Gemini 4 Argon vs Claude Opus 5.5: Winning the Benchmarks, Losing the Builds
Google released Gemini 4 Argon on 13 October 2026 and called it a new generation rather than an update. The benchmark table backs the language: Argon leads nearly every row against the strongest models on the market, and the API price shown at launch — $2 per million input tokens and $10 per million output — undercuts Anthropic's flagship by half.
So the obvious question is whether the thing you actually ship comes out better. Six builds, one prompt each, same prompt to both models, run side by side:
| # | Build |
|---|---|
| 1 | A luxury perfume landing page |
| 2 | A 15-second intro animation |
| 3 | A Minecraft-style game |
| 4 | A sales dashboard from a CSV |
| 5 | Find and fix a checkout pricing bug |
| 6 | A physics sandbox |
Argon lost all six. That is the finding, and the rest of this post is the detail behind it — including the two places where Argon was genuinely better, and the dated fine print under the price that made it look cheap.
About this write-up
Written from the video on the channel. The six runs, the costs, the timings and the picks are what happened on screen.
What is labelled rather than asserted: the benchmark figures quoted in the first section are Google's own, taken from the announcement page in the video, not independent measurements. The per-test costs and durations come from each tool's own session accounting, so two different accountants are being compared. Every build was judged by eye, once, by one person — this is a hands-on comparison, not a controlled benchmark. The product names were badly garbled in the auto-captions and were checked against both vendors' documentation before being printed here.
Prefer to watch it on YouTube?
Watch on YouTube — or subscribe to the channel for the next one.
What Google says Argon is
Worth having in front of you, because it sets up the gap:
| Gemini 4 Argon | |
|---|---|
| Released | 13 October 2026 |
| Built | A new generation, not a revision of the previous model |
| Input | Text, images, audio, files |
| Context | 1M tokens, up from 64K on the previous generation |
| Architecture | Scaled mixture of experts — only the experts relevant to your prompt are activated |
| Effort levels | low, medium, high |
| Price at launch | $2 / $10 per million tokens, in and out |
| Where | Google AI Studio, Vertex AI |
The effort levels matter for a fair comparison, and both sides were set to medium throughout.
On Google's benchmark table Argon leads a composite index at 68.9%, takes 91.9% on the coding measure, holds 99.7% recall at 128–256K context and 84.2% out at a million, and sits around 68% on cybersecurity — ahead of the three rival flagships on nearly every row.
One number in that section does not fit the others, and it comes from an independent aggregate rather than from Google: on that index Argon lands seventh out of 352 models, behind Opus 5.5. A vendor table and an independent index are answering different questions, and here they already disagree before a single build is run.
How the comparison was run
Each model was driven through its own vendor's tool, which is how most people will actually meet it:
| Model | Tool | |
|---|---|---|
| Gemini 4 Argon (medium) | Antigravity 2.24.0 | |
| Anthropic | Claude Opus 5.5 (medium) | Claude Code 2.4 |
The rules were deliberately plain: identical prompts, pasted into both; no skills, no plugins, no custom instructions. Each test ran in its own session so that cost and wall time could be read off per build. Six tests, run in parallel across the two tools, about two hours of model time in total.
Keeping skills and plugins out is the right call for a model comparison, and worth saying out loud: it is what makes the result a statement about the models rather than about two people's setups.
Test 1 — A luxury perfume landing page
The prompt specified brand, sections, visual direction, tone and output.
Both produced a reasonable page with a similar structure — hero, body, product information, footer. The split was in the detail:
- Argon — layout fine; the generated imagery looked artificial, and one product card was not clickable.
- Opus 5.5 — an animation on entry, a perfume-smoke effect, and imagery that read as more convincing.
| Cost | Wall time | Input | Output | |
|---|---|---|---|---|
| Argon | ~$0.20 | 8 min | 112K | 68K |
| Opus 5.5 | ~$2 | 6 min | 38K | 44.6K |
Opus cost about ten times as much and still finished two minutes sooner, on a third of the input tokens. Pick: Opus 5.5.
Test 2 — A 15-second intro animation
Title card, technical constraints, output format and style were all specified.
Opus produced the better piece — a glow effect and a composition the reviewer preferred. Argon's was serviceable.
| Cost | Wall time | |
|---|---|---|
| Argon | ~$0.31–0.45 | 4 min |
| Opus 5.5 | $1.55 | 6 min |
Here Argon was both cheaper and faster. It still lost on the result. Pick: Opus 5.5.
Test 3 — A Minecraft-style game
The prompt covered world, player, UI and visuals. This is the test where the gap stopped being a matter of taste:
- Argon — the UI looked fine and the block-selection keys worked, but the spacebar did nothing. No jump. No audio either.
- Opus 5.5 — jump worked, audio was there, and the interface looked better.
| Cost | Wall time | |
|---|---|---|
| Argon | $0.55 | 6 min |
| Opus 5.5 | $2 | 8 min |
A platformer where you cannot jump is not a cheaper build; it is an unfinished one. Pick: Opus 5.5.
Test 4 — A sales dashboard from a CSV
Real data in, with requirements for KPIs, revenue breakdowns, filters, design and technical constraints.
- Argon — the metrics were there: revenue by month, product, city and channel, plus a data note and a working search. Basic, but correct.
- Opus 5.5 — dark theme, a wider range of chart types, the data table beside the chart rather than below it, colour switching, and an order list. Spot-checking the figures against the source showed them matching.
| Cost | Wall time | |
|---|---|---|
| Argon | $1.78 | 16 min |
| Opus 5.5 | $1.84 | 7 min |
This is the most interesting row in the post. The costs are within four cents of each other — and Opus did it in under half the time. On the one test with real data to chew through, Argon's price advantage evaporated entirely. Pick: Opus 5.5.
Test 5 — Find and fix a checkout pricing bug
A deliberate logic error in checkout pricing, with sample code. The only questions were whether each model found it and how fast.
Both found and fixed it, and both produced the correct total of 797.
| Cost | Wall time | |
|---|---|---|
| Argon | $0.27 | 2 min |
| Opus 5.5 | $0.69 | 1 min |
So: Argon at a third of the cost, Opus at twice the speed, same correct answer. This is the one test where the cheaper model has a real claim, and the pick went to Opus on speed — on a bug this simple, that margin is small enough to decide either way. Pick: Opus 5.5.
Test 6 — A physics sandbox
Sand, water, oil, stone, wood, fire, smoke, steam and plants in one page, with the question being whether each element behaves the way it does in the world.
- Argon — sand fell through water correctly and oil behaved, but fire burned water. Which settles it.
- Opus 5.5 — fire burned wood and plants, water put the fire out, oil carried it, and fire hitting water went out.
| Cost | Wall time | |
|---|---|---|
| Argon | $0.38 | 5 min |
| Opus 5.5 | $1.19 | longer than Argon |
A sandbox whose rules are wrong is a demo of the wrong rules. Pick: Opus 5.5.
The scoreboard
Scored out of ten on three axes, these are the reviewer's own numbers:
| Argon 4 | Opus 5.5 | |
|---|---|---|
| Quality | 6.5 | 9.5 |
| Speed | 5.5 | 8.0 |
| Cost / tokens | 7.5 | 6.5 |
| Builds won | 0 | 6 |
Read the cost row carefully. It is the only one Argon wins, and it is the row that is about to change.
The discount has an expiry date
The $2 / $10 pricing is introductory. Published rates put it in effect to 30 November 2026, after which Argon moves to $4 / $20 per million tokens — which is exactly Claude Opus 5.5's standard list price.
That reframes every cost column above. The gap being measured here is not a structural advantage of one model over another; it is a launch promotion with a date on it. From December, the same six builds cost roughly the same on either side — and the quality column is the only one left standing.
If a cheaper model is the plan, that plan needs re-checking on 1 December.
What I would take from this
Vendor benchmarks and shipped work measure different things. Argon leads almost every row of Google's table and lost six builds out of six. The independent index in the same section, which put it behind Opus 5.5, turned out to predict the hands-on result better than the vendor's own numbers did.
The failures were functional, not aesthetic. A spacebar that does nothing and fire that burns water are not matters of taste. On a quick look both builds are finished; it takes using them to find out that one is not.
Cheap per token is not cheap per task. On the dashboard, the model with half the token price cost the same money and took twice as long. Price per million tokens only becomes price per job after you know how many tokens the job takes and how many attempts it needs.
The one place Argon clearly wins is simple, well-specified work. The bug fix cost a third as much and was correct. That is a real result, and it points at the sensible way to use a model like this: cheap and fast where the task is narrow, the expensive one where the output is the product.
Related
- Gemini 3.7 Flash: What Rank 23 Actually Buys You — the same question from the other end: how little you can pay before quality becomes the issue.
- Claude Fable 5 vs Claude Opus 5: One Prompt, Two Models, One Excel File — two models from one vendor, one prompt, and what you give up by paying less.
- Google Antigravity 2.0: The Settings Screen Is the Story — the tool that ran the Google side of this comparison.