Google made Gemini 4 Argon broadly available yesterday, 13 days after unveiling it to a small group of security defenders. The headline is the price: $2 per million input tokens and $10 per million output tokens. The fine print is the date. That rate ends on November 30. From December 1 it's $4 and $20.
Cached input is even more eye-catching at $0.10 per million tokens, a 95% discount. Same deal, same expiry.
What the numbers actually say
At the intro price, Argon undercuts Claude Opus 5.5 by half and GPT-6 Astra by a factor of five. But the intro price is just matching Claude Sonnet 5.5 and GPT-6.1 Sol, which already sit at $2 and $10. After December, Argon lands exactly where Opus 5.5 and Meta's Muse Wonder are priced.
So the real question isn't whether Argon is cheap. It's whether it's worth Opus money, because that's what it will cost you in about seven weeks.
The independent scores are good but not dominant. Vals AI has it first on its index at 68.9%, narrowly ahead of Sonnet 5.5 and Opus 5.5. Artificial Analysis puts its Intelligence Index at 53, which trails Opus 5.5 (58) and Sonnet 5.5 (56). On Terminal-Bench 4.0, Google's own numbers show Argon at 57.4% against 66.4% for Opus 5.5. It wins on agent-style tasks like Zapier's AutomationBench and Finance Agent v2. It loses on terminal-heavy coding.
That split is the whole story. A model can lead one benchmark and trail another by ten points, and neither tells you how it handles your prompts.
How we'd test it
Don't run a bake-off on the promo rate and call it done. Do this instead:
- Replay real traffic. Take a few hundred actual requests from your product, with the real system prompts and tools. Not a public benchmark.
- Cost it at $4/$20. Compute cost per completed task, not cost per token. A cheaper model that needs two retries isn't cheaper.
- Check the cache math. If your workload has long, stable prefixes, the cached rate matters more than the headline. Confirm what happens to it after December 1 before you rely on it.
- Keep a switch. Route by task type behind your own interface so moving a workload costs a config change, not a sprint.
If Argon wins on your tasks at the December price, great, you've found a real option. If it only wins at $2, you've found a discount, and discounts end.
The agent workloads are where it might earn its place
Google reports that more than 100,000 of its own employees use Argon through an internal Antigravity build, with a 35% rise in agentic code submissions and 30% fewer rollbacks over four weeks. That's a vendor's number about its own staff, so treat it as a hint. Still, the agent benchmarks are where Argon looks strongest, and agent loops are exactly where token spend gets ugly. It's worth testing there first.
One more detail: a reviewer who rates the model highly also says Google's Antigravity tooling isn't yet as strong as Claude Code or Codex. The model and the harness around it are separate decisions. Test them separately.
The takeaway
Introductory pricing is a marketing tool, and it works. Teams adopt in October, build around the cost in November, and get a surprise on the December invoice. Price every model at its list rate, measure on your own traffic, and treat anything cheaper as a bonus you can lose.
We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!