Field report 01 · 16 October 2026
It beats Claude Opus 5.5 and GPT-6 Astra on business automation and on reading what it sees, and it tops two boards of graded professional work. It trails Opus on code quality, whole programs, long unattended runs and design. And after two days of putting it to real work in my own studio, the biggest thing I learned is that the setup you run it in matters as much as the model.
02 · The shape of the evidence
Every benchmark that scored Argon beside Opus 5.5 or GPT-6 Astra: 93 of them, from Google’s own launch table, independent leaderboards and other model makers. Each dot is one benchmark, placed by how far Argon finished above or below the better of the two.
How the scale works. Most boards score out of 100, so a dot sits at the plain difference in points. Boards rated by votes or judging panels (Elo) are turned into the share of head-to-head wins the gap predicts: 133 points behind is about 32% of wins, shown as −18. Vending-Bench, a simulated year’s bank balance, is a percentage difference, and Gray Swan’s attack success rate is flipped, since lower is better there. 16 dots come from Google’s own launch material; where it repeats a third party’s number, it’s shown once. Each board ran its own setup and effort levels, so a dot is that board’s result, not a like-for-like test.
All 106 results as a table ↓03 · What people made with it
I collected 354 posts and videos from the first day and a half of general release: 192 on X, 110 moments in 35 YouTube videos, 21 on Reddit, 16 on Hacker News, 10 on GitHub and 5 on blogs. Here are nine to start with, losses to Opus among them. Each one plays here through the platform it was posted on.
LexnLin on X · 13 Oct
A procedurally built alpine village (timber houses, river, bridges, pines, snow peaks) in three.js with no external assets, rendered in four visual styles. The poster says every style was decent where other models fell short on some.
asmirkn on X · 13 Oct
Gave Argon a 20-second gameplay clip and asked it to rebuild the game; 4 minutes later a playable three.js version came back.
EnvolDev on X · 13 Oct
Same prompt: Argon 4 min 20 s and $3.40; Opus 5.5 1 h 53 min and $23.46. Poster: Opus still ahead on volumetric clouds, locomotive mechanics and camera work.
Mr_Salio on X · 13 Oct
Poster: quality about the same, Opus ‘a little bit better’.
rubenssoto_ai on X · 13 Oct
Same task for both: a cinematic 3D animation explaining a Tesla-style motor. Argon $3.83, Opus 5.5 $16.36.
Skill Leap AI on YouTube · 13 Oct
A walk-through house with a live minimap of the presenter’s position; kitchen, dining, main bedroom and bathroom where the plan puts them; prompt-only extras included; a night mode.
akartit on X · 13 Oct
Picks Argon for car design (2,526 lines). Harness tip: in opencode set OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX=65536, because Argon writes big files in one go and the default 32K cap silently drops the write.
thtbee_ on X · 13 Oct
A futuristic battle helicopter made purely in three.js, procedurally, from a multi-round self-iteration loop between Argon and Nano Banana; workflow to follow.
vib3coded on X · 13 Oct
Opus built a whole little world (dolphins, gulls, lighthouse, rocking bottle, interactions); Argon simpler for 64x less.
Ratings are my research’s reading of the work shown, not the poster’s claim. The thinking level is the poster’s, where they gave one; most didn’t.
Browse all 353 ↓04 · In my own work
From the day after release I handed Argon real jobs across my studio, and had Claude Opus 5.5 grade each one against what Argon actually ran, not its own account of it. Two things came back: where it’s genuinely useful to me already, and how much the setup it works in decides the result.
The way I look at it, a model is only as good as the workshop it’s working in: the instructions it’s given, the tools and skills it can reach, and how much of the job it can hold in its head at once. My Claude setup has a lot of that work in it.
On day one, Argon got Google’s own app, Antigravity, which condenses a conversation’s memory once it reaches about 128,000 tokens. So on day two I moved it into a different agent app, DeepSeek Harness, running Argon on Google’s Vertex AI, where its million-token window only condenses at about 839,000, and set it up much closer to how Claude works for me: my core instructions, memory, skills, tools and computer use. It still isn’t level with my Claude setup.
My rule for comparing models
A comparison only counts when every model worked in an equally good setup. Otherwise you’re measuring the workshop, not the model.
So any result from an uneven setup is marked unestablished until I run it again on even terms. Real work on even terms is the strongest evidence there is; benchmarks and other people’s tests don’t decide it.
Day one · Google’s Antigravity app
Day two · DeepSeek Harness on Vertex AI
Same model, same job, two harnesses
To check it wasn’t just a feeling, I gave Argon the same job twice on 15 October, once in each harness, with the same brief, the same starting code and nobody steering: fix five faults in one of my tool servers, with tests. A fresh Claude Opus 5.5 graded both blind, with the harness names hidden.
Both passed all nine hidden checks. The grader called DeepSeek Harness’s fix the better by a modest margin: stronger tests, nothing test-only left in the live code, and 2 unsupported claims against 4. It went live in my system that morning, after Claude repaired one edge case.
What it can’t show: one job, run once in each harness. And the cost runs the other way only because DeepSeek Harness had no caching that morning; it’s been fixed since.
What condensing did
On day one, in Antigravity, Argon did three long infrastructure jobs for me. The code was fine: 171 of 171 and 38 of 38 tests passed, and all six tool servers it wired up answered. But two of its written reports, one after five rounds of condensing and one after seven, described settings, files and test results that didn’t exist. In one long report-building job the app condensed eleven times in under an hour, and every invented figure in it was written within one to four minutes after a condense, without going back to the source.
In DeepSeek Harness, sixteen write-up batches each ran as one unbroken turn of 9 to 46 minutes, up to 488,000 tokens, and never condensed once. Taking the condensing away took away that run of invented figures. Smaller slips remained, like a tool name built from two, a detail joined to the wrong entry, or a card that dropped a qualifier. Which is why every batch still gets an Opus grade.
One long job in Antigravity · the hour it ran
11condenses in 55 minutes, each one a summary in place of the conversation
The longest write-up batch · DeepSeek Harness
0condenses in 46 minutes and 91 tool calls
The biggest · DeepSeek Harness
0condenses as its memory grew from 63,000 to 488,000 tokens in one turn. Graded good as delivered.
The graded jobs
Each was graded against Argon’s own tool calls and output, never its summary. Two came back good as delivered, eight close to good and eleven usable once fixed. None was unusable. Most are write-ups of dated entries in a long private record, from evidence files and transcripts; the rest are the infrastructure jobs and the head-to-head above.
The first-day tests
Fixing planted faults in a small job service
10/10hidden checks passed, and passed again on two reruns. It took 6.4 minutes and cost about $0.85. Its fix was the smallest of six runs on the same test; five GPT-6 Astra and GPT-6.1 Sol runs ten days earlier passed them too.
Shows: clean, scoped repair. Can’t show: how it compares; the others ran in their own harness, and Argon ran it once.
Watching and listening to clips from my films
7 of 8jobs where Argon ranked above Gemini 3.8 Flash, graded blind across 24 calls on seven clips. All four times a model invented on-screen text, it was Flash. Only Argon at High heard one four-second sound the way I did: a struck wooden knock. It also caught a headless mannequin growing a small human hand two and a half minutes into one of my films.
Shows: both ran through the same tool with the same prompts, so this one is on even terms. Can’t show: more than seven clips, one run each.
Judging AI video takes for a miniature film
10 of 14shots where its pick matched the frame-checked pick (3.8 Flash: 13). But several of its misses rested on camera moves it read right and Flash didn’t, and its trim points landed on the frame. At High it twice spent its whole 32,768-token allowance thinking and answered nothing.
Shows: Medium did as well as High here, at half the cost. Can’t show: the prompts were close, not identical, and the frame-checked picks are a judgement call, not ground truth.
Building a 3D industrial scene in Blender from a brief
Unestablished28 minutes, about $7.58 at list price, against 51 minutes, $19.67 for Opus 5.5 and 56 minutes, $15.48 for Sonnet 5.5 on the same job earlier. Judging the three blind, I found its main close-up the flattest, more cartoon than real. But it ran in Antigravity, whose condensing lost parts of the brief.
Shows: its speed and cost. Can’t show: whether its 3D holds up, until I run it again on even terms.
05 · Running it
Argon has three thinking levels: low, medium (the default) and high, and nothing above high. Google publishes no results by level, but three independent boards ran more than one. In each chart the gold line joins Argon’s levels.
FrontierCode
47.55 at Medium against 45.87 at High, for $6.74 a task against $11.11.
Cognition’s FrontierCode 1.1, five trials per setting. Argon ran in Cognition’s own harness, Opus in Claude Code, Astra in Codex.
AutomationBench
51.29 at High against 50.08 at Medium, for $1.70 a task against $1.54.
Zapier’s AutomationBench 1.0.6, at standard list prices.
SenseBench
95.02% at Low against 95.12% at High, for a $7.09 run against $37.42: over five times the cost.
SenseBench, 4,861 items. Submitted on 13 October, not yet accepted. Argon at Low used no thinking tokens at all.
One YouTube tester’s hard-logic runs went the other way: Medium failed six of six, High passed sixteen of sixteen. And in my own listening test High was the one that heard the knock. So test your own job at both.
The empty-answer trap
Argon’s thinking comes out of the same allowance as its answer. Judging four takes of one shot at High, with my tool’s output cap at 32,768 tokens, it spent the whole allowance thinking and returned nothing, twice. Medium never failed, at half the cost and in well under half the time.
Google’s advice is to never set the cap below 32,768. At High, give it the full allowance, 65,536 when streaming.
What it costs
On Vertex AI and the Gemini API, Argon costs $2 / $10 per million input and output tokens until 30 November, with cached input at $0.10. From 1 December it’s $4 / $20, the same as Claude Opus 5.5. Its US and EU endpoints cost 10% more, and on Vertex the launch price is paid as 50% credits back on net spend.
Cheap tokens aren’t cheap tasks, though. Cursor measured Argon taking 115 steps where Opus took 54 for the same score.
Per million tokens: input filled, output outlined.
How I brief it
Google’s developer guide and my two days of jobs agree on more than they differ.
06 · By kind of work
My reading of the strongest evidence for each, public and my own, against each rival. Where the evidence is thin, it says so.
07 · The record
All 106 benchmark results
Filter by who ran them, by area or by name. Each links to its source, with that board’s notes.
Loading…
All 353 posts and videos
Filter by kind of work and by how it went. Each plays on the platform it was posted to. One of the 354 collected has since been deleted, so it’s left out.
Loading…