Social·Community
Blind test: TRAMA 2.0 deliberative system vs the same frontier model used alone
A developer ran a blind test pitting TRAMA 2.0, his multi-perspective deliberative system built on a frontier model, against the same model used alone and other systems on one public-spending question. Claude Opus 5.5 judged the answers against six pre-set criteria. The plain model call won on reasoning; the deliberative system won only on verifiability, and web search mattered more than deliberation.
1 engagement1 sourceSat, Oct 3, 2026, 12:11 PM
AI Overview
- Builder's TRAMA 2.0 uses three internal faculties and logging; blind-judged by Claude Opus 5.5 against six preset criteria on one question.
- Frontier model alone won on reasoning, finding the ESCo-financed option that dissolved the LED-vs-programme dilemma in under 60 seconds.
- TRAMA with web search produced the most verifiable answer (cited sources, discarded a checked figure) but made a calculation error in about 4 minutes.
- TRAMA without search assumed data unavailable instead of looking up typical values; a lighter rival model organized well but refused to choose.
- Author's lessons: search mattered more than deliberation, and faculties on one model missed the arithmetic mistake; a 3–5 question test without tools is next.
Entities
Related
Social·Community
Sonnet 5.5 vs Opus 5.5 vs Sonnet 5 in Claude Code and pi: results and behavioral differences on 10 coding tasks
Multiple independent developers benchmarked Anthropic's new Sonnet 5.5 against Opus 5.5 and Sonnet 5 in Claude Code and pi-agent. Sonnet 5.5 matched Opus 5.5's pass rates while running roughly twice as fast and costing one-third to one-half as much, and one tester with skills loaded could not distinguish the two models. A separate comparison found Opus 5.5 itself takes far fewer steps and tokens than Opus 5.
Multiple independent developers benchmarked Anthropic's new Sonnet 5.5 against Opus 5.5 and Sonnet 5 in Claude Code and pi-agent. Sonnet 5.5 matched Opus 5.5's pass rates while running roughly twice as fast and costing one-third to one-half as much, and one tester with skills loaded could not distinguish the two models. A separate comparison found Opus 5.5 itself takes far fewer steps and tokens than Opus 5.
Updated Tue, Sep 29, 2026, 04:23 PM — A skills-focused rerun on 2026-09-29 found Sonnet 5.5 indistinguishable from Opus 5.5 across 3 skills and 3 runs each.
11148 engagement3 sourcesredditThu, Sep 24, 2026, 08:15 PM
Social·Community
Opus 5.5 vs Sol 6.1 build Lake Bled castle in three.js under a 15-minute constraint
A Reddit user prompted two frontier models to build a simplified, rotatable Lake Bled Castle in three.js under a 15-minute limit, then published a side-by-side comparison. Claude Opus 5.5 finished in 5m46s with a 27 KB file; Sol 6.1 took 11m2s but cost less (~$0.96 vs ~$1.56 estimated API-equivalent) and produced a 2.08 MB uncompressed file.
A Reddit user prompted two frontier models to build a simplified, rotatable Lake Bled Castle in three.js under a 15-minute limit, then published a side-by-side comparison. Claude Opus 5.5 finished in 5m46s with a 27 KB file; Sol 6.1 took 11m2s but cost less (~$0.96 vs ~$1.56 estimated API-equivalent) and produced a 2.08 MB uncompressed file.
34 engagement1 sourceredditFri, Oct 2, 2026, 05:21 PM
Social·Community
Developer runs Claude and Codex debates in a shared markdown doc to pressure-test decisions, seeks convergence structure
A developer describes running Claude and Codex as a debate pair for real decisions: both models append rounds to a shared markdown doc, argue with evidence, and converge while the human relays. The setup catches real errors — one model verified provider pages and debunked another's near-certain-income claim — but suffers from a 2,500-line ever-growing transcript, simultaneous posting, and manual coordination. The author asks for round limits, stop rules, doc templates, and orchestration tools.
A developer describes running Claude and Codex as a debate pair for real decisions: both models append rounds to a shared markdown doc, argue with evidence, and converge while the human relays. The setup catches real errors — one model verified provider pages and debunked another's near-certain-income claim — but suffers from a 2,500-line ever-growing transcript, simultaneous posting, and manual coordination. The author asks for round limits, stop rules, doc templates, and orchestration tools.
12 engagement1 sourceredditWed, Sep 23, 2026, 08:05 PM
Social·Community
Five frontier AIs were told to engineer and 3D-print the strongest bridge they could with 500 g of plastic. Claude Opus 5.5's design held ~130 lb, nearly 5× the runner-up
A Reddit post describes a contest in which five frontier AI models were tasked with engineering and 3D-printing the strongest bridge possible using 500 g of plastic. Claude Opus 5.5's design held roughly 130 lb, nearly five times the runner-up's load. The post offers a hands-on comparison of frontier models' physical engineering output rather than benchmark scores.
A Reddit post describes a contest in which five frontier AI models were tasked with engineering and 3D-printing the strongest bridge possible using 500 g of plastic. Claude Opus 5.5's design held roughly 130 lb, nearly five times the runner-up's load. The post offers a hands-on comparison of frontier models' physical engineering output rather than benchmark scores.
1.6k engagement2 sourcesredditSun, Sep 27, 2026, 09:01 PM