A four-model Claude ensemble scored 78% on Terminal-Bench 2.1, at double the cost of first place

The run placed seventh with a bill of $1,178, roughly twice what the top entry spent. The division of labor: Fable 5 plans and delegates, Haiku 4.5 scouts, Opus 5 edits code, Sonnet 5 verifies the work.
Three security tasks went unsolved because the executor refused to run them: zero out of 14 delegated attempts. In a control run without delegation, Opus 5 solved the same tasks six times out of six. Credit those three and the ensemble lands at 80.5% and third place.
Verification did most of the heavy lifting. Tasks that went through the verifier passed 91% of the time, versus 43% without the check. Opus 5 in the executor role accounted for 58% of total spend, or $689.
First place on the board still belongs to Claude Code with Fable 5 at 83.8%.
Up next
Meta releases Muse Glimmer weights for local agents
Meta Superintelligence Labs open-sourced Muse Glimmer, a 30B agent model, under Apache 2.0. Full precision needs 55GB+ of memory; quantized, it drops under 20GB and runs on one consumer GPU. Meta claims 3.1x faster decoding on an RTX 5090.
Related stories
- Claude Code and two free models split on Next.js picks
- Claude Code matches Codex on SWE-Bench Pro at twice the cost
- Auditing one Claude session: 7.5M tokens, 4,038 tool calls
- Claude Code adopters merged 24% more PRs, study finds
- Claude Code sends 33k tokens before it even reads your prompt; OpenCode uses 7k
- Microsoft’s MAI Code 2 claims 96% of Opus 5.5 at 5% of price
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

