Support

Benchmarks · Professional work · Agents & computer use

Corpo Index

Star0

14 consulting-style tasks, each ending in a finished deliverable such as a note, minutes, a workbook or a deck, marked out of 10 by two blind graders. The mean of two roles: the model working alone, and the model writing one work order for a fixed GPT-6 Luna worker. Each model runs in its own harness. From local-bench.ai.

Quality 68 of 100 · Private test set · Has measured 10 of today's 10 default lineup models

23 models · 23 results · updated 2026-11-19 · local-bench.ai ↗

Leaderboard

Opus 5.5 xhigh

Benchmark quality

How the score works
  • Overall
    68

    68 of 100

    How useful it is for telling today’s models apart

  • Headroom20%
    29

    Top models average 68 of 100

    Can top models still improve on it?

  • Test set20%
    100

    Private

    Could labs train on the questions?

  • Recency15%
    100

    36 days old

    How new is it?

  • Grading15%
    65

    Not researched

    How are answers scored?

  • New models15%
    100

    10 of 10 top models

    How fast does it test new releases?

  • Efforts10%
    0

    One effort per model

    Does it test each model at several reasoning efforts?

  • Breadth5%
    57

    23 models

    How many models has it measured?

How each part was worked out
Headroom 29 of 100 · 20% · +5.8
From the spread of published scores and the room left above the top quarter of models, one score per model.
Test set 100 of 100 · 20% · +20
Private test set. The 14 consulting-style tasks were built for the benchmark and are not published; the method page describes them but shows none. Source
Recency 100 of 100 · 15% · +15
Released 2026-10-17 (from the publisher's pages, not yet checked by hand).
Grading 65 of 100 · 15% · +9.8
Judged from its unit and description; the grading method has not been researched.
New models 100 of 100 · 15% · +15
Has measured 10 of today's 10 default lineup models. Too few new models to time how quickly it adds them.
Efforts 0 of 100 · 10% · 0
Measured 0 of the 23 models that come in more than one reasoning effort at two or more of them.
Breadth 57 of 100 · 5% · +2.9
23 models measured.

Not counted in the score: 0 stars · used by 1 published page. Both lift it in the directory’s recommended order. Test set, release date and grading were researched from the publisher’s pages (2026-11-19); tell us if one is wrong.

Score vs subscription cost per task

Highlight models

Click a point on the chart to highlight that exact effort; click it again to remove it.

11 more models under All models in settings.

More from this benchmark

1 published measures

All results

23 rows
#ModelEffortCorpo IndexSubscription $ / taskAPI $ / task
1Opus 5.5xhigh70.6 ±5.1$0.023$1.80
2GPT-6.2 Astraxhigh70.1 ±6.2$0.051$0.76
3GPT-6.2 Solxhigh67.2 ±5.8$0.008$0.24
4Fable 5.2xhigh66.9 ±6.8$0.19$5.18
5Sonnet 5.5xhigh66.3 ±5.7$0.012$0.89
6GPT-6.1 Solxhigh66.0 ±6.3$0.009$0.28
7GPT-6 Astraxhigh65.4 ±6.7$0.047$1.32
8GPT-6 Solxhigh63.2 ±5.3$0.009$0.28
9Gemini 4 Carbonhigh63.0 ±5.5$0.15$2.54
10Gemini 4 Neonxhigh62.5 ±6.5$0.035$1.66
11Fable 5.1xhigh62.4 ±5.4$0.52$3.35
12Gemini 4 Hydrogenxhigh61.3 ±4.9$0.016$0.77
13Haiku 5.5xhigh61.0 ±6.4$0.004$0.22
14GPT-6.1 Terraxhigh60.1 ±5.4$0.002$0.060
15DeepSeek V4.1 Flashhigh59.4 ±5.8$0.012$0.12
16GPT-6 Lunaxhigh56.2 ±4.6$0.001$0.020
17Hy4 previewhigh55.1 ±4.5$0.13$0.38
18GLM 5.3 Flashhigh54.8 ±5.0$0.007$0.040
19Muse Emberxhigh53.9 ±4.7$0.040$0.040
20MiniMax M3.1xhigh53.7 ±5.0$0.045$0.090
21Hy4 Flashhigh52.4 ±4.7$0.010$0.010
22Sonnet 5xhigh50.9 ±5.2$0.010$0.98
23Opus 4.6max49.1 ±5.6$1.26$1.26