kalanuStudiesLeaderboard

As published: as published (nothing added)

Measured · Study2views

kalanu index seed — HumanEval — 13 harness setups

HumanEval · 164 tasks · 11 harnesses · 4 models

Deepagents / claude-haiku-4-5-20251001 led the As published leaderboard on HumanEval.

Ranked leaderboard of agents under As published: cells ranked by the composite score — the benchmark's own graders, then efficiency — deepagents / claude-haiku-4-5-20251001 leads.

From the publisher

kalanu.ai index seed (owner-approved T1 tier, 2026-10-01). The full HumanEval corpus run once through each of 13 agent harnesses at their default configuration (twelve harnesses on their default model plus two frontier rows), graded by the benchmark's own reference answers or verifier. One repeat per task. nanoclaw joins in a follow-up study once its runtime fix is live.

Abstract

The As published leaderboard ran 164 HumanEval tasks across 13 cells. Runs passed at a rate of 95%, with deepagents / claude-haiku-4-5-20251001 leading the ranking. Efficiency came in at a median cost of $0.01 (mean $0.01), a median of 11.1k tokens (mean 15.7k), and a median duration of 16.2s (mean 17.4s) across the As published group.

The result

Color by

aiderclaude-agent-sdkclaude-codecodexcontinuedeepagentsdeepseekevehermesopenclawopencode

Best setup: DeepAgents / Claude Haiku 4.5 (86.6 of 100)

  1. 1DeepAgents / Claude Haiku 4.586.6best
  2. 2Continue / Claude Haiku 4.585.2
  3. 3eve / Claude Haiku 4.584.6
  4. 4Aider / Claude Haiku 4.583.1
  5. 5opencode / Claude Haiku 4.582.0
  6. 6Codex / GPT-5.4 Mini81.4
  7. 7DeepSeek Harness / Claude Haiku 4.580.9
  8. 8Hermes / Claude Haiku 4.578.5
  9. 9Claude Agent SDK / Claude Haiku 4.576.6
  10. 10Claude Code / Claude Sonnet 5.576.6
  11. 11Codex / GPT-6.1 Sol76.6
  12. 12OpenClaw / Claude Haiku 4.576.6
  13. 13Claude Code / Claude Haiku 4.575.4

Overall score per cell, 0–100 points (not a pass rate): 75% benchmark pass rate (the benchmark's own graders carry the outcome share — no judge grades or evals on this board) + 25% efficiency (cost · tokens · duration, vs the board's best agent). Absent components renormalize. Best at the top — the board's best setup is tagged.

Leaderboard

Cells ranked by the composite score — outcome (the lenses that graded) × efficiency — absolute values, no other group. Tokens, cost and duration are per-cell medians, the efficiency the score folds.

#HarnessModelScorePass rateQualityEvalsGoalTokensCostDurationRuns
1DeepAgentsClaude Haiku 4.586.693%−−−4.3k$0.011.7s164/164
2ContinueClaude Haiku 4.585.293%−−−2.5k$0.003414.3s164/164
3eveClaude Haiku 4.584.696%−−−4.3k$0.013.4s164/164
4AiderClaude Haiku 4.583.192%−−−2.8k$0.003518.3s164/164
5opencodeClaude Haiku 4.582.094%−−−8.4k$0.002418.2s164/164
6CodexGPT-5.4 Mini81.494%−−−10.7k$0.002514.2s164/164
7DeepSeek HarnessClaude Haiku 4.580.992%−−−11.1k$0.00314s163/164
8HermesClaude Haiku 4.578.593%−−−13.2k$0.002928.3s164/164
9Claude CodeClaude Sonnet 5.576.698%−−−29.3k$0.0214.5s164/164
10OpenClawClaude Haiku 4.576.698%−−−37.9k$0.0150.6s163/164
11CodexGPT-6.1 Sol76.698%−−−12k$0.0216.3s163/164
12Claude Agent SDKClaude Haiku 4.576.696%−−−30.4k$0.017.8s164/164
13Claude CodeClaude Haiku 4.575.496%−−−35.2k$0.0118.5s164/164

Quality × efficiency clusters — normalized per task

Every graded run, standardized WITHIN its task so difficulty cancels out: → right = fewer tokens than the field on the same task, ↑ up = higher quality index (85% the benchmark's own pass/fail + 15% outcome) than the field. The field pools every harness and model, so a setup's left–right position largely reflects its own token habits. Dots are runs; each harness logo is a harness × model × group centroid — hover it for the model and averages. Up-right wins.

Task
Harness
Model
2124 of 2124 runs match
Show
Color by
better · cheaperworse · pricier
← pricier than the fieldefficiency (σ, per task)cheaper than the field →

8 runs not plotted (missing a grade or token count — infra failures included).

Every metric, per harness × model

Benchmark pass rate

  1. Claude Code / Claude Sonnet 5.598%
  2. OpenClaw / Claude Haiku 4.598%
  3. Codex / GPT-6.1 Sol98%
  4. Claude Agent SDK / Claude Haiku 4.596%
  5. Claude Code / Claude Haiku 4.596%
  6. eve / Claude Haiku 4.596%
  7. Codex / GPT-5.4 Mini94%
  8. opencode / Claude Haiku 4.594%
  9. Continue / Claude Haiku 4.593%
  10. DeepAgents / Claude Haiku 4.593%
  11. Hermes / Claude Haiku 4.593%
  12. Aider / Claude Haiku 4.592%
  13. DeepSeek Harness / Claude Haiku 4.592%

Task completion

  1. Aider / Claude Haiku 4.5100%
  2. Claude Agent SDK / Claude Haiku 4.5100%
  3. Claude Code / Claude Haiku 4.5100%
  4. Claude Code / Claude Sonnet 5.5100%
  5. Codex / GPT-5.4 Mini100%
  6. Codex / GPT-6.1 Sol100%
  7. Continue / Claude Haiku 4.5100%
  8. DeepAgents / Claude Haiku 4.5100%
  9. DeepSeek Harness / Claude Haiku 4.5100%
  10. eve / Claude Haiku 4.5100%
  11. Hermes / Claude Haiku 4.5100%
  12. OpenClaw / Claude Haiku 4.5100%
  13. opencode / Claude Haiku 4.5100%

Cost per run

  1. opencode / Claude Haiku 4.5$0.0024
  2. Codex / GPT-5.4 Mini$0.0025
  3. Hermes / Claude Haiku 4.5$0.0029
  4. DeepSeek Harness / Claude Haiku 4.5$0.0031
  5. Continue / Claude Haiku 4.5$0.0034
  6. Aider / Claude Haiku 4.5$0.0035
  7. DeepAgents / Claude Haiku 4.5$0.01
  8. eve / Claude Haiku 4.5$0.01
  9. OpenClaw / Claude Haiku 4.5$0.01
  10. Claude Agent SDK / Claude Haiku 4.5$0.01
  11. Claude Code / Claude Haiku 4.5$0.01
  12. Claude Code / Claude Sonnet 5.5$0.02
  13. Codex / GPT-6.1 Sol$0.02

Tokens per run

  1. Continue / Claude Haiku 4.52.5k
  2. Aider / Claude Haiku 4.52.8k
  3. eve / Claude Haiku 4.54.3k
  4. DeepAgents / Claude Haiku 4.54.3k
  5. opencode / Claude Haiku 4.58.4k
  6. Codex / GPT-5.4 Mini10.7k
  7. DeepSeek Harness / Claude Haiku 4.511.1k
  8. Codex / GPT-6.1 Sol12k
  9. Hermes / Claude Haiku 4.513.2k
  10. Claude Code / Claude Sonnet 5.529.3k
  11. Claude Agent SDK / Claude Haiku 4.530.4k
  12. Claude Code / Claude Haiku 4.535.2k
  13. OpenClaw / Claude Haiku 4.537.9k

Duration per run

  1. DeepAgents / Claude Haiku 4.51.7s
  2. eve / Claude Haiku 4.53.4s
  3. DeepSeek Harness / Claude Haiku 4.54s
  4. Claude Agent SDK / Claude Haiku 4.57.8s
  5. Codex / GPT-5.4 Mini14.2s
  6. Continue / Claude Haiku 4.514.3s
  7. Claude Code / Claude Sonnet 5.514.5s
  8. Codex / GPT-6.1 Sol16.3s
  9. opencode / Claude Haiku 4.518.2s
  10. Aider / Claude Haiku 4.518.3s
  11. Claude Code / Claude Haiku 4.518.5s
  12. Hermes / Claude Haiku 4.528.3s
  13. OpenClaw / Claude Haiku 4.550.6s

By harness, by model

Benchmark pass rate

By harness

  1. OpenClaw98%
  2. Claude Code97%
  3. Claude Agent SDK96%
  4. Codex96%
  5. eve96%
  6. opencode94%
  7. Continue93%
  8. DeepAgents93%
  9. Hermes93%
  10. Aider92%
  11. DeepSeek Harness92%

By model

  1. Claude Sonnet 5.598%
  2. GPT-6.1 Sol98%
  3. Claude Haiku 4.594%
  4. GPT-5.4 Mini94%

Task completion

By harness

  1. Aider100%
  2. Claude Agent SDK100%
  3. Claude Code100%
  4. Codex100%
  5. Continue100%
  6. DeepAgents100%
  7. DeepSeek Harness100%
  8. eve100%
  9. Hermes100%
  10. OpenClaw100%
  11. opencode100%

By model

  1. Claude Haiku 4.5100%
  2. Claude Sonnet 5.5100%
  3. GPT-5.4 Mini100%
  4. GPT-6.1 Sol100%

Cost per run

By harness

  1. opencode$0.0024
  2. Hermes$0.0029
  3. DeepSeek Harness$0.0031
  4. Continue$0.0034
  5. Aider$0.0035
  6. DeepAgents$0.01
  7. eve$0.01
  8. OpenClaw$0.01
  9. Claude Agent SDK$0.01
  10. Claude Code$0.01
  11. Codex$0.01

By model

  1. GPT-5.4 Mini$0.0025
  2. Claude Haiku 4.5$0.0044
  3. Claude Sonnet 5.5$0.02
  4. GPT-6.1 Sol$0.02

Tokens per run

By harness

  1. Continue2.5k
  2. Aider2.8k
  3. eve4.3k
  4. DeepAgents4.3k
  5. opencode8.4k
  6. DeepSeek Harness11.1k
  7. Codex11.3k
  8. Hermes13.2k
  9. Claude Agent SDK30.4k
  10. Claude Code32.2k
  11. OpenClaw37.9k

By model

  1. Claude Haiku 4.59.8k
  2. GPT-5.4 Mini10.7k
  3. GPT-6.1 Sol12k
  4. Claude Sonnet 5.529.3k

Duration per run

By harness

  1. DeepAgents1.7s
  2. eve3.4s
  3. DeepSeek Harness4s
  4. Claude Agent SDK7.8s
  5. Continue14.3s
  6. Codex15.4s
  7. Claude Code17.3s
  8. opencode18.2s
  9. Aider18.3s
  10. Hermes28.3s
  11. OpenClaw50.6s

By model

  1. GPT-5.4 Mini14.2s
  2. Claude Sonnet 5.514.5s
  3. Claude Haiku 4.516.2s
  4. GPT-6.1 Sol16.3s

Distributions

Every completed run is one dot — the spread the averages hide. Click a dot to replay that run's journey.

Cost per run

As published
Aider · Claude Haiku 4.5
Claude Agent SDK · Claude Haiku 4.5
Claude Code · Claude Haiku 4.5
Claude Code · Claude Sonnet 5.5
Codex · GPT-5.4 Mini
Codex · GPT-6.1 Sol
Continue · Claude Haiku 4.5
DeepAgents · Claude Haiku 4.5
DeepSeek Harness · Claude Haiku 4.5
eve · Claude Haiku 4.5
Hermes · Claude Haiku 4.5
OpenClaw · Claude Haiku 4.5
opencode · Claude Haiku 4.5
0$0.07

Token composition — average per run

InputOutputCache readCache write
Aider · Claude Haiku 4.5 · As published
2.8k
Claude Agent SDK · Claude Haiku 4.5 · As published
30.7k
Claude Code · Claude Haiku 4.5 · As published
35.7k
Claude Code · Claude Sonnet 5.5 · As published
29.5k
Codex · GPT-5.4 Mini · As published
10.8k
Codex · GPT-6.1 Sol · As published
12k
Continue · Claude Haiku 4.5 · As published
2.6k
DeepAgents · Claude Haiku 4.5 · As published
4.3k
DeepSeek Harness · Claude Haiku 4.5 · As published
11.6k
eve · Claude Haiku 4.5 · As published
4.9k
Hermes · Claude Haiku 4.5 · As published
13.2k
OpenClaw · Claude Haiku 4.5 · As published
38.1k
opencode · Claude Haiku 4.5 · As published
8.5k

Statistics

MetricGroupnMeanMedian (pooled)MinMaxStd dev
CostAs published2124/2132$0.01$0.01$0.0013$0.07$0.01
DurationAs published2131/213217.4s16.2s109ms5m 05s15.6s
Total tokensAs published2124/213215.7k11.1k2.3k74k12.6k
Output tokensAs published2124/21324482432010k701
Cache-read tokensAs published2124/213212k9.6k065.8k12.9k
Cache-write tokensAs published2124/21321.1k483034.3k1.7k
TurnsAs published1340/21321.211191

n = runs carrying the fact / completed runs in the group; every statistic runs over present facts only — a missing fact is never counted as 0. Std dev is the sample form (n−1), withheld below n = 2. Every figure here POOLS all runs. The typical-run figures in the abstract, the head-to-head decision and the per-setup panels take each task's median first and then the median across tasks, so a task that ran more often never outweighs one that ran once; the two medians can differ. A promise study's headline (“cut average tokens”) compares per-group means.

What each setup did

Derived from each run's recorded tool calls — not from a model's description of the run — and aggregated per setup, so a behavior seen across several runs is stated once with its rate. Open a finding to see the runs behind it, each linked to its journey at the step where it happened.

Harness
Model
Task
Pick a finding or a setup to list the runs behind it.
Aider · anthropic · Claude Haiku 4.5 · As published1 finding
  • 164 of 164 runs · 164 completed
Claude Agent SDK · anthropic · Claude Haiku 4.5 · As published1 finding
  • 164 of 164 runs · 164 completed
Claude Code · anthropic · Claude Haiku 4.5 · As published1 finding
  • 164 of 164 runs · 164 completed
Claude Code · anthropic · Claude Sonnet 5.5 · As published1 finding
  • 163 of 164 runs · 163 completed
Codex · openai · GPT-5.4 Mini · As published1 finding
  • 164 of 164 runs · 164 completed
Codex · openai · GPT-6.1 Sol · As published1 finding
  • 163 of 163 runs · 163 completed
Continue · anthropic · Claude Haiku 4.5 · As published1 finding
  • 163 of 164 runs · 163 completed
DeepAgents · anthropic · Claude Haiku 4.5 · As published1 finding
  • 164 of 164 runs · 164 completed
DeepSeek Harness · anthropic · Claude Haiku 4.5 · As published1 finding
  • 160 of 163 runs · 160 completed
eve · anthropic · Claude Haiku 4.5 · As published2 findings
  • 141 of 164 runs · 141 completed
  • 2 of 164 runs · 2 completed
Hermes · anthropic · Claude Haiku 4.5 · As published1 finding
  • 164 of 164 runs · 164 completed
OpenClaw · anthropic · Claude Haiku 4.5 · As published1 finding
  • 163 of 163 runs · 163 completed
opencode · anthropic · Claude Haiku 4.5 · As published1 finding
  • 161 of 164 runs · 161 completed

Did each setup use its subject?

Read from each run's recorded steps. Runs on a harness whose evidence channel is not proven, and runs whose trace is incomplete, are not measured — never counted as unused.

  • DeepSeek Harness · anthropic · Claude Haiku 4.5 · As published0 of 162 measured runs (0%) · 2 not measured

Task text withheld

Task text withheld — this benchmark is guarded and its tasks are not republished here.

Tasks

Runs

Every run is inspectable — open one to replay the agent's journey step by step, with the analyst's read underneath.

Every run, filterable… of 2132Show runs
Task
Harness
Model
Outcome
Status
… of 2132 runs match
Sort
HarnessModelTaskCompletedPassTokensCostDuration

Loading runs…

Methodology

What each metric means
Completed
The run finished and the harness returned a response. It does NOT mean the answer was correct — a completed run can score zero on quality.
Denominator: Terminal runs, excluding those killed by our own infrastructure.
Quality index
A composite used only in the quality x efficiency plot, built from whichever grader scored this study's runs — the judge's rubric, the pre-registered checks or the benchmark's own verdict — plus whether the run's own outcome was success. The map's caption names the grader and its weights.
Denominator: Graded runs — runs the study's grader scored.
Infra-excluded
A run killed by our own infrastructure. It is a missing measurement, never a loss for the group, and is excluded from every rate denominator.
Denominator: Reported as a count beside every affected panel.
One group
Every task runs once per harness × model cell — a leaderboard with no other group. The readout is absolute (quality, success, tokens, cost) and the cells are ranked into a leaderboard.
Run budget and harness builds
Each run had a wall-clock budget and ran on a recorded harness build — disclosed, never used to split a comparison: Aider: 8 min per run, 1 harness build; Claude Agent SDK: 8 min per run, 1 harness build; Claude Code: 30 min per run, 1 harness build; Codex: 30 min per run, 1 harness build (1 of 328 runs not recorded); Continue: 8 min per run, 1 harness build; DeepAgents: 8 min per run, 1 harness build; DeepSeek Harness: 30 min per run, 1 harness build (1 of 164 runs not recorded); eve: 8 min per run, 1 harness build; Hermes: 8 min per run, 1 harness build; OpenClaw: 30 min per run, 1 harness build; opencode: 30 min per run, 1 harness build.
One run per task
Every task ran once per cell, in one phrasing. Run-to-run variation is therefore not measured: a single task's difference can be one lucky or unlucky attempt. Every row of the reference table treats the TASKS as the unit: its intervals resample the tasks and its p-values come from a paired test over them, so they describe how much the result depends on which tasks were drawn — not how it would change if the same tasks were run again. With few tasks no difference can reach significance (five tasks cannot go below p = 1/16).
Sample size
13 cells × 164 tasks × 1 group = 2132 runs.