Gemini 4 Argon goes generally available at $2 per million tokens, and the price doubles on December 1

Google made Gemini 4 Argon generally available on October 13, 2026, and the number that matters most to anyone building on it is a date: the introductory price of $2 per million input tokens ends on November 30, and on December 1 it doubles to $4. If you are asking what Gemini 4 Argon costs, the honest answer is that it costs half of what it will cost in seven weeks.

The launch widens access well beyond the cyber defenders who got the model first. It also lands in a pricing market where the rivals' flagship rates have not moved much, which makes the intro window the real story.

What shipped on October 13

Koray Kavukcuoglu, Google DeepMind's senior vice president and chief AI architect, framed the release in the launch post this way: "We're giving developers, enterprises, scientists, and consumers access to our latest frontier model."

Paying API customers get Argon through Google AI Studio and the Gemini API, with a 1 million token unified input and output limit that Google calls industry-leading. Google AI Ultra subscribers get it in Antigravity, the company's agentic development platform. The launch post lists the Gemini app and AI Mode in Search for Ultra and Pro subscribers, while the AI Studio pricing page states that the model "is not available on the Free tier or on the Google AI Pro plan." Those lines describe different surfaces and they read awkwardly together, so check which plan covers your use before you budget.

Scientists can apply for six months of free access to Antigravity's top usage tier, with 10,000 places available on a rolling basis. Cyber defenders in the Fairwind Program keep access without cyber guardrails, and Google says it began expanding that to hundreds of additional partners on launch day. We covered the first, restricted rollout in our report on the cyber-defenders-first release.

The price, tier by tier

Google's pricing page lists four paid tiers, each priced per million tokens, and every one of them doubles on December 1, 2026:

• Standard: $2 input and $10 output now, then $4 and $20.
• Batch and Flex: $1 input and $5 output now, then $2 and $10.
• Priority: $3.60 input and $18 output now, then $7.20 and $36.
• Cached input on Standard: $0.10 now, then $0.20, with a 95 percent discount on cached tokens in the launch post.

Set that against what rivals publish today. OpenAI's pricing page lists GPT-6 Astra at $10 input and $50 output per million tokens, rising to $20 and $75 above 272,000 input tokens. Anthropic's pricing page, read on October 14, lists Fable 5.1 at $10 and $50, Opus 5.5 at $4 and $20, and Sonnet 5.5 at $2 and $10. So Argon at its introductory rate matches Sonnet 5.5 and costs a fifth of the two top-priced flagships. At its December rate it matches Opus 5.5.

Gizmodo, which covered the launch, put the same comparison in plain terms: Argon at $2 and $10 against GPT-6 Astra at $10 and $50. The comparison is fair on list price. It says nothing yet about cost per finished task, because a model that needs more thinking tokens or more retries can erase a price edge.

What Google says it can do, and what is unproven

Google reports that Argon ranks first on the Vals Index, which covers finance, coding, legal and tax work, and scores 87 percent on ERQA, an embodied reasoning test. It cites private evaluations by Gamow Labs for rare disease diagnosis assistance (57 percent) and wet-lab reasoning (52.2 percent), and says the Gray Swan prompt injection benchmark showed 0.7 percent attack success across 15 adaptive attempts.

Inside Google, the company says agentic code submissions in Antigravity rose more than 35 percent and code rollbacks fell by nearly a third. It also describes Argon refactoring about 6 million lines of a 40 million line Chrome codebase in weeks rather than months.

Every one of those figures comes from Google or from a partner Google chose. The Gamow evaluations are private. Gizmodo notes the launch carried no independent benchmark scores and reports doubts among some Google employees about coding performance. Treat the Vals Index claim as a lead to verify on the index itself, not as a settled ranking.

The safety claims

Google says it evaluated Argon under its Frontier Safety Framework and built a layered defense of jailbreak resistance, monitoring and enforcement. It reports that external red teams could unlock less than 1 percent of the biological sub-tasks tested, that violations on an adversarial harmful-request set fell nearly twofold against earlier Gemini models, and that shortcut attempts on coding benchmarks dropped to 2 percent.

It also says an automated monitor reads the agent's reasoning and actions in real time, and that training avoided penalizing the model's thoughts so its reasoning stays monitorable. That last design choice is the one safety researchers have been asking for, and it is worth watching whether the model card shows the evidence behind it. For why that matters, see our explainer on chain-of-thought monitorability.

What we would do

If you are a builder, test Argon on your own tasks during the introductory window and log cost per completed task, not per token. Do not lock in a margin on the $2 rate, because the December rate is already published and it is double.

If you are following the race, the useful signal is not the price cut. It is that Google chose to open a frontier model to paying customers with a time-limited discount and a flat doubling, a pricing shape that tells you it wants usage data fast. Wait for independent benchmark results before believing the Vals Index headline.