Models · · 2 min read

DeepSeek V4 Flash scores 61.4% on ARC-AGI-2 at $0.042 a task — the same band as Kimi K3 for 1/38th the cost

ARC Prize published verified ARC-AGI results for DeepSeek V4 Flash 0731 on July 31, 2026: 89.0% on ARC-AGI-1 and 61.4% on ARC-AGI-2 at max reasoning effort, for $0.02 and $0.042 per task. Every other model in that 59-62% band on ARC-AGI-2 costs between $0.18 and $1.68 per task. Here are the exact numbers, what they do not tell you, and the routing decision they change.

ARC Prize published verified ARC-AGI results for DeepSeek V4 Flash 0731 on July 31, 2026. (Source: ARC Prize verified results, retrieved 2026-08-08)

Key facts:

  • At max reasoning effort, DeepSeek V4 Flash 0731 scores 61.4% on ARC-AGI-2 Semi-Private at $0.042 per task.
  • On ARC-AGI-1 Semi-Private it scores 89.0% at $0.02 per task.
  • Three reasoning variants were tested. Low scores 46.0% on ARC-AGI-2 at $0.021 per task; High scores 56.0% at $0.045; Max scores 61.4% at $0.042.
  • ARC-AGI-3 was not evaluated. The interactive-agent benchmark is blank for this model.
  • The frontier ceiling on ARC-AGI-2 is 92.5%, held by GPT-5.6 Sol (Max) at $1.44 per task. Claude Opus 5 (Max) is at 90.4% for $2.06. (Source: ARC Prize leaderboard, retrieved 2026-08-08)
ARC Prize verified scores table for DeepSeek V4 Flash 0731: Max variant 89.0% on ARC-AGI-1 and 61.4% on ARC-AGI-2; High 87.0% and 56.0%; Low 84.0% and 46.0%; ARC-AGI-3 not evaluated for any variant.
The verified scores table as published. Note that ARC-AGI-3 is blank for all three variants. (Table: ARC Prize, 2026-07-31)

The number that actually matters is the horizontal position

ARC Prize plots every verified system on cost-per-task against score. On that chart DeepSeek V4 Flash sits alone in empty space — roughly a full order of magnitude to the left of everything else at its score.

ARC Prize ARC-AGI-2 leaderboard scatter plot, log-scale cost per task on the x-axis and score on the y-axis. DeepSeek V4 Flash 0731 is highlighted in yellow at roughly $0.02 to $0.04 per task and 46% to 61.4%, far to the left of Kimi K3 (High), Gemini 3.6 Flash (Medium), Claude Opus 4.8 (High), Claude Fable 5 and GPT-5.6 Sol (Max), which cluster between $0.30 and $10 per task.
DeepSeek V4 Flash 0731 (yellow) against every other verified system on ARC-AGI-2. The x-axis is log-scale cost per task. (Chart: ARC Prize, 2026)

Here is the 59–62% band on ARC-AGI-2, sorted by cost. All numbers are from the ARC Prize leaderboard, retrieved 2026-08-08.

System (variant)ARC-AGI-2Cost / taskvs DeepSeek
DeepSeek V4 Flash 0731 (Max)61.4%$0.042—
GPT-5.6 Luna 2026-07-30 (Max)59.6%$0.1774.2×
Gemini 3.6 Flash (High)60.4%$0.60614×
GPT-5.6 Luna (Max, 2026-07-09 run)59.5%$0.67016×
Kimi K3 (Max)60.4%$1.5938×
Claude Opus 4.8 (Low)62.2%$1.6840×

What this means if you’re building with DeepSeek V4 Flash

1. This is a routing signal, not a “switch everything” signal. ARC-AGI measures abstract grid reasoning from a handful of examples. It does not measure agentic coding, tool use, or long-horizon work. DeepSeek V4 Flash’s ARC-AGI-3 row is empty, so there is no verified interactive-agent number for it at all. Do not read 61.4% as “close to Opus 5 for real work” — read it as “the cheap tier is now good enough to try first on structured reasoning.”

2. Effort setting is now a bigger lever than model choice. Inside DeepSeek V4 Flash, going from Low to Max costs 2× and buys +15.4 points on ARC-AGI-2 (46.0% → 61.4%). Going from DeepSeek Max to Kimi K3 Max costs 38× and buys −1.0 points. If you have a Flash-class model in your stack and you are running it at default effort, raising effort is the cheapest quality win available. The same pattern shows up in the Kimi K3 guide and in GLM-5.2’s local setup.

3. The two-tier fallback finally pencils out. At $0.042 a task you can run Flash at max effort, verify the answer, and escalate the ~38% it misses to a frontier model, and still land far under a frontier-only bill. Previously the cheap tier scored badly enough that the escalation rate ate the savings — Gemini 3.5 Flash-Lite (High), the other explicitly cheap tier ARC verified in July, scores 10.3% on ARC-AGI-2 at $0.139 per task. That is 6× lower score at 3.3× higher cost. (Source: ARC Prize, 2026-07-21)

4. Read the cost column as list price, not your bill. ARC’s figure is measured API spend during their test window at the prices in effect then. DeepSeek’s published rates are $0.14 per million input tokens and $0.28 per million output — see our DeepSeek V4 Flash guide for the cache-hit discount and the peak-hour multiplier that neither the ARC chart nor the sticker price shows.

One caveat on the methodology: the headline percentages are semi-private set results, while the per-task pass/fail tables on the same page are for the public eval sets (120 ARC-AGI-2 tasks, 400 ARC-AGI-1 tasks). They are different task pools. If you are trying to reproduce a specific task result, use the public tables; if you are comparing models, use the semi-private headline.

For the frontier side of this comparison, see our guides to Claude Opus 5 and Qwen3.8-Max’s official numbers, and our earlier writeup of DeepSeek V4 Pro.

Sources:

Source: ARC Prize (verified results)