Gemini 4 Argon: Google's New Frontier Model, Minus the Fine Print

Argon walks into the AI race

For most of the past year Google looked like it was running the frontier model race in flip-flops. Gemini 3.5 Pro, widely reported to have missed its internal targets, never shipped at all. Now Google has skipped ahead with Gemini 4 Argon, and it has arrived with a pile of first-place benchmark results and a very aggressive price tag.

The name is a fun one. Argon comes from the Greek argos, meaning "lazy" or "idle", which is a bold thing to call a model you want people to pay for. Naming has never been Google's strong suit. The model, though, deserves a closer look.

What's actually new

Argon isn't another fine-tune of the old Gemini. Google trained it from scratch, and that shows in how differently it behaves. Here's what's been shared so far:

• Architecture: a sparse mixture-of-experts transformer, trained on Google's own TPUs with JAX.

• Knowledge cutoff: March 2026, which points to a training run of at least six months.

• Rollout: announced on 30 September for a hand-picked group of security researchers, who reportedly got access without the usual guardrails. Everyone else gets the guarded version, which Google says refuses far less often than earlier Gemini models.

• Modalities: an omni model that accepts image, video and audio, and can generate them too.
  

Read the bold numbers, then read the rest

The headline results are self-reported, so we treat them as claims for now. Taken at face value, they're strong. Argon leads on a long-horizon software engineering benchmark, tops an index that weights finance, legal and coding tasks by their share of US GDP, and scores 91.7% on LVBench for long-video understanding.

The rows Google didn't bold are more revealing. On Terminal-Bench 4.0, Claude Opus 5.5 beats it 66 to 57. GPT-6 Astra wins another frontier benchmark by more than ten points. On LMArena, where people vote on answers blind, Argon takes first place for text and drops to ninth for web development.

First for text, ninth for web dev. That gap tells you more than any single headline score.

The price is the real headline

Until 30 November, Argon costs $2 per million input tokens and $10 per million output tokens. For a model in this class, that's cheap enough to change some build-versus-buy maths. The developer launch was bumpier. Argon replaced Gemini 3.1 in Antigravity, Google's agentic IDE, but only for top-tier AI Ultra subscribers at first. Pro users got it a few days later, with smaller quotas.

How it holds up on real tasks

Early hands-on comparisons back up the benchmark split. When asked for a promo video, Argon produces something polished that still looks obviously AI-made. Claude can't output raw video, but it gets to a stronger result another way: it writes the code that renders the motion graphics. That makes the result much easier to tweak, inspect and version than a diffused clip.

A tougher test is a realistic, interactive 3D model in Three.js. Claude and GPT-6 Astra both return convincing first drafts with fixable flaws. Argon's interface looks excellent, but the model itself comes out low-poly despite the prompt asking for realism, and it adds a few anatomical details nobody requested. It looks good, but it didn't follow the brief.

What this means for teams building with AI

One thing is clear: no single model wins everywhere. Argon looks like a great pick for long-context reading, video understanding and anything where cost per token matters. For agentic coding, the people who use these models every day still lean towards Claude and GPT-6.

That's how we build at Birdhouse. Rather than tying a project to one vendor, we:

• Route by task: Birdhouse OS sends each job to the lightest model that can handle it and escalates only when it needs to.

• Test on your workload, not the leaderboard: a new model earns a place in production after it beats the current one on the client's own tasks.

• Review adversarially: one model's output gets checked by another, and a person signs off on anything that matters.

• Prefer artefacts you can inspect: code that renders a video beats a video you can't edit.

Argon is a real comeback, and the price will shake up the market. Use it where it's strong and test it before you trust it everywhere.