HarnessTax — Terminal-Bench 2.0

Shown as HarnessTax (Arena) publishes it — their metrics, their models. A model name opens its page here; the arrow opens the original board.

Published by

HarnessTax (Arena)

Licence

not stated

Updates

frozen

Fetched

2026-10-05

Entries

21

Filter modelsOrder byTerminal-Bench 2.0Cost per rolloutMatched models only

21 scored

#	Model	Terminal-Bench 2.0 ↑	Cost per rollout ↓
1	GPT 5.6 SolPiOpenAI	83.3%71.1–94.4	$0.42
2	GPT 5.6 SolCodexOpenAI	78.9%65.6–91.1	$0.76
3	GPT 5.6 LunaPiOpenAI	76.7%62.2–90.0	$0.05
4	Claude Fable 5Claude CodeAnthropic	75.6%61.1–88.9	$1.55
5	Kimi K3PiKimi	73.3%60.0–85.6	$0.38
6	Claude Fable 5CodexAnthropic	72.2%56.7–86.7	$0.98
7	Claude Opus 4.8CodexAnthropic	72.2%56.7–86.7	$0.85
8	Claude Opus 4.8PiAnthropic	72.2%58.9–84.4	$0.76
9	GPT 5.6 LunaCodexOpenAI	72.2%57.8–85.6	$0.06
10	Claude Fable 5PiAnthropic	71.1%54.4–85.6	$1.08
11	GPT 5.6 SolClaude CodeOpenAI	71.1%55.6–85.6	$1.35
12	GPT 5.6 LunaClaude CodeOpenAI	70.0%55.6–83.3	$0.10
13	Kimi K3CodexKimi	70.0%55.6–83.3	$0.45
14	Claude Opus 4.8Claude CodeAnthropic	68.9%52.2–84.4	$0.90
15	Kimi K3Claude CodeKimi	66.7%52.2–80.0	$0.52
16	Claude Sonnet 4.6PiAnthropic	65.6%48.9–82.2	$0.61
17	Claude Sonnet 4.6CodexAnthropic	63.3%46.7–78.9	$0.55
18	Claude Sonnet 4.6Claude CodeAnthropic	62.2%46.7–76.7	$0.67
19	Claude Haiku 4.5PiAnthropic	47.8%32.2–63.3	$0.25
20	Claude Haiku 4.5Claude CodeAnthropic	41.1%27.8–55.6	$0.26
21	Claude Haiku 4.5CodexAnthropic	31.1%15.6–47.8	$0.21

What each column measures

Terminal-Bench 2.0

Higher is better · Agentic coding

Cost per rollout

Lower is better