Gemini 4 Argon Arrives With a Price Cliff Set for November

Google released Gemini 4 Argon on 13 October at an introductory price of $2 per million input tokens — around a fifth of what its frontier rivals charge.

Output tokens run $10 per million, and DeepMind says cached input tokens are discounted 95%. The Verge reports those introductory rates double at the end of November, when the introductory period finishes.

What you can use today, and where

Argon is globally available to developers and enterprises through the API in AI Studio and the Gemini platform. Google AI Ultra subscribers get it in Google Antigravity, Google's agentic development platform, and both Ultra and Pro subscribers see it in the Gemini app and AI Mode in Search.

There is a gap worth noting before you plan around it: The Verge reports it is not clear when — or if — Argon will reach Plus subscribers or free users. If your product assumes a free Gemini tier, nothing has changed for you yet.

As with other recent frontier releases, the public rollout followed a limited release to a selected set of "trusted cyber defenders" for risk testing. Google says it is now expanding access to hundreds of additional partners in its Fairwind Program.

The benchmark claims, and who is making them

Google claims Argon outperforms GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across coding, cybersecurity, science, finance and law. DeepMind says the model ranks #1 on the Vals Index, which measures performance across economic sectors including finance, coding, legal and tax work.

Two more numbers come from the company's own post. On the Embodied Reasoning (ERQA) benchmark, which tests spatial problem solving in 3D environments, DeepMind says Argon scores 87% using pure reasoning. On private evaluations built from clinical and lab data at Gamow Labs, Google reports 57% on RareBench for rare-disease diagnosis assistance and 52.2% on LabBench for wet-lab reasoning.

Every figure in this section is vendor-reported. Google names one outside index, Vals, but the scores themselves come from its own runs, and none of this has been independently reproduced yet.

The internal deployment numbers are the more interesting signal

Google published specifics on how its own teams used Argon, which are easier to sanity-check than a leaderboard position.

• Chrome refactor: DeepMind says Argon helped separate core webpage handling from a roughly 6-million-line legacy folder inside Chromium's 40-million-line codebase, finishing an architectural refactor across hundreds of files — "months of work in weeks," per the company.
• Agent Development Kit: working autonomously overnight, Argon produced optimisations now live in google/adk-python that Google says cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%.
• Antigravity usage: Google reports teams increased agentic code submissions by more than 35% while reducing code rollbacks by nearly a third.
• Quantum chemistry: DeepMind says the open-source xTB package became up to 2.3x faster at predicting molecular structures and 1.8x faster at conformer search, with changes contributed back to the official repository.

Google also says it used Argon to halve the computing power behind YouTube's comments section, and to translate an internal version of its AlphaQubit quantum error-correction decoder into a single optimised hardware kernel that nearly doubled inference throughput.

The cheap-token claim has a catch

Per-token price is not per-task cost. The Verge makes this point directly: a lower per-token rate "doesn't necessarily translate to lower costs in practice, as models can use different numbers of tokens for the same tasks."

That caveat matters more for a reasoning model than for a chat model. If Argon spends more tokens thinking through a multi-step task than the model you run now, a fifth of the list price can still land at a similar invoice. The only way to settle it is to replay a real workload through both and compare the token counts, not the rate cards.

The 95% cached-input discount cuts the other way, and is the number to pay attention to if your workload repeats a large system prompt or document across many calls.

Safety and agent-behaviour figures

Google reports that under its Frontier Safety Framework, external bio red teaming could unlock less than 1% of the bio sub-tasks it tests for, and that violations on an adversarial stress-test set dropped nearly two-fold against previous Gemini models.

One figure deserves attention from anyone running agents in CI: Google says it trained Argon to respect system boundaries, report negative findings honestly, and pause before irreversible steps — and that on coding benchmarks where models have no incentive to hold back, shortcut attempts dropped to 2%. The company also describes an automated misalignment monitor as a defence-in-depth layer. Again, these are Google's own measurements.

What to watch next

Three dates and numbers are worth putting in a calendar. The end of November, when introductory pricing ends and the real rate takes effect. The first independent benchmark runs, which will test whether the Vals Index position holds outside Google's framing. And whatever Meta's Muse Wonder Full, announced a day before Argon, turns out to cost — Argon is Google's first major frontier release since Gemini 3 in late 2025, and the gap between frontier launches is now measured in days.

If you are evaluating a switch, start with the cached-input discount against your own prompt structure. That is the one lever where Argon's pricing is unambiguous.