Google made Gemini 4 Argon generally available on 13 October 2026 and called it its most powerful model. Here is what the independent numbers say if you're deciding whether to build on it.
Access and limits
- Model ID:
gemini-4-argon - Available via the Gemini API (AI Studio, Gemini platform), plus Antigravity for Ultra subscribers
- Context: 1M tokens, shared across input, thinking and output
- Max output: 262k tokens
Pricing
| Period | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Until 30 Nov 2026 (intro) | $2 | $10 |
| From 30 Nov 2026 | $4 | $20 |
Cached input is 95% off. Budget at the full price - the intro rate ends in six weeks.
The leaderboard
On the independent Vals AI index (finance, law, tax, code):
| Model | Vals Index |
|---|---|
| Gemini 4 Argon | 68.9% |
| Claude Sonnet 5.5 | 67.0% |
| Claude Opus 5.5 | 67.0% |
| Claude Fable 5.1 | 65.8% |
A 1.9-point lead with about a point of noise. First place is real, but it's a pack, not a breakaway.
Coding benchmarks
| Benchmark | Score | Rank |
|---|---|---|
| IOI (competitive programming) | 100% | 1st, tied with GPT-6 Astra |
| Vibe Code Bench (app from a description) | 91.9% | 2nd, behind Claude Sonnet 5.5 |
| Code Migration | 68.2% | 2nd |
| Terminal-Bench 4.0 | 57.6% | 6th |
| SWE-bench Multimodal | 56.0% | 4th |
| ProgramBench (large engineering tasks) | 2.5% | 7th |
| CUA-bench (operating a computer UI) | 4.8% | 8th of 9 |
Strong on writing and migrating code. Weak on agentic computer use - don't pick it to drive a UI.
Latency and cost per task
On the Vals index, one task averages $15.68 and ~46 minutes. Long agentic tasks cost more: about $58 per Code Migration test and $194 per CUA-bench test. Argon is built for long, deep reasoning. For low-latency or high-volume calls, a smaller model will serve you better.
Vendor claims (not independently verified)
- 35%+ more accepted code changes and ~1/3 fewer rollbacks inside Antigravity
- Up to 2.3x faster xTB quantum chemistry runs
- 0.7% prompt-injection success rate on Gray Swan's IPI benchmark
Plausible, but they're Google's own numbers.
Takeaway
- Top of the general leaderboard ≠ best on your task. Run your own evals before switching.
- Use Argon where long context and deep reasoning pay for the latency: big codebases, migrations, document-heavy analysis.
- Look elsewhere for UI automation and fast responses.
We test models on real market tasks and publish the results, misses included: neuportal.ai/experiment. Full write-up: Gemini 4 Argon: Google Is Back on Top - by Two Points.
Top comments (0)