.0%	61.4%	—
High	87.0%	56.0%	—
Low	84.0%	46.0%	—

The "Max" variant, which uses the most aggressive reasoning settings, outperforms the "High" and "Low" configurations across both benchmarks. The model does not have a reported ARC‑AGI‑3 score at the time of writing.

Task‑level Pass/Fail (ARC‑AGI‑2 Public Eval)

Out of 120 public evaluation tasks, the Max variant passes 73 tasks, the High variant passes 71, and the Low variant passes 58. The detailed pass/fail matrix shows consistent strengths on tasks involving logical deduction and code reasoning, while a subset of visual‑puzzle‑style tasks remain challenging.

Cost Efficiency

DeepSeek V4 Flash 0731 charges $0.02 per ARC‑AGI‑1 task and $0.04 per ARC‑AGI‑2 task under the "max effort" pricing tier. This translates to roughly $0.02‑$0.04 per 1 M tokens when using the cached pricing model referenced by the provider. Community members repeatedly highlight the negligible operational cost:

"I'm running it in Oh My Pi with a second instance running as "advisor" and even with 5‑6 active sessions I'm struggling to spend more than 5 bucks per day." – @LaurensBER

"DeepSeek is my cheap and cheerful Chinese model of choice for API use. Even cheaper, and now better than Pro. ... Not the best, but stronger across the board than any model within an order of magnitude of its price." – @SwellJoe

Real‑World Usage Insights

Speed and Deployability

Users report that the model runs at ~8 k tokens/s prefill and ~250 tokens/s generation on a dual‑RTX 6000 Blackwell system, scaling to ~1 k tokens/s with 64 concurrent streams via vLLM. This throughput enables interactive chat without noticeable latency and supports large‑scale agent deployments.

"On 2x RTX Pro 6000 Blackwell, its ~8k tok/s prefill and ~250 tok/s on a single stream. I saw 1000 tok/s with ~64 concurrent streams on vLLM. That's fast enough that you can interactively chat with it without switching tabs while you wait." – @ak_t

Programming and Tool Use

The model is praised for its coding assistance and tool‑calling capabilities, often outperforming larger proprietary models in day‑to‑day development workflows.

"I strongly recommend trying this for programming tasks. It is strong (not Fable strong though) with a much better "persona" than Opus, and very different blindspots. ... On balance I actually prefer DeepSeek for programming now, because of the way it talks." – @mosura

"One of the best things about this version is that it is trained in the codex harness. It feels just as good as OpenAI models in using codex tools, but extremely cheap and with 1M context." – @tarruda

Low‑Resource Accessibility

Even the quantized GGUF Q8 K XL variant fits comfortably on a 256 GB DRAM server, allowing users without GPUs to run the model overnight on CPU‑only hardware.

"If you have no GPU at all and are okay with setting up a workflow that handles slow token per second rate, give it a task and check back in 4‑6 hours, it works great." – @walrus01

Performance Caveats

While the benchmark scores are impressive, several users note regressions and edge‑case failures:

Infinite loops and self‑talk: Some users experienced the model entering repetitive loops or generating unrelated content when used with certain agents.
Verbosity shift: A recent update appears to have increased token usage and made responses more verbose, affecting cost and conciseness.
Visual‑puzzle reasoning: Despite strong ARC‑AGI‑2 scores, the model lacks multimodal inputs, raising questions about how it solves visual puzzles purely via text reasoning.

"I've had tons of issues with it getting in infinite loops and talking to itself without executing tool calls, wasting tons of tokens... It's still worth it but the agentic performance has degraded in my experience." – @nylonstrung

"The model became a lot more pretentious and over‑explanatory. Heavy prompt reworking helped but maybe that's just the cost of being better at coding and ARC‑AGI?" – @momojo

Community Sentiment

Overall sentiment on Hacker News is strongly positive, with the post receiving 779 upvotes and 466 comments. Users celebrate the model’s price‑to‑performance ratio, its local deployability, and its utility for automation (e.g., CI test generation, log monitoring, social‑media feed re‑ranking). A minority express disappointment over inconsistent behavior and potential upcoming price hikes.

"DeepSeek has announced an upcoming "significant increase" in price, so this line may have to move to the right soon." – @modeless

"Kimi K3 was an interesting model only a month ago, and now we're looking at the same performance for 1/20th of the price. Wild how fast this is advancing." – @542458

Conclusion

DeepSeek V4 Flash 0731 sets a new benchmark for high‑accuracy, low‑cost LLM inference in mid‑2026. Its 89.0% ARC‑AGI‑1 and 61.4% ARC‑AGI‑2 scores demonstrate competitive reasoning ability, while its sub‑cent per‑task pricing, fast CPU/GPU throughput, and strong coding assistance make it a practical choice for both hobbyists and enterprise agents. Users should remain aware of occasional regressions and monitor upcoming pricing changes, but the model currently represents one of the most cost‑effective pathways to near‑state‑of‑the‑art language‑model performance.