Opus + Sol Orchestrated Team: This BEATS every AI MODEL right now! (Tested)

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Opus + Sol Orchestrated Team: This BEATS every AI MODEL right now! (Tested) · Channel: AICodeKing · 13:55 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, AI developer tools, AI models

Summary

The video compares five AI lead-and-worker teams across six coding tasks, evaluating both their applications and the teams’ orchestration decisions. Opus leading Soul achieves the highest comparable overall score, while other pairings excel on specific tasks; incomplete plans, broken features, and provider quota interruptions affect several results. The author also reviews recorded token usage and estimated costs, emphasizing that missing usage data prevents a reliable cost-per-score comparison. The conclusion favors Opus with Soul as a starting setup, while stressing the need to check task ownership, reproduce failures, and test complete user journeys.

Key points

Five lead-worker pairings were tested on six tasks, with up to three simultaneous workers, 20 lead rounds, and 50 assignments per team.
The comparison used one high-effort run per task on a shared machine, without equal compute budgets, matched solo tests, or reversed roles.
Opus with Sonnet achieved the highest emulator score, while Astra-led emulator plans omitted DS and 3DS execution paths.
Room Designer results were strong overall; Astra with Sonnet earned the highest room score, and the tester-repair-review loop in Opus with Soul was highlighted.
The keyboard task showed that a polished editor is insufficient if it cannot communicate with the tested device; both Astra-led teams omitted that connection.
Astra with Sonnet earned the highest Pokémon score by completing the tested route, while two interrupted or incomplete runs scored zero.
Opus with Soul led the Fitbit task, but none of the teams demonstrated a personal Fitbit account connection or a complete version of the requested product.
Opus with Soul scored 44.2 out of 60 overall, narrowly ahead of Opus with Sonnet at 43.7; quota interruptions limit comparisons of capability.
Recorded orchestration usage totaled about $405.49, but only eight of 30 tasks had complete captured usage, so final costs and value comparisons remain unknown.
The author recommends checking that every required feature has an owner before work begins, then reproducing failures and testing the normal user journey and final package.

Chapters

Start	Chapter	Summary
00:02	Test setup and emulator	The author introduces five AI teams and six coding tasks, then compares their emulator implementations and notes that some plans omit DS and 3DS execution.
02:16	Docker alternative and Room Designer	The Docker tests do not establish a complete replacement, while the Room Designer results show stronger applications and examples of teams catching and repairing defects.
04:23	Keyboard Configurator and Pokémon game	The keyboard comparison highlights missing device communication in Astra-led plans, and the Pokémon tests contrast a completed route with incomplete or quota-interrupted runs.
07:34	Fitbit desktop app	The teams deliver useful health-app features but retain gaps in provider connections, imports, data handling, and verification of the native desktop experience.
09:42	Overall scores, usage, and costs	Opus with Soul leads the comparable total, but incomplete usage records and quota interruptions prevent firm capability or value rankings.
12:51	Conclusions and recommendations	The author favors Opus leading Soul as a starting setup and recommends assigning every required feature, reproducing failures, and testing the finished package.

Quotes

There are no matched solo tests or reversed roles, so this comparison doesn’t isolate how much delegation itself helps. — 01:10

That’s the kind of review loop I want from a team. — 03:20

So, the teams improve local editors while the required connection remains absent. — 05:27

Every team’s final cost remains unknown, so these subtotals can’t identify the cheapest setup or a score per dollar winner. — 12:21

For my next project, I’d check that every required feature has an owner before the workers start. — 13:24

Concepts and entities

Opus 5.5, Sonnet 5.5, GPT-6.1, GPT-6, Astra, Sol, Muse Wonderful 1, Bambood, Claude Code, Codex, Muse CLI, orquestração, AICodeKing

Full transcript (original language)

00:02 Hi. Welcome to another video. So, I tested five AI teams to see which lead and worker setup I’d actually use for a coding project. One model plans the work and checks the results, while another handles the assignments. That’s orchestration, and the lead’s decisions can shape the whole application. I followed their conversations and tested the applications. So, let’s see where the teamwork helped and where it failed. The models are Opus 5.5, GPT 6.1, Soul, GPT 6, Astra, Sonnet 5.5, and Muse Wonderful 1. The five pairs are Opus

00:37 with Soul, Astra with Sonnet, Opus with Sonnet, Astra with Soul, and Opus with Museful. The first model is always the lead. Everything ran through Bambood using Claude Code, Codex, and the native Muse CLI. Each team had up to three simultaneous workers, 20 lead rounds, and 50 assignments without an added time limit. The six tasks were an emulator, Docker alternative, Room Designer, Keyboard Configurator, Pokémon Opening, and Fitbit Desktop App. Each task was worth 10 points. Each pair gets one high-effort run per task on a shared machine without equal computation

01:10 budgets. There are no matched solo tests or reversed roles, so this comparison doesn’t isolate how much delegation itself helps. The emulator asks for Game Boy, DS, and 3DS support in Rust, including saving and normal playback features. Both Astra-led teams planned Game Boy execution, but only file inspection for DS and 3DS, leaving those execution paths missing. Astra with Sonnet gets 1.9 out of 10, and Astra with Soul gets 2.2. But, improving Game Boy support doesn’t make a DS file inspector run DS game. Opus with Soul earns 6.2 with connected execution paths for all

01:43 three systems. DS and 3DS examples demonstrate rendering with separate 3DS input checks. Opus with Sonnet earns 7.1, the highest emulator score. It passes all eight shared Game Boy diagnostics and the Game Boy Color City Builder supports useful controls and saving. The DS and 3DS examples also progress further, although these small homebrew tests don’t prove that most commercial games work. Opus with Muse full earns 5.8 with some output across all three systems, but blank or incomplete results in broader

02:16 examples. Provider quota errors interrupt workers and the lead completes more of the work. So, Opus with Sonnet produces the strongest emulator here. While the Astra teams show how an incomplete plan limits everything that followed. For the Docker alternative, importing an image is only part of the job. You also need process isolation and resource controls and these macOS hosted reviews didn’t establish working isolation through the delivered Linux paths. That limits the scores to five even where the image tools work. Opus with

02:47 Soul, Opus with Sonnet, and Astra with Sonnet all get 5.0. Astra with Soul gets 4.7 with working registry and Docker save workflows, but no import for the standard OCI archive format. None of the teams demonstrates the requested efficiency advantage over Docker. Opus with Muse full gets 3.0 after an extraction test, changes permissions on two owned directories outside the intended extraction route. Their contents stay unchanged and both remain inside the application storage parent. This doesn’t demonstrate a guest escaping Linux or overriding arbitrary host files.

03:20 That run also suffers quota interruptions and none of these results establishes a complete Docker replacement. The room designer produces much more usable applications. In the Opus and Soul project, a tester finds that selecting a chair makes the viewport go blank. The lead suspects an old builder screenshot problem, but the tester reproduces the failure. A worker repairs the selection highlight and the lead checks the rebuilt app. That’s the kind of review loop I want from a team. Opus with Soul earns 8.9 while Opus with Sonnet and Astra with

03:52 Soul both earn 9.0. The Opus and Sonit version still lets oversized furniture protrude from a smaller room. Opus with Muse full earns 8.9 with custom shapes, colored lighting, and working project restoration. Color changes leave three extra textures retained, although normal editions followed by undo remain stable. This run finishes without an observed quota failure, even though its usage records still have gaps. Astra with Sonit earns 9.3, the highest room score. The lead gives the room model, editor,

04:23 and renderer separate owners, then sends specific corrections as those parts connect. Editing and project restoration work and repeated object additions and removals return to the original resource counts. A very thin object can overhang the wall slightly, but this is the room designer I’d favor from these results. The keyboard task asks for a yellow themed alternative to VA that configures QMK keyboards. A nice editor isn’t enough if the keyboard never receives the changes. I checked configuration messages and resulting state through an independently controlled virtual keyboard. That’s limited compatibility

04:54 evidence, not proof of support for every physical keyboard or firmware version. Opus with Soul earns 7.8 and Opus with Sonit earns 7.4. Both send configuration messages that change the tested device state with editing compatibility and backup limitations reducing their scores. Both Astra lead versions earn 2.8 and their assignments explain why. Astra with Soul assigns selection and metadata, while Astra with Sonit explicitly excludes opening the device and sending configuration messages. The benchmark prohibited changing real connected hardware, but it didn’t prohibit

05:27 implementing that communication path. So, the teams improve local editors while the required connection remains absent. Opus with Muse full earns 7.7 with working keycodes, macros, lighting, and restoration in a fresh browser profile. It still accepts an incomplete acknowledgement and invalid macro imports can partly change the configuration. Quota failures also interrupt the workers before the lead completes further repairs. Opus with Sol leads this task, and the bigger lesson is to include the device connection before spending more work on

06:00 appearance. For Pokémon, I wanted an ordinary playable opening through the first gym with progress that survives saving and loading. Opus with Sol earns 8.6, and Opus with Sonnet earns 8.7. Both complete that route when browser storage works, but change names, simpler artwork, and simplified mechanics reduce fidelity. The credited tests use ordinary play, not the agents’ heavily leveled parties or automatic healing. Opus with Museful scores zero because it delivers no playable game.

06:30 There is a contract and a graphics helper, but no entry point, world battles, or saving. Subscription quota errors affect all six workers used during the run, eventually blocking even a request to read one line and reply. That zero describes the delivered project, not what Muse could do with available capacity. Astro with Sonnet earns 9.1, the highest Pokémon score. The lead catches a walk-through that tells players to heal Imputer without providing a healing center. It assigns the missing door, nurse, and return route. The independent route completes the parcel quest, catches a wild

07:03 Pokémon, and reaches Brock through the forest. Normal losses have the money and give no badge, while winning from a legitimate save awards the badge and expected money. Progress survives another reload when storage works. It’s still a reduced game without evolution or move use limits. A misleading save success message when storage is blocked also costs points because reloading loses that temporary checkpoint. Astro with Sol also scores zero with a completely empty folder after an HTTP

07:34 429 error exhausts retries. The logs don’t establish whether account quota or another rate limit caused it. Astro with Sonnet leads the completed game comparison, while those interrupted runs need working provider access before we can judge their full potential. The Fitbit task asks for a desktop analytics app with a Rust back end. Opus with Soul earns 7.7, while Opus with Sonnet earns 6.5. Soul’s provider clients work against controlled test inputs representing expected provider responses, while Sonnet’s external connection receives credit from source inspection only. Neither result

08:06 establishes a personal Fitbit account connection or clinical accuracy. In the Opus and Sonnet app, a workout’s calculated load can appear in its detail view, but disappear from the daily summary. In the Opus and Soul app, demo data replaces imported daily measurements, and clearing the demo removes those days. Importing the original synthetic CSV again restores the exact values, so the demonstrated loss is recoverable. Opus with Muse full earns 6.3 with substantive Rust calculations, but broken connect and import buttons that

08:39 fail before reaching the back end. Direct back end calls don’t fix those broken buttons. This run also suffers quota interruptions. Both Astro led health apps earn 7.6. Astro with Sonnet catches demo refresh deleting user records. A worker also finds an old interface inside the package tab, so the lead rebuilds it and checks the embedded files. A repaired source file doesn’t help if you launch an older packet. Its final app supports workout and journal entries with data surviving export, import, and restart, but it accepts only its own JSON format without a Fitbit connection

09:10 or actual Fitbit export path. Its load calculation treats missing days as zero, so incomplete history can produce misleading results. Astro with Soul repairs retries that could create duplicate workouts after a lost response. An independent test loses that response, then retries after reopening the dialogue, and exactly one workout remains. The pending draft still doesn’t survive an application reload. That app also relies on its own JSON format rather than a working Fitbit connection. It shows zero steps on a workout only date and renders a thousand

09:42 workouts as a thousand table rows. The Astra health checks use the rust services in a browser, so the complete native desktop experience remains unverified. Opus with Soul leads this task, but every result still has gaps against the requested product. Now, here’s the overall result. Opus with Soul earns 44.2 out of 60 or 73.67%. Opus with Sonnet follows at 43.7 or 72.83% with five quota affected tasks. Astra

10:13 with Soul earns 26.3 or 43.83% including its interrupted Pokémon run. I keep those delivered scores visible, but the interruptions prevent a fair capability ranking across all five teams. These percentages measure earned points, not future success rate. For tokens and cost, the totals include leads, workers, corrections, and recorded additional model calls. Setup and evaluation are excluded. Resumed usage counts only once and missing records remain unknown. The total is about 936.11

10:44 million recorded tokens including 926.74 million input and 9.37 million output. Most input is repeated cached context and reasoning already belongs to output, so it doesn’t count twice. Opus with Sonnet records about 444.81 million tokens compared with 143.73 million for Opus with Soul. Astra with Sonnet records 107.06 million, Opus with Muse full records 155.06 million, and Astra with Soul records 85.45 million. These counts include repeated context

11:17 and don’t directly measure useful code. Using the saved API rates, Soul and Sonnet each cost $2 for fresh input, 0.10 for cached reads, and $10 for output per million tokens. Opus costs $4.20 and $20 for those categories, while Astra cost $10, $1 and $0.50. So, Opus is twice Soul’s unit price, while Astra is five times its fresh input and output price, and 10 times its cached read price. Cash write and long context rules also affect the calculations.

11:49 For Muse full, I used the requested standard wonder assumption, $4 for fresh input, $0.20 for cached reads, and $20 for output. That isn’t a confirmed full tariff or subscription invoice. And there is no confirmed full cash rate right here. For the subtotals, native Claude cost records replace the Claude token estimates where available with token estimates for the other providers. Those alternatives don’t get added together, and the records don’t establish every hosted fee. Opus with Soul records about $56.97,

12:21 while Opus with Sonnet records $124.22. Astra with Sonnet records $68.75. Opus with Muse full records $80.22, and Astra with Soul records $75.22. That’s about $405.49 of recorded orchestration usage, but only eight of the 30 tasks have complete captured usage. Every team’s final cost remains unknown, so these subtotals can’t identify the cheapest setup or a score per dollar winner. A lower recorded amount can also mean

12:51 some usage is missing. For my own starting setup, I’d choose Opus leading Soul. It has the highest comparable total and stays useful across all six tasks, with Opus leading Sonnet a close alternative. I’d favor the Sonnet team for emulation, while Soul’s team scores higher on the keyboard and health applications. Astra with Sonnet is worth considering for the visual editor and small game work demonstrated here. But, I’d inspect its initial assignments carefully because its emulator and keyboard plans omit required work. Opus with Muse full and Astra with Soul

13:24 need complete runs before I’d make a fair overall comparison. For my next project, I’d check that every required feature has an owner before the workers start. Then I’d reproduce disputed failures, test the normal user journey, and check the final package. More runs and complete costs are still needed before making broader reliability or value claims. Overall, it’s pretty cool. Anyway, let me know your thoughts in the comments. If you like this video, consider donating through [music] the Super Thanks option or becoming a member by clicking the join button. Also, give this [music] video a thumbs up and

13:55 subscribe to my channel. I’ll see you in the next one. Until then, bye. >> [music]