Gemini 4 Argon: Cheap, But Don't Underestimate It
A new model announcement alone doesn't cut it. The real questions are whether it's worth using for your work, and whether it's worth the money. Short answer: Argon sits at the bottom of its class on price, while Google's announced benchmarks put it at the top. Independent scoreboards, though, are still raising an eyebrow. Below: why opinions split, what it can do, and who's already using it.
The short version
Price — $2 in · $10 out per 1M tokens, bottom of its class (promo until Nov 30)
Performance — top-tier on Google's numbers, e.g. 77.9% on DeepSWE
Catch — on the independent AAII index it trails Anthropic and Meta
Argon at a glance
- Google's frontier model "Gemini 4 Argon" is out — the first model of the Gemini 4 series.
- It's built to sustain deep reasoning across complex, long-running workflows, with frontier-level performance claimed for software engineering, legal and financial knowledge work, and cyber defense.
- Rollout is staged: through the Fairwind Program, vetted security professionals get it first, with access widening over time.
- The output token limit jumped from 64K to 1M, and the Gemini API supports up to 1M tokens combined across input and output.
Sources: Google Blog · AI Times
Next to its peers
Price (checked manually on 2026-10-14, USD per 1M tokens)
| Model | Input | Output |
|---|---|---|
| GPT-6 Astra | $10 | $50 |
| Claude Opus 5.5 | $4 | $20 |
| GPT-6.1 Sol | $2 | $10 |
| Gemini 4 Argon |
Argon's price is a promo until Nov 30; from Dec 1 it's $4/$20. Cached input tokens are 95% off. Prices change — recheck before you rely on this.
Sources: Google AI pricing · OpenAI pricing · Claude pricing
Benchmarks (Google's announcements)
- 77.9% on DeepSWE v1.1, the long-horizon software engineering eval — a state-of-the-art claim.
- Led the Vals Index (economic-impact-weighted across finance, coding, legal, tax), plus Vals Finance Agent v2 and Harvey's Legal Agent Benchmark.
- 51.3% on Zapier's AutomationBench, #1 for end-to-end business task execution.
- 91.7% on LVBench for long-video understanding.
Source: Google Blog
The cold water
- On the Artificial Analysis Intelligence Index (AAII), it landed behind Anthropic's and Meta's frontier models, and it placed 5th overall on Arena.ai's Agent Arena.
- Outside expert Sam Witteveen (co-founder & CEO, Red Dragon AI) called it a really good model but wouldn't agree it's the best on the market — and he separated the model from the tooling, saying he doesn't think Antigravity is as good as Claude Code or Codex. Its strength, he noted, shows in work combining visual understanding, code generation, and creative production.
Source: AI Times
What Argon actually does
- Coding, end to end: everyday debugging to large codebase migrations and algorithm design — not just generating code, but investigating existing systems and testing the fixes. The human job shifts from writing to reviewing.
- Long-running work: multi-step jobs that hold context across documents and tools. Financial research, legal research and drafting, tax work. Keeping context unbroken is the whole point.
- Vision-heavy knowledge work: expert chart analysis, pulling details out of long videos, follow-ups based on document sets. It saves you watching the whole video from the start.
- Cyber defense: offered first to vetted security professionals through the Fairwind Program.
- Science: quantum algorithm optimization, quantum chemistry calculations — it cuts compute wait time and speeds up experiments.
- Where: paid API, Google AI Studio, the Gemini app and Search AI Mode (Pro and Ultra subscribers), Antigravity (Ultra only). Google also said it will give 10,000 scientists the top Antigravity tier free for 6 months (eligibility and selection criteria apply).
Sources: Google Blog · AI Times
Already out in the field
Inside Google
- Video decoder: it optimized the Rust port of the libgav1 open-source video decoder, replacing some 32,000 lines of SIMD code, and matched near-optimized C++ speed at 2.7x the previous Rust version — with memory safety and the same quality.
- Large-scale code conversion: from core libraries of tens of thousands of lines (re2, libgav1) up to the 800K+ line Fuchsia OS Zircon kernel, C/C++ codebases are being converted to Rust — with strict automated and manual audits, tests, and code review before anything ships.
- Data center memory: an Argon agent team analyzed profiling data across the fleet and autonomously found and applied memory optimizations, freeing 300+ TiB, with total savings projected at 500 TiB up to 1 PiB.
- Quantum algorithms: it optimized the space-time resources (qubits x gates) of bottleneck subroutines and beat previously published baselines by 40%+ in minutes.
- Chromium cleanup: it helped split out ~6M lines of legacy code inside the ~40M-line Chromium codebase across hundreds of files — months of expected work done in weeks. Teams on Antigravity report 35%+ more agent-based code submissions and rollbacks down by about a third.
Sources: Google Blog · AI Times
The ADK case — the wins and the rejects
- On Google's open-source Agent Development Kit (ADK, ~800 files, 205K lines of Python), Argon worked unattended for 11 hours in Antigravity, ran 16,892 public unit tests, and submitted 6 optimization proposals across 32 files.
- Only 2 were accepted. The other 4 were rejected for risks like returning stale data, depending on private implementations, or adding complexity with no payoff. The accepted changes cut in-memory session prep time by ~40% and package loading by 21%, while time-to-first-request in a fresh process barely moved.
- In the dev team's words: agents explore; maintainers remove what's unnecessary.
Source: AI Times
Companies and research
- Early participants: Deloitte, LexisNexis, Ambience Healthcare, Isomorphic Labs.
- LexisNexis reported positive results for legal grounding checks, contract review, and litigation analysis. Isomorphic Labs uses it to extract information from complex PDFs and patents, speeding up drug discovery work.
- The quantum chemistry package xTB, with Argon's help, predicts molecular structures up to 2.3x faster and searches conformers 1.8x faster.
- Gemma inference optimization: on NVIDIA A100 with vLLM, throughput went from ~37,900 to 108,000+ tokens/sec — a 2.85x end-to-end improvement.
- Demis Hassabis showcased asking the model to build a wizard chess game, highlighting its grasp of physical movement and 3D modeling.
Regular users are still exploring
- Korean user reports include a battery-tech meeting write-up that came back with made-up content and heavy token use, and errors in report summaries that used to work on older models.
- On the flip side, its Korean writing gets praise. It's the early phase where what works and what doesn't is still split.
Sources: Damoang user report · Clien
Korea's reaction is split
- Yonhap and AI Times covered the launch on 2026-10-14, and Korean-language X mentions jumped 10x+ versus pre-launch (per collected data).
- Praise: Korean writing quality. Doubts: coding and accuracy question marks, complaints about token appetite, and confusion from Pro subscribers who can't see it yet.
Puzz's one-liner
Bottom of its class on price, top of the class on Google's benchmarks — independent scoreboards and field reports are still split. The fastest way to judge is to plug it into your own work. Just note: the half-price promo ends Nov 30, and from Dec 1 it's $4/$20. The ones who try first and judge first usually come out ahead. Ribbit.