AI HOT (Curated Pool)Sep 24, 2026, 09:24 UTC+8

Claude Opus 5.5 tops Coding Agent Index, but per-task cost rises to $13.04

Artificial Analysis：Claude Opus 5.5 登顶 Coding Agent Index，但单任务成本升至 $13.04

Artificial Analysis tested Claude Opus 5.5 under Claude Code max effort and it scored 66 on the Coding Agent Index, up from Opus 5's 60. All three subtests improved: Terminal-Bench 4.0 63.1%, DeepSWE v1.1 68.4%, SWE-Atlas-QnA 66.4%. The trade-off: per-task cost jumped from $3 to $13.04. The post doesn't break down how max effort drove the cost increase.

Why it matters: Claude Opus 5.5 tops the Coding Agent Index with a 6-point jump to 66, but $13.04 per task is the hard number. Anthropic substantive update + independent third-party benchmark + concrete data — all three HKR axes hit. Not scoring higher because this is a single benchmark, not ...

Read the original ↗Export Markdown

Take

The score-vs-cost gap is what makes this worth clicking. Opus 5.5 under Claude Code max effort hit 66 on the Coding Agent Index, up from Opus 5's 60, with all three subtests improving — Terminal-Bench 4.0 at 63.1%, DeepSWE v1.1 at 68.4%, SWE-Atlas-QnA at 66.4%. But per-task cost quadrupled from $3 to $13.04.

I'd discount this a bit because the summary doesn't explain what "max effort" actually means — more self-correction loops, a longer thinking budget, or parallel agent runs with a best-of-N pick. Each of those drives cost differently. If the 6-point gain came from burning way more tokens, the cost-efficiency is actually worse.

For reference: Opus 5 got 60 points for $3, Opus 5.5 gets 66 for $13.04. That's steep marginal cost. Whether the upgrade is worth it depends on how sensitive your coding tasks are to that 6-point difference — probably not for small bug fixes, maybe yes for complex refactors. What's missing is a cost breakdown: is the model itself pricier to run, or did the agent loop count just get cranked up?

Source: AI HOT (Curated Pool) · x.com
Published: Sep 24, 2026, 09:24 UTC+8 · 6 days ago