Samson's comparison starts with a practical question: which model gets a creative idea into a working game, 3D scene, explainer, or animation with an acceptable amount of time and spend? The video tours results made by several creators, then shows Samson connecting Opus 5.5 to outside creative tools.
Watch Samson's Comparison
Credit: AI Samson's video, published 25 September 2026. The video includes a paid Higgsfield sponsorship link. The model and demo observations below are attributed to Samson and the linked creators.
Five Comparisons Worth Making
| Test | What Samson shows | What to inspect yourself |
|---|---|---|
| Game worlds | Opus-made Zelda-like and Minecraft/Star Wars scenes, a crowd-combat browser game, and a fishing world. | Can you play for ten minutes, recover from errors, and understand the rules? |
| Rocket launch | A side-by-side visual comparison of Opus, Astra, Sol, and Luna. | Check animation continuity, scene logic, and repeatability across runs. |
| Product launch page | Four models design a page for the fictional Pebble console, with creator-reported build costs. | Test page interactions, accessibility, responsive layout, and total repair time. |
| Night train and strategy scenes | Visual detail differs across model outputs. | Judge against the same prompt, assets, constraints, and budget. |
| 3D-to-video pipeline | Opus builds an underwater scene and directs an outside video model through Higgsfield. | Separate scene construction from video rendering and check continuity between shots. |
A compelling screenshot is evidence of an output. It says less about how many passes it took, whether the mechanics hold up, or whether someone else can extend the work. For a useful comparison, repeat the same task, cap the spend, record failures, and review the finished artifact blind.
The Strongest Ideas Go Beyond Game Screenshots
Samson's examples become more interesting when a generated scene lets someone do something. The interactive camera lens lab by Ryan Sael lets learners change aperture and focus and observe the result. A walk-through house demo turns visual references into a navigable space. Victor Mustar's LEGO Microduck pairs a design with a parts plan and build instructions; its physical build still needs verification.
The Minecraft and Star Wars crossover and Voxel Musou battle show a different direction: creators can prototype a visual language and player interaction together. Bubu's playable build and process notes make that example more inspectable than a short clip alone.
Price Per Token Is Only One Part of Cost
Samson cites a launch-page comparison in which the least expensive model produced a usable first pass for a few cents. That is valuable for exploration. Yet a finished product also includes retries, tool calls, asset creation, debugging, review, and the time required to correct mistakes. His larger fishing-world example illustrates how a deep build can consume substantial credits even when the result looks impressive.
Anthropic's launch material reports a 40% lower cost than Opus 5 on typical workloads and lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens. That is vendor pricing for a model endpoint, not the bill for a finished game. Samson also compares Astra with the smaller Sol and Luna models; effort settings, harnesses, and tool access can change both quality and cost.
A Workflow for Your Own Creative Test
- Pick one artifact. A playable browser scene, a lens explainer, or a short 3D shot is enough.
- Freeze the brief. Give each model the same assets, rules, quality bar, and time limit.
- Generate cheaply first. Use a lower-cost model for rough layout and logic where it clears your quality bar.
- Escalate the hard part. Bring in Opus or Astra for complex mechanics, spatial reasoning, or difficult repairs.
- Use specialist tools for specialist work. In Samson's sponsored example, Opus directs the scene and Higgsfield's connected video tools render the shot.
- Run the artifact. Test on mobile and desktop, vary the inputs, measure frame rate, and inspect the files another builder would inherit.
A Good First Tour of the Gallery
- Rocket launch comparison by Tony / @EnvolDev: four models attempt the same animated task.
- Pebble launch-page comparison by @noclipepe: useful for comparing visual polish against reported spend.
- Camera lens lab by Ryan Sael: a live educational interaction rather than a static render; original X post.
- Voxel Musou by Bubu: playable crowd combat with build notes and an X demo.
- LEGO Microduck by Victor Mustar: a claimed parts list and instruction booklet that invites a real-world build test.
- Minecraft x Star Wars by ChrisGPT: an early visual experiment, with no playable build linked in the source post.
All 77 Linked Examples
The list below is a snapshot of Samson's explorable example gallery, checked 28 September 2026. It contains 77 entries across 13 categories. Some entries share a source or describe a reported capability; an Explore build link appears only when Samson lists a separate demo destination. The source and creator links let you inspect the original context.
Games & visual builds16 entries
-
Minecraft-style voxel game
Fully playable Opus 5.5 build.
-
WWI flight simulator / dogfighting game
Playable browser game.
-
Real-time strategy game
Resource gathering, construction, soldiers, exploration and hostile camps.
-
Dark Souls-style Unreal game
Opus 5.5 + Unreal recreation with multiple worlds/bosses.
-
Mario Maker-style game
One of Alex's Opus 5.5 Unreal experiments.
-
Rocket roguelike
Another playable-game experiment made with Opus.
-
Pokémon-style card game
Opus added features and also helped edit the trailer.
-
Generate 12 different puzzle games
Claude judges the set and develops the winner.
-
Tidekeeper puzzle game
The winning game from that 12-game tournament.
-
CrossFire-style 3D shooter
One-shot browser recreation, with Opus researching the original map before building it.
-
QQ Speed-style 3D racing game
One-prompt build with procedural assets.
-
Pelican cycling game/world
Claude built it, stress-tested it for a simulated 320 seconds and then debugged memory/graphics problems itself.
-
Build a game from a single prompt
Anthropic says an Opus 5.5 build beat other Claude models on graphics and polish.
-
Recreate San Francisco in Unreal Engine
Including cars, pedestrians, pets and traffic.
-
Generate an animated stick-figure heist film
Synchronized around a music-driven story concept.
-
Realistic Minecraft × Star Wars
ChrisGPT posted a 1-minute-49-second video of an Opus 5.5 Minecraft and Star Wars crossover. The creator describes these as early results and does not link a playable build in the post.
Playable Opus games5 entries
-
Playable game · @_MaxBlade
Open the original creator post to inspect this example.
-
Playable game · @MengTo
Open the original creator post to inspect this example.
-
Playable game · @edwinarbus
Open the original creator post to inspect this example.
-
Playable game · @bridgemindai
Open the original creator post to inspect this example.
-
Voxel Musou · Zhao Yun vs 300 soldiers
BubuAi built a playable Dynasty Warriors-inspired browser game with Opus 5.5 ultracode. It has combos, dodges, a Musou attack and roughly 300 voxel soldiers. The creator publishes the source and describes multiple build and critique rounds.
Astra vs Opus games3 entries
-
Astra vs Opus game comparison · @higgsfield_ai #1
Open the original creator post to inspect this example.
-
Astra vs Opus game comparison · @higgsfield_ai #2
Open the original creator post to inspect this example.
-
Astra vs Opus game comparison · @higgsfield_ai #3
Open the original creator post to inspect this example.
3D & simulation3 entries
-
A buildable LEGO Microduck
Victor M says Opus 5.5 designed a life-size Microduck from 1,113 real LEGO pieces, checked 3,204 connections with no collisions, and produced a 141-page, 237-step instruction booklet. The physical build remains unverified.
-
Walk through a house from one photo
Higgsfield describes a Blender house model and web viewer made from a photo and floor plan, with views through the walls and interior.
-
Blender claymation from one prompt
Alex Albert says Opus 5.5 used Blender through claude.ai to make a claymation from a single prompt. The Instagram roundup also mentions this example.
Motion graphics & animation10 entries
-
Motion graphics / animation · @chhddavid
Open the original creator post to inspect this example.
-
Motion graphics / animation · @devteamdrew
Open the original creator post to inspect this example.
-
Motion graphics / animation · @donaldjewkes
Open the original creator post to inspect this example.
-
Motion graphics / animation · @kevin_t_ngo
Open the original creator post to inspect this example.
-
Motion graphics / animation · @Michaelzsguo
Open the original creator post to inspect this example.
-
How ChatGPT picks the next word
A code-drawn prize wheel turns token prediction into a 26-second explainer.
-
The moon tucks the stars in
A gentle bedtime animation made entirely from code and synthesized sound.
-
The half second after you press Enter
A moving network map explains DNS, TLS and page loading in 30 seconds.
-
I’m Upping My P(doom) music video
A lyric-led music video built as a p5.js animation with its source code published.
-
A house builds itself in 3D
The creator shows a Three.js code animation that progresses from the first line of a house to the completed structure.
Fun websites2 entries
-
Fun website · @petergyang
Open the original creator post to inspect this example.
-
Interactive camera lens lab
Ryan Sael’s interactive camera lesson lets viewers change focus, focal length and aperture to see how the optics respond. The linked source describes an 86-minute build.
Model comparisons3 entries
-
Four models build the same rocket launch
Tony gave Claude Opus 5.5, GPT-6 Astra, Sol and Luna the same “Build a rocket launch” prompt and posted the results together. His visual judgment is a creator opinion, not a controlled benchmark.
-
A product launch for a console that does not exist
noclipepe gave Opus 5.5, GPT-6 Sol, Luna and Grok 4.7 the same one-shot brief for a fictional handheld console site. The creator reports approximate costs of $1.80, $1.10, $0.04 and $2.90 respectively; those are this run’s reported costs.
-
Four models build a night train
Tony says he gave Opus 5.5, GPT-6 Astra, Sol and Luna the same “Build a Night train” prompt. His caption also refers back to a rocket launch, so use the video itself to assess the train comparison.
Serious coding10 entries
-
Migrate a 680,000-line codebase
In less than a day.
-
Audit and repair a 200,000-line codebase
In under three hours.
-
Translate HAProxy from C into Rust
Completed in 9.5 hours while passing nearly all regression tests.
-
Optimize an entire web app's loading speed
Anthropic reports success on 39/40 pages.
-
Manage a 40-PR rebase
Stripe reported one Opus session coordinating a dozen other sessions and resolving the conflicts.
-
Solve terminal-based software-engineering tasks autonomously
Its main Terminal-Bench use case.
-
Make repo-wide code changes that would actually be mergeable
Tested through FrontierCode.
-
Operate as an IDE coding agent
In VS Code. GitHub reports fewer steps/tokens on agentic jobs.
-
Use Claude as a command-line coding agent
Via GitHub Copilot CLI.
-
Recover automatically when a multi-step coding task goes wrong
GitHub specifically highlighted improved error recovery.
Bug finding & verification5 entries
-
Formally verify software using Lean
Boris Cherny used it against the Claude Agent SDK.
-
Use TLA+ to model concurrency and state-management problems
Formal method example from the earlier linked research.
-
Find race conditions humans missed
That formal-verification experiment resulted in 16 bug-fixing PRs.
-
Generate mathematical proofs about software behavior
The same experiment generated 1,529 Lean theorems.
-
Have one AI sub-agent independently audit another AI's code
The Pelican demo used exactly this workflow and reportedly found 12 issues.
Agents & computer use8 entries
-
Control a computer GUI to finish tasks
Measured by OSWorld. Opus 5.5 scored 81.8% on Anthropic's partial evaluation.
-
Run multi-hour autonomous tasks
Rather than responding turn-by-turn. This is explicitly one of Opus 5.5's intended workloads.
-
Coordinate multiple sub-agents
On one job, as demonstrated in the Stripe 40-PR rebase case.
-
Automate business workflows
Measured via Zapier's AutomationBench.
-
Deploy software autonomously to Cloudflare Pages
Including CDN/cache work and domain setup — done in the three-game demo repository.
-
Debug a production-only failure
The Pelican game reportedly developed an occasional black screen after deployment; Claude reproduced the NaN failure, located it and redeployed a fix.
-
Research the web before coding
Then use what it learns to reproduce mechanics accurately. The CrossFire/QQ Speed builds used this workflow.
-
Automatically stress-test its own creation
For memory leaks and long-session failures.
Research & professional work6 entries
-
Carry out long-form research and produce reports
Anthropic tested research-report reliability; Snowflake reports 16/18 met a bar rejecting fabricated quotes/numbers.
-
Scientific research agents
Tested through Terminal-Bench-Science.
-
Analyze charts visually and extract information with tools
Anthropic reports 89.0% on Chartography.
-
Solve multidisciplinary research problems using external tools
Evaluated on Humanity's Last Exam.
-
Analyze massive document/code collections in one context
Thanks to the 1M-token context window.
-
Produce extremely large structured outputs
Up to 128K output tokens, with 300K in Batch beta.
Data, finance & enterprise4 entries
-
Analyze company data directly inside Snowflake
Keeps data within Snowflake's governance perimeter.
-
Build AI agents over enterprise databases
Using Snowflake Cortex Agents.
-
Financial modelling and merger analysis
One reported internal merger-analysis task took Opus 5.5 63 minutes versus 93 for Opus 5, at roughly half the cost.
-
Professional knowledge-work tasks
Across areas such as finance, law and business analysis — evaluated with GDPval-AA.
Writing & huge content2 entries
-
Rewrite complex technical explanations in a strict house style
Anthropic specifically highlights improved instruction-following and clearer, less jargon-heavy writing.
-
Work through enormous projects without continually re-feeding context
Arguably the broader use case underlying Opus 5.5's combination of 1M context + agentic coding + persistent adaptive reasoning.
Video Chapters
| Time | Topic | Time | Topic |
|---|---|---|---|
| 00:00 | Opus vs Astra | 01:57 | Zelda, Star Wars, and games |
| 02:42 | Complex game economics | 04:39 | Rocket launch comparison |
| 05:53 | Hidden costs | 07:16 | Visual and physics results |
| 08:52 | Interactive learning | 11:17 | Connected creative tools |
| 15:49 | Safety discussion |
Sources and Creator Links
- AI Samson: Here's What Opus 5.5 Can Actually Do, primary video and transcript.
- Samson's 77-example collection, including creator profiles, original X posts, and playable builds.
- Anthropic: Claude Opus 5.5 release and pricing.
- OpenAI: GPT-6 Astra announcement and API changelog for GPT-6 Sol and Luna.
- Artificial Analysis for current independent model comparisons; its rankings change over time.
- Higgsfield, sponsor link supplied with the video.
- Samson's newsletter and business contact.