Google announced Gemini 4 Argon on September 30, for cyber defenders in its Fairwind program only. Paid API access opened on October 13. Google’s own benchmark table puts it behind Claude Opus 5.5 on coding: 57.4 against 66.4 on Terminal-Bench 4.0, 55.0 against 62.3 on FrontierSWE. Reviews agreed. On Hacker News, meowface put it plainly:
— meowface, Hacker NewsFor agentic coding, Claude Code with the latest Fable or Opus is still just better.
My bench says the opposite, and I am honestly puzzled by it. I expected Argon to land behind Opus, somewhere near Sonnet, the way the vendor table and the reviews said it would. Instead it got the best result I have measured on real work. So I looked for the mistake: a test leak, a drifted bench, a harness that favours it. I did not find one. The bench also says that accuracy is not the whole story.
The Sticker
| Per million tokens | Input | Cached input | Output |
|---|---|---|---|
| Gemini 4 Argon, until November 30 | $2 | $0.10 | $10 |
| Gemini 4 Argon, from December 1 | $4 | $0.20 | $20 |
| Claude Opus 5.5 | $4 | $0.20 | $20 |
| Claude Sonnet 5.5 | $2 | $0.20 | $10 |
| GPT-6.1 Sol | $2 | $0.10 | $10 |
The dates are on Google’s pricing page. After the introductory period, Argon costs exactly what Opus costs.
The Same Six Changes
Same bench as the Opus 5.5 vs GPT-6 Sol and Sonnet 5.5 runs: six merged changes from a production monorepo, each replayed on a clean snapshot with no git history, graded by the tests the original fix shipped with. Same prompts, medium effort, 20 runs. Argon ran in Claude Code through OpenRouter. A leak check found no target test in any snapshot and no reads outside it.
| 20 runs each | Argon | Opus 5.5 | Sonnet 5.5 | GPT-6.1 Sol |
|---|---|---|---|---|
| Runs that passed every test | 16 | 13 | 11 | 11 |
| Runs that broke a test that passed before | 0 | 0 | 0 | 5 |
| New tests fixed | 98% | 93% | 88% | 94% |
| Original-fix files also changed | 90% | 85% | 82% | 86% |
| Output tokens | 1.37M | 495k | 247k | 167k |
| Total cost at list price | $57.68 | $45.81 | $11.76 | $6.73 |
| Change | Argon | Opus 5.5 | GPT-6.1 Sol |
|---|---|---|---|
| A: 15 files, 1 workspace | 4/4, $1.84 | 4/4, $1.12 | 0/4 |
| B: 17 files, 3 workspaces | 4/4, $3.21 | 0/4, $3.36 | 3/4 |
| C: 26 files, 3 workspaces | 0/3, $6.34 | 0/3, $4.76 | 0/3 |
| D: 15 files, 1 workspace | 2/3, $1.34 | 3/3, $0.98 | 2/3 |
| E: 11 files, 3 workspaces | 3/3, $2.75 | 3/3, $1.70 | 3/3 |
| F: 13 files, 2 workspaces | 3/3, $2.07 | 3/3, $1.85 | 3/3 |
- Change B decides it. In a rerun on the current Claude Code, Opus failed the same two tests in all four runs: the ones that check a complete list of codes. Argon passed them every time. Sonnet passed one run of four.
- It keeps what works. On Change A, Sol rewrote a user-facing message in all four runs and dropped the instruction it carried. Argon kept it every time, like both Claude models.
- It writes a lot. 1.37M output tokens is 2.8x Opus and 8x Sol. One Hacker News commenter pointed out that “argon” comes from the Ancient Greek for “lazy or inactive”. It is the busiest model I have tested. Artificial Analysis measured the opposite on its index: 62k output tokens per task for Argon, 119k for Opus. In a coding agent at medium effort, Argon explores and rereads more.
- The bill follows the tokens. At half the token price, Argon cost 1.26x Opus. After December 1, the same 20 runs cost about $115, or 2.5x Opus.
— AravPhi, Hacker Newsthey named it argon because after using it your token ar gon
Simon Willison made the point on launch day: “what really matters is how many tokens a model takes to achieve a goal within an agentic harness.” Argon is the clearest case yet. Half the sticker price, more than the total cost.
The Cache Trap
My first Argon run cost $17.60. It read 8.4M input tokens and only 59k came from cache. Claude Code adds a system-role message on most turns. OpenRouter moves every system message into Gemini’s system prompt, so the start of the prompt changed on every call and Gemini’s implicit cache never hit. Argon has no explicit caching to fall back on.
I fixed it in claude-launcher 0.11.1: a local proxy moves those messages into the previous user turn. The same run then cost $2.62, with 94% of input from cache. LiteLLM has an open issue with the same cause on Vertex. If you run Argon in a third-party harness, check your cache hit rate before you judge the price.
Is It the Harness?
Gemini users blame Google’s tools. LewisOrton on Hacker News: “Gemini models are in fact very good. Google’s harness(es) for Gemini are nightmares, terrible.” So I ran Change B in Google’s Antigravity CLI too, and reran Opus on the current Claude Code.
| Change B | Claude Code | Antigravity |
|---|---|---|
| Argon | 4/4, mean 991 s | 4/4, mean 1,740 s |
| Opus 5.5 | 0/4 | 0/1 |
Argon solved B in both harnesses. Opus missed the same two tests in both. Antigravity was 1.75x slower. Its Ultra plan also let Opus run exactly once before a quota error, which is one way to win a benchmark. That cell is a single run. The model decides this change, not the harness.
Vision: Last on Screenshots
Argon’s launch pitch leans on multimodal work. I ran a second bench against Opus 5.5 and GPT-6 Astra, three runs per task.
| Mean score (0 to 1) | Argon | Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Screenshot to HTML, 6 pages | 0.754 | 0.831 | 0.836 |
| SVG from a text prompt, judged blind | 0.614 | 0.503 | 0.733 |
Argon placed layout least precisely and came last on five of six pages. A blind panel of three models ranked Astra first on SVG and Argon second, and no judge favoured its own model. Chart-to-SVG and image-to-JSON were too easy: every model scored near 1.0. That matches jjcm’s test on Hacker News: “I’d put it at B tier, just under gpt 6.1 sol. It isn’t A tier like Opus 5.5 or Astra.”
What Argon Doesn’t Solve
- Cost. It is the most accurate model on this bench and, from December 1, the most expensive one by a wide margin.
- Speed. It was slower than Opus in both benches: 991 s against 455 s per run on Change B, and 3x slower on screenshots.
- Access. Antigravity needs Google AI Ultra on a personal account. My Workspace account was refused. Google will take my money, just not from that address.
- Sample size. Six changes from one codebase, 20 runs. A different codebase puts a different change in the slot where B sits.
Lessons
- The vendor table undersold it. Google’s own numbers put Argon behind Opus on coding. On real changes it was ahead.
- Measure cost per fix, not per token. Argon’s half price became a higher bill.
- Check the cache. A harness that breaks the prompt prefix multiplied my cost by more than six.
- Run your own tasks before December 1. That is the cheapest Argon will be.
Related Posts
