ARK benchmark

ARK Capability 100

Measured across 100 capability tests.

Run date
September 30, 2026
Dataset version
v1
Concurrency
1
Lane
FULL
Failed requests
0

ARK100 score

Overall results

TierScore / 100AA Intelligence
ARK Lite (Thinking)62/100~7
ARK Flash (Thinking)75/100~16
ARK Pro (Thinking)93/100~34

ARK100 score chart

  • ARK LITE62 out of 100 correct
  • ARK FLASH75 out of 100 correct
  • ARK PRO93 out of 100 correct

Intelligence vs cost per task

Intelligence vs cost per task

AA Intelligence versus cost per taskHigher AA Intelligence and lower cost/task are more desirable. External cost/task values are published by Artificial Analysis. ARK cost/task values are estimates produced by repricing each tier's underlying AA workload to ARK's public token prices; they are not measured ARK AA runs. This scatter focuses on models with published cost per task at or below $0.30, so low-cost differences remain readable.AA Intelligence vs Cost per TaskAA Intelligence$0$0.10$0.20$0.30010203040Cost per task (USD)GPT-6 Luna (Non-reasoning) · Effort: non-reasoning · AA Intelligence: 18 · Cost / task: $0.0100GPT-6 Luna (Low) · Effort: low · AA Intelligence: 22 · Cost / task: $0.0045GPT-6 Luna (Medium) · Effort: medium · AA Intelligence: 30 · Cost / task: $0.0200GPT-6 Luna (High) · Effort: high · AA Intelligence: 33 · Cost / task: $0.0300GPT-6 Luna (Xhigh) · Effort: xhigh · AA Intelligence: 35 · Cost / task: $0.0400GPT-6 Luna (Max) · Effort: max · AA Intelligence: 38 · Cost / task: $0.0700DeepSeek V4.1 Flash (Non-reasoning) · Effort: non-reasoning · AA Intelligence: 25 · Cost / task: $0.1500DeepSeek V4.1 Flash (Max) · Effort: max · AA Intelligence: 39 · Cost / task: $0.2700MiMo-V2.5-Pro · Effort: reasoning · AA Intelligence: 26 · Cost / task: $0.0500ARK Lite (Thinking) · AA Intelligence: ~7 · Estimated cost / task: ~$0.0006 · AA cost basis: exact model/provider proxy (Llama 4 Scout) (approximate) · ARK Capability 100: 62/100ARK Flash (Thinking) · AA Intelligence: ~16 · Estimated cost / task: ~$0.0057 · AA cost basis: exact model cost (Nemotron 3 Ultra 550B A55B (Reasoning)) · ARK Capability 100: 75/100ARK Pro (Thinking) · AA Intelligence: ~34 · Estimated cost / task: ~$0.0173 · AA cost basis: exact model cost (Qwen3.8 27B (Xhigh)) · ARK Capability 100: 93/100
Higher AA Intelligence and lower cost/task are more desirable. External cost/task values are published by Artificial Analysis. ARK cost/task values are estimates produced by repricing each tier's underlying AA workload to ARK's public token prices; they are not measured ARK AA runs. This scatter focuses on models with published cost per task at or below $0.30, so low-cost differences remain readable.
ARK tiers and external comparators on one plot: cost per task runs along X from zero and AA Intelligence up Y. Effort variants of one family are joined by a line within that family only. ARK cost / task is an estimate; external cost / task is published by Artificial Analysis. This scatter focuses on models with published cost per task at or below $0.30, so low-cost differences remain readable.
ModelAA IntelligenceCost / task
ARK Lite (Thinking)~7~$0.0006
ARK Flash (Thinking)~16~$0.0057
ARK Pro (Thinking)~34~$0.0173
GPT-6 Luna (Non-reasoning)18$0.0100
GPT-6 Luna (Low)22$0.0045
GPT-6 Luna (Medium)30$0.0200
GPT-6 Luna (High)33$0.0300
GPT-6 Luna (Xhigh)35$0.0400
GPT-6 Luna (Max)38$0.0700
DeepSeek V4.1 Flash (Non-reasoning)25$0.1500
DeepSeek V4.1 Flash (Max)39$0.2700
MiMo-V2.5-Pro26$0.0500

ARK cost/task is an estimate obtained by repricing each underlying serving model's published AA cost/task by the ratio between ARK public input+output token price and that model's published AA input+output token price, then taking the same routing-value weighted tier average used for ARK's AA aggregate.

Domain breakdown

Domain heatmap

ARK100 domain results: percent answered exactly correctly, per domain and ARK tier.
DomainARK LITEARK FLASHARK PRO
Coding semantics86.67%93.33%93.33%
Debugging and root-cause analysis50%90%100%
Algorithms53.33%80%93.33%
Mathematics60%75%100%
Formal logic90%90%100%
Science reasoning50%70%90%
Data and statistics50%40%90%
Long-context reasoning60%60%100%
Multimodal40%40%40%

How this benchmark was run

Methodology

Questions
100
Run date
September 30, 2026
Concurrency
1
Lane
FULL
Failed requests
0
Full methodology
  • 100 questions, run September 30, 2026 at concurrency 1 on the FULL lane, with 0 failed requests.
  • Dataset SHA256 ef28dec9ee8ea055b16d4427e5a5a62bcec0e42a651f98d2721532dd951d560c (version v1), identical for every tier run.
  • Closed book: tools=false, web=false, llmJudge=false.
  • Scoring: exact match after deterministic type normalization, no LLM judge, no partial credit. 100 of 100 gold answers were independently verified first. Gold answers never enter an inference request.
  • Dataset file: evaluation/datasets/ark-capability-qa-v1.json.
  • Vision limitation: 4 image-bearing cases were omitted by perception in this run, so the multimodal row is a text-path result and is not published as a vision capability score.