Can an Army of Claude Haiku 5.5 Subagents Beat Opus 5.5?

Summary
In this video, creator Dubibubi evaluates whether Claude Opus 5.5 orchestrating a swarm of cheaper Claude Haiku 5.5 subagents can match or beat a solo Claude Opus 5.5 instance. Through three benchmarks—coding an Angry Birds browser game clone, generating a canvas-based animated samurai short film, and reviewing an application codebase for bugs—he compares quality, execution speed, and token cost. Solo Opus 5.5 decisively beats the multi-agent Haiku team in output quality and overall efficiency, winning the comparison 10 to 2.

What is shown

• Setup & Delegation Skill [00:47–02:20]: Dubibubi configures Claude Code with high effort for both sides, equipping the team setup with a custom /delegate skill that directs Opus 5.5 to plan while spawning Haiku 5.5 subagents to execute tasks.
• Test 1 (Angry Birds clone) [02:41–04:20]: Demonstrating gameplay of “Squawkpult” built by the Opus + Haiku team versus “Feather Fling” built by Solo Opus, showing physics, sling aiming, destruction, and level selection.
• Test 2 (Samurai animation film) [06:02–08:56]: Showing parallel subagent execution tracking [06:02], followed by playback of Solo Opus’s film “The Red Thread” [06:36] and the Haiku team’s film “Ensō” [08:00], both rendered entirely with JavaScript canvas and featuring ElevenLabs voiceover.
• Test 3 (Code Review & Bug Hunting) [10:44–13:38]: Running both setups on Dubibubi’s application codebase (ACE), scoring detected bugs using GPT-6 Astra (+1 for confirmed bugs, -1 for hallucinations).
• Usage Tracking Dashboard & Scoring [02:22, 04:28, 09:45, 11:46, 13:54]: Detailed API token usage, run duration, and cost analytics across all three tests.

Claims & numbers

• The presenter notes Claude Haiku 5.5 costs $0.10 input / $0.50 output per million tokens, making it 40x cheaper than Opus 5.5 ($4/$20 per million tokens) [00:01].
• In Test 1 (game clone), Solo Opus completed in 43m 20s costing $12.78, while the Opus + Haiku team took 1h 29m and cost $11.62 (saving 9.1%) [04:28].
• In Test 2 (animation), Solo Opus completed in 1h 39m costing $25.04, whereas the Opus + Haiku team took 2h 38m and cost $44.79 (78.9% more expensive) due to heavy orchestration token overhead [09:47].
• In Test 3 (code review), the Opus + Haiku team found 103 confirmed bugs and 14 rejected/hallucinated bugs (net score +89) in 49m 19s costing $55.16 (61% savings) [11:51, 12:13]. In contrast, Solo Opus took 38m 49s (or 18m 33s active working time), cost $141.54, and found 122 confirmed bugs with 0 rejected/hallucinated bugs (score +122) [13:08, 11:51].
• The presenter references historical benchmarks on the same code review eval: Claude Fable 5.1 scored ~21, and Sonnet 5.5 scored 85 [12:38, 12:47].
• Across the three tests, Solo Opus won 10 points out of 12 (winning quality in all three tests, speed in all three, and cost in test 2), while the Haiku team won 2 points (cheapest in tests 1 and 3) [13:54].

Notable quotes

• “For every dollar Opus spends, Haiku spends two and a half cents.” [00:08]
• “This test has just made me realize how freaking good Opus 5.5 is... This felt like a movie.” [09:00]
• “If a job splits cleanly into pieces, like hunting bugs, give it to a Haiku team. If it needs taste or the full picture, let Opus do it alone.” [14:47]

Assessment
This is an independent hands-on review and benchmarking video comparing multi-agent delegation against a solo frontier model. The demonstrations of code execution, interactive games, animated scripts, and dashboard metrics are shown live and authentic, though portions of long multi-hour generation runs are naturally cut for time.

Described by gemini-3.8-flash on 2026-10-10 from the video’s audio and frames.