Gemini 4 Argon — Google retakes the lead at half Anthropic's price, but almost nobody can use it yet

Google announced Gemini 4 Argon on 30 September 2026, a model that leads or ties on 13 of the 18 tests the company chose and costs, at launch, half of what Anthropic charges for Claude Opus 5.5.

The stake is not a ranking won but a comeback: Google dropped the model it had promised for June, reorganised its research lab in August and spent months behind Anthropic and OpenAI. This article's verifiable thesis: Argon is a model that clearly leads on long tasks and office work, not on programming in the terminal, and its aggressive price cannot, for now, be paid by almost anyone. The dominant scenario: a broad launch by the end of the year, with a price response from rivals.

What Google announced, and with which numbers

The official announcement, signed by Koray Kavukcuoglu, senior vice president of Google DeepMind, calls Argon “the next era of frontier intelligence”, built for “deep reasoning across complex, long-horizon workflows”. The fields named are software engineering, office work in law and finance, and cyber defence.

The model is larger than the “Pro” line so far, according to Reuters, and the output limit rises to one million tokens in a single answer, from 64,000. Google has published neither the size of the input context window nor a model card.

The launch price is $2 per million input tokens and $10 per million output tokens, and cached input tokens cost 95% less, i.e. $0.10. After the launch period, whose length is not stated, the price doubles to $4 and $20.

Access, however, is restricted. The model goes first to “trusted cyber defenders” through the Fairwind programme, launched on 2 September with more than 650 partners — national cybersecurity authorities, critical infrastructure operators, technology and security firms —, according to TechWire Asia. Use is limited to employees working in cybersecurity, with multi-factor authentication. Google AI Ultra subscribers and paying API customers come next, with no date announced.

The price: half of Anthropic, a fifth of OpenAI

The comparison with rivals shows how aggressive the launch price is. On output, where most of an agent's work is paid for, Argon charges $10 per million, Claude Opus 5.5 charges $20, and GPT-6 Astra and Claude Fable 5.1 charge $50 each. Even the post-launch price, $4 and $20, equals that of Opus 5.5 rather than exceeding it.

In lei, at the NBR rate of 4.6474 published on 30 September, the difference becomes concrete. A typical agent task, with 200,000 tokens read and 50,000 written, costs $0.90 on Argon (about 4.18 lei), $1.80 on Opus 5.5 (8.37 lei) and $4.50 on GPT-6 Astra (20.91 lei). A single one-million-token answer, possible only on Argon, would cost $10, about 46.47 lei.

One detail shows where Google wants to strike: the launch output price, $10, is below that of its own Gemini 3.1 Pro, $12, even though Argon is a larger model. Rival prices are the list prices published by Anthropic and OpenAI via OpenRouter; batch discounts and large contracts are not part of the comparison.

Performance: where it wins and where it loses

The table below gathers the figures published by Google, as reproduced by DataCamp, VentureBeat and MarkTechPost. A dash means no result was published for that model.

The pattern is clear. Argon leads on long tasks and office work — law, finance, automation, long video, reasoning over large documents —, exactly where a one-million-token output limit matters. On pure programming the result is split: it wins DeepSWE but loses FrontierSWE v2 to GPT-6 Astra and Terminal-Bench 4.0 to Opus 5.5, by 9 percentage points. Reuters sums it up the same way: behind on “two of the four” coding tests.

On agent security, Google reports an attack success rate for indirect prompt injection (Gray Swan IPI) of 0.7%, against 1.0% for Opus 5.5 and 8.5% for GPT-6 Astra; here, lower is better. On CWE-bench, the vulnerability-finding test, Argon ties Astra at 68%, one point above Opus 5.5.

The differences that matter beyond the ranking

The first difference is answer length. One million output tokens, more than 15 times the 64,000 of earlier Gemini models, means the model can write a whole code migration or a long legal file in a single pass, without an external agent splitting it into chunks. Google calls it the largest limit in the industry; every split adds cost and a risk of errors at the seams.

The second is the release model. Google is releasing the model to cyber defenders first and taking part in the US government's voluntary pre-release access process, according to the company. Fairwind partners and internal security teams receive Argon without the usual cyber guardrails, so they can use its full capability for defensive work, according to TechWire Asia. The first partner named publicly, the security firm Wiz, used Argon to find a critical vulnerability in software used by hospitals worldwide, one that earlier frontier models had missed, according to The Next Web.

The third is documentation. As of the announcement, Google had not published a model card, a capability-threshold assessment or a misuse evaluation — only a list of four protection areas: cyber and CBRN misuse, prompt injection, alignment monitoring through the chain of reasoning, and isolated environments for risky tasks.

The actors

The mechanism: from the June promise to the September announcement

The chain has four dated links. First: Sundar Pichai had announced Gemini 3.5 Pro for June 2026; according to Reuters, Google no longer plans to release it. Second, on 5 August: Demis Hassabis gives up the role of DeepMind chief executive, becomes the lab's chairman and Alphabet's chief scientist, and Koray Kavukcuoglu takes over day-to-day leadership, according to Fortune.

The third link is September's competition: OpenAI launches GPT-6 Astra on 3 September and Anthropic, on 22 September, Claude Opus 5.5, with its price cut by 20% from Opus 5. The fourth, on 30 September: Google answers with a model larger than the Pro line, at a launch price half that of Opus 5.5.

What ties the links is the starting position. A lab that has fallen behind cannot win customers merely by matching rivals on tests, because switching providers is costly; it needs an extra reason. The launch price and the one-million-token output limit are that reason. Alphabet shares (class A) closed the day of the announcement up 3.23%, according to TradingKey, which attributes the move to the whole AI ecosystem and the cloud without naming Argon; from a single day, the announcement's effect cannot be isolated.

The counter-hypothesis: the ranking and the price belong to the seller

Three objections must sit next to the figures above. First: the tests are chosen by Google, and some are run by the company. Only a few come from external evaluators — Vals, Harvey —, and on Vals' general index the lead is small, 68.9% against 67.0% for Opus 5.5. Moreover, the rivals' figures in Google's table do not always match their own: on AutomationBench Google gives Opus 5.5 42.5%, while Anthropic's launch page gave 40.0%.

Second: a price nobody can pay is not yet a price. As long as the model sits behind the Fairwind programme, the cost comparison is theoretical, and the length of the launch price is not announced. Third: a larger model costs more to run, so a launch discount may be a temporary subsidy rather than a new market level.

Weighed, the objections do not overturn the thesis, but they narrow it. The lead on office work and law is too large to be merely a choice of tests — 19.6% against 5.4% on Harvey —, while the lead on programming is small and inconsistent. The prudent conclusion: Argon leads on long tasks, not on everything, and the price will matter only at the broad launch.

What it means for Romania and for Europe

For a software firm in Romania, Argon changes nothing today, because it cannot be bought. When it can, the arithmetic is simple: at the launch price, a team spending $1,000 a month on Opus 5.5 would pay about $500 for the same volume of tokens on Argon, i.e. roughly 2,324 lei instead of 4,647 lei, at the NBR rate of 4.6474.

Google has not said whether the broad launch will include the European Union from day one. AI Act obligations for general-purpose models have applied since 2 August 2025, and a frontier model launched without a public model card would enter the European market with its documentation still to be completed. The European timetable remains, for now, an open question.

What comes next: verifiable predictions

The signal to watch is the date Argon appears on the API pricing page, together with the term of the launch price: only then does this article's cost comparison become a market comparison.

Technical annex: what the tests measure

DeepSWE and FrontierSWE measure the resolution of real, long-running software engineering tasks in real code repositories. Terminal-Bench measures an agent's work on the command line. Vals Index is a composite index from the firm Vals, which rates models on office tasks. Harvey Legal Agent measures complete legal tasks, from research to drafting. LVBench measures the understanding of long video recordings. CWE-bench measures finding vulnerabilities classified in the CWE catalogue. Gray Swan IPI measures how often an attack through instructions hidden in content read by the agent succeeds.

A token is, on average, about three quarters of a word in English and less in Romanian, where words split into more pieces. Prices are list prices, without batch or contract discounts.

Limits of the analysis

All scores come from Google or are taken from Google's table; no independent evaluator has yet run the model on a full set of tests. Methodological differences between labs mean that one or two percentage points are not a reliable difference; the table here is a selection from the 18 tests Google published. Google has not published the model's size, the input context window, the post-launch cache price or the length of the discount. Rival prices are list prices as of 1 October 2026.