SCIENCE & TECH
Google has thrown down the gauntlet in the AI arms race. Its newest frontier model, Gemini 4 Argon, is now rolling out to developers, businesses and paying subscribers, and the company is pitching it as a model built for the hardest jobs: multi-step coding, finance, legal work, cybersecurity and science.
The sticker price is the real jaw-dropper. During an introductory period, Google is charging $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95 percent. But the party ends on 30 November 2026, when the public-access rates double to $4 and $20.
And Google is not alone in slashing prices. According to a report from Yahoo Finance, OpenAI launched GPT-6.1 Sol a day earlier at exactly the same $2 and $10 rates, after cutting prices from its previous GPT-6 Astra. The big AI labs are now in a genuine price war, and customers are the winners, at least for now.
What Google Says Argon Can Do
Koray Kavukcuoglu, Google DeepMind's senior vice president and chief AI architect, announced Argon on 13 October 2026. Google says it is available globally through the Gemini API in Google AI Studio, to Google AI Ultra subscribers inside its Antigravity tool, and to Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search.
Google's headline claims are packed with numbers. It says Argon ranks number one on the Vals Index, a benchmark measuring performance across economic sectors such as finance, coding, legal and tax work. It reports 87 percent on ERQA, a test of embodied reasoning, which Google calls the highest among frontier models. In private evaluations by Gamow Labs, it scored 57 percent on RareBench, which tests help with rare-disease diagnosis, and 52.2 percent on LabBench, which tests wet-lab reasoning.
Developers will be watching the output limit closely. Through the Gemini API, Google offers a unified input-output limit of one million tokens plus a new continuation protocol for very long responses. Reports differ on the exact input window; one analysis said Google had not stated a figure, while another put it at two million tokens. What is clear is that the output limit has leapt from the 64,000 tokens of earlier Gemini models to around one million, which matters if you want an AI to write an entire software project in one go.
The Benchmark Showdown
Coders will care about DeepSWE v1.1, a software-engineering benchmark. Google's own table, as summarised by Yahoo Finance, has Argon at 77.9 percent, ahead of Anthropic's Claude Opus 5.5 at 74.2 percent and OpenAI's GPT-6 Astra at 74.1 percent. On CWE-bench v1, which measures vulnerability remediation, Argon ties for first place at 68 percent.
Google also reports practical wins from its own engineers: a more than 35 percent increase in agentic code submissions and a nearly one-third drop in code rollbacks. Partners and researchers cite a 2.3-times speed-up in the xTB quantum chemistry package, a 2.85-times inference speed-up for Gemma 4 E2B on vLLM, and nearly doubled throughput for an internal quantum error decoder.
But hold your applause. Nearly all of these figures come from Google itself, and Yahoo Finance's analysis notes they are vendor-reported and unverified. A separate review by security firm NeuralTrust, which sells AI-security products of its own and so has an interest in the topic, painted a mixed picture. It found Argon trailing Opus 5.5 on Terminal-Bench 4.0, 57.4 percent versus 66.4 percent. Independent scorer Artificial Analysis gave Argon 52.6 against 57.6 for Opus 5.5, while Vals still ranked it first of 41 models.
In plain English: Argon is excellent, and it is clearly in the top tier, but whether it is the best depends on which test you pick and who is running it.
Safety: The Fine Print Everyone Wants to Read
As models grow more capable, so do worries about misuse. Google says Argon runs behind a multi-layer defence under its Frontier Safety Framework, including jailbreak resistance, monitoring and enforcement. External red-teamers testing for biological risks could unlock less than 1 percent of the sub-tasks tested, and policy violations on an adversarial test set fell nearly twofold compared with earlier Gemini models.
Google also says Argon was trained to respect system boundaries and to pause before taking irreversible actions, and that attempts to take shortcuts on coding benchmarks dropped to 2 percent. An automated misalignment monitor watches the agent's reasoning and actions in real time. On the Gray Swan indirect prompt-injection benchmark, Google says attack success fell to 0.7 percent across 15 adaptive attempts.
Prompt injection, where hidden instructions in a web page or email trick an AI agent into misbehaving, is one of the hottest security problems in the industry. NeuralTrust cautioned that Google had not published every score behind its claims and urged enterprises to run their own misuse tests rather than take vendors at their word. That is sound advice for any company handing an AI agent the keys to its systems.
Another eyebrow-raiser: through Google's Fairwind Program, approved cyber defenders can use Argon without the usual cyber guardrails. Google is expanding that programme to hundreds more defence partners. The logic is that good guys need the sharpest tools to find holes before criminals do, but it is a reminder that these systems cut both ways.
Science Gets a Sweetener
Google is also courting researchers. Up to 10,000 scientists can apply for free access to Antigravity's highest usage tier for six months, subject to eligibility and capacity. Separately, Google announced that up to 200 European researchers will receive access to advanced models and 3 million dollars in computing credits, supporting the European Union's RAISE programme, according to the TechStartups daily roundup.
The Wider Battlefield
The Argon launch landed on a busy news day. The same roundup noted Nvidia pledging 250 million dollars over five years in donated computing access for AI safety and robotics research in the UK. At a UK parliamentary hearing on 13 October, OpenAI and Anthropic reportedly backed binding AI safety requirements, while Google DeepMind and Meta were more cautious. Meanwhile, Yahoo Finance reports that open-weight Chinese models now account for about 61 percent of top-model token traffic on OpenRouter, at an average of 83 cents per million tokens, which it says is roughly 7.3 times cheaper than proprietary rivals.
That last number is the one that should keep Silicon Valley executives awake. A launch price of $2 per million input tokens looks cheap, until you notice that some rivals are charging a fraction of it. The price war is not just Google versus OpenAI; it is the whole closed-model business against a wave of cheap open alternatives.
Expert Perspective
Read the coverage carefully and a consistent theme emerges: the model is strong, the price is aggressive, and the evidence is still being gathered. Yahoo Finance's analysis called the pricing move the real strategic story, while stressing that the benchmark lead is vendor-reported. NeuralTrust, meanwhile, called Argon a credible model with a launch-price advantage but noted that independent results are mixed and that some details were unclear at launch, including how long the introductory rate would last and how reasoning tokens are billed. Google's own post, by contrast, lists the November 30 date explicitly, a reminder that details are still shifting a few days into the rollout.
There are even conflicting reports on timing. Google's blog says the announcement came on 13 October, while NeuralTrust's article refers to a 30 September announcement limited to the Fairwind cyber defenders, with wider access to follow. The likely explanation is a staged rollout, but companies would be wise to check what their own accounts can actually access today.
What This Means for You
If you are a developer or run a business, the next six weeks are a window of unusually cheap access to a top-tier model. If you are planning a big project, testing Argon before 30 November, when prices double, could save real money. Just do your own evaluations on your own tasks rather than trusting a leaderboard, and build in security checks if the model will act on your behalf.
If you are an everyday user, you may notice smarter answers in the Gemini app and AI Mode in Search if you pay for Pro or Ultra. Expect more AI agents that book, code, research and file things for you, and expect to see competitors answer with their own upgrades, because this price war is only just beginning.
And if you worry about where this is heading, the safety details are worth following. The fact that rival labs are publicly arguing over binding rules in Parliament shows that the guardrail debate is moving from conference panels into law.
The Bottom Line
Gemini 4 Argon is Google's loudest shot yet in the battle for the AI crown, and its low launch price turns up the heat on OpenAI, Anthropic and a swarm of cheap open-source rivals. Whether it truly wears the crown will be settled not by slides but by independent tests and the verdict of the developers who use it every day. The discount clock is ticking, the benchmarks are being re-run, and the AI world is watching.
