Pat Simmons put Gemini 4 Argon beside Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra on four demanding builds. His test is more useful than a single leaderboard position because he links the exact prompts, all 16 live results, and full token, time, and cost tables. It is also narrower than the title might suggest: these are visual, code-heavy projects, not trials of legal analysis or everyday office work.
In Pat's blind ranking, Opus 5.5 finished first and Argon fourth in all four builds. Argon used the least estimated API spend: $39.54 total at Google's introductory rates versus $189.04 for Opus, $203.08 for Fable, and $409.96 for Astra. Those are Pat's API-equivalent estimates, not what he paid on subscriptions or a general model leaderboard. For this kind of first-pass visual work, the cheaper run did not produce his favorite result. Source: Pat's cost summary.
Watch the Four-Build Test
Credit: The demonstrations, rankings, prompts, skills, and cost tables are Pat Simmons's work. Read his original AI for Mortals article for full prompts and openable builds, or follow his newsletter. Google's product claims and current API terms are sourced separately below.
What the Test Can and Cannot Tell Us
Pat says each model received the same prompt per task, with no subsequent steering from him, and he ranked anonymous columns before revealing the model names. The coffee builds used his fictional Clanker Coffee kit. For the video, every model could use his Launch Video Clone skill; for the website, all four used his clone-workflow skill. The full instructions are on Pat's article, not reproduced here.
Blind judging helps with brand bias, and open builds let readers check the artifacts. But this is still one run per model per creative task, with subjective quality judgments and differing agent trajectories. A high token count can reflect more tool rounds, caching, or a longer attempt; it is not a pure measure of model intelligence. Pat's own conclusion is that these unusually hard visual tasks may not predict knowledge work. The website-clone exercise is a comparison artifact, not permission to publish an imitation of someone else's brand.
Four Builds, With the Results Open
1. Interactive 3D cold-brew bottle
What Pat saw
The brief asked for a photoreal bottle with orbit and zoom, a cap that opens, liquid poured over ice, and convincing materials. Pat found Argon's shape and pour less realistic. He ranked Opus's lighting, condensation, and product finish first; Argon came fourth.
Practical read
Argon finished in 36 minutes at an estimated $6.85; Opus took 114 minutes at $29.95. The cheapest preview may be useful for direction-finding, but a client product shot needs a human check of proportions, label legibility, materials, and interaction before it is usable. Full prompt and table.
2. Clanker Coffee launch video
What Pat saw
All four had the same fictional brand and video skill. Pat thought Argon's result had too much text and felt unlike a short commercial, despite stronger 3D work than its bottle test. Opus's pacing, product imagery, and motion made it his first choice. Argon placed fourth.
Practical read
The estimated cost was $10.14 for Argon in 80 minutes and $44.98 for Opus in 318 minutes. A finished launch video also needs claims review, sound, readability at phone size, and licensed assets; the model bill is only part of that production cost. Full prompt and table.
3. Browser car-soccer game
What Pat saw
The specification pushed for photoreal stadium art, vehicle and ball physics, a computer opponent, camera modes, controls, and sound. Pat preferred the feel and visual finish of Opus's game. Argon's car and scene looked weaker to him, and it ranked fourth.
Practical read
Argon took 54 minutes at $10.59 estimated API cost; Opus took 155 minutes at $46.24. A playable demo is not a tested game. Check controls, frame rate, opponent behavior, collision edge cases, and original art before any public release. Full prompt and table.
4. Motion-heavy website reconstruction
What Pat saw
The reference was Oryzo's 3D website. Pat asked the agents to recreate its scrolling scenes and interactions using their own assets. He found Opus closest overall, although its hand scene still fell short. Argon's textures, transitions, and hand were notably weak to him; it ranked fourth.
Practical read
Argon took 77 minutes at $11.95. Opus took 129 minutes at $67.87. Astra spent 499 minutes and an estimated $357.84, an example of how an agent can work much longer without winning. Use this as a diagnostic exercise with permission; do not ship copied design or brand identity. Full prompt and table.
Cost per Job Matters More Than Price per Token
Here are the API-equivalent estimates from Pat's written tables, in US dollars. He says the actual runs used subscriptions, and that his $99/month Google plan hit a daily limit after about one Argon build. The estimates exclude human review, brand preparation, external media services, hosting, and the value of revisions. The last task is especially sensitive to one long Astra run. Source: creator's cost tables.
| Build | Argon | Opus 5.5 | Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| 3D bottle | $6.85 | $29.95 | $17.96 | $8.99 |
| Launch video | $10.14 | $44.98 | $54.26 | $28.44 |
| Car-soccer game | $10.59 | $46.24 | $32.30 | $14.70 |
| Website reconstruction | $11.95 | $67.87 | $98.56 | $357.84 |
| Four-build total | $39.54 | $189.04 | $203.08 | $409.96 |
Rounding note: The total row reproduces Pat's aggregate table. Adding its rounded task rows gives $39.53 for Argon and $409.97 for Astra, one cent away from those aggregates. The time totals are 247 minutes for Argon, 715 for Opus, 548 for Fable, and 670 for Astra. Pat's ranking favors Opus's results, but those took substantially more build time. Creator's totals and methodology.
Google's Benchmarks Are a Different Question
Google's launch post reports strong performance on its chosen coding, legal, finance, automation, and long-video evaluations, including 77.9% on DeepSWE v1.1 and 51.3% on AutomationBench. These are Google-reported benchmark results. Pat did not test those domains. His four builds show a gap between published breadth and his preferred visual artifacts, not that the benchmark numbers are invalid or that Argon is fourth at every job.
As of 15 October 2026, Google's developer documentation lists the gemini-4-argon API and AI Studio access, with introductory pricing through 30 November: $2 per million input tokens and $10 per million output tokens, including thinking tokens; implicit cache hits receive a 95% input discount. On 1 December, those base rates are scheduled to double to $4 and $20. The API doc also says Argon is not on its Free tier or the Google AI Pro plan. That is an API entitlement statement, not a blanket description of every Gemini app surface. Check your actual account and limits before committing to a client workflow.
Verdict for builders: If the job is an interactive product demo, launch film, game, or intricate 3D site, open Pat's live builds and judge the accepted output, revision effort, elapsed time, and total tool cost together. Argon was the economical first pass in this sample; Opus was Pat's quality winner. A new task or repeated trials could change that order.
Jump to the Video
| Time | Segment | Time | Segment |
|---|---|---|---|
| 00:00 | Introduction | 00:46 | Google's launch demos |
| 03:15 | Pricing and benchmarks | 06:59 | 3D product bottle |
| 14:50 | Product launch video | 26:48 | Car-soccer game |
| 33:01 | Website reconstruction | 43:14 | Verdict |
Business Idea: Test a Launch-Asset Service
Pat's test suggests a practical offer: make one original, high-quality launch asset for a buyer who already has a product and approved brand materials. The buyer pays for a useful, reviewed deliverable, not for model access or a cheap first draft.
Test a product-launch asset offer
Pick one buyer, one asset, and a measurable review standard.
Sources and Limits
Creator: Pat's full video and original written account, prompts, 16 builds, and detailed tables. Official: Google's launch post and developer pricing and access guide. Rankings are Pat's visual judgments from four single runs, not independent repeated evaluations. Prices are time-sensitive API-equivalent estimates; plan entitlements and actual invoices differ. All third-party creative work still needs rights and human quality review.