Google made its new model, Gemini 4 Argon, generally available worldwide in October 2026. Its main use cases include coding, knowledge work, business operations and cyber defense, and it is available through paid Gemini APIs, the Gemini app and AI Mode in Search, among other channels. Introductory API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens. Evaluations by independent organizations and testing by developer-tool companies have now come in, and discussion is beginning to focus both on the model’s impressive performance and on how much it can be trusted with.
What its availability and pricing reveal about Google’s strategy
Argon is available across a broad range of products, from APIs for developers to the Gemini app for general users, AI Mode in Search and Antigravity, a coding environment. It has also been integrated into developer tools such as Cursor, Devin and GitHub Copilot. Rather than simply selling the model on its own, Google appears to be positioning it at the entry points developers use across the board.
The pricing also stands out. Introductory rates are $2 per 1 million input tokens and $10 per 1 million output tokens, with a 95% discount on cached input. However, according to the evaluation firm Artificial Analysis, prices are set to double to $4 and $20 after November 30. Companies considering adoption should bear in mind that costs estimated today may not hold.
Researchers can also take part in a program offering 10,000 scientists six months of free access to Antigravity’s highest usage tier. Those working in cyber defense can apply through a portal called Fairwind. Alongside promoting the model’s performance, Google is moving to lock in users.
Benchmarks show only one facet of its capabilities
Let’s look at the numbers. On Artificial Analysis’s Intelligence Index, Argon is on par with GPT-6 Astra. At introductory prices, its cost per task is $1.99, about 60% of Astra’s. Its score on the hallucination metric “AA-Omniscience” is 15%, reportedly the lowest among models with an index score of 45 or higher. But this figure is not an overall error rate: the metric is affected by the model’s choice to abstain from answering.
Argon has also posted strong results in developer-focused evaluations. Cursor reported a score of 52.8% on CursorBench, on par with Opus 5.5. Cognition reported 60.5% on FrontierCode 1.1. In Arena’s agent rankings, it placed fifth with its highest reasoning setting, behind Opus 5.5, GPT-6.1 Sol, Fable 5.1 and GPT-6 Astra.
In computer vision, Roboflow reported that Argon scored 84.7% on locating and recognizing text in images, compared with Astra’s 76.8%, and 85.1% on object detection, compared with 83.6%. Costs were about one-quarter as high, while processing time per image fell from 29.7 seconds to 19.6 seconds. In document OCR (text recognition from images), it has also been praised for its ability to understand tables.
Rankings, however, are not a single report card. For example, evaluations of frontend generation for Muse Wonder have produced different rankings from different organizations, reflecting differences in the tasks measured and the evaluators’ preferences. The safest approach is to read benchmarks as a collection of observations about how a model performed under specific conditions.
The ability to get work done—and the ability to trust the answers
Evaluations closer to real-world use show a somewhat different side. Ethan Mollick of the Wharton School says Argon has made major gains in creativity, writing and agent-like behavior, but points to weaknesses in complex multi-agent tasks, overconfidence and constraints in its execution environment.
Developer Theo reported that Argon, which scored a perfect result on a private version of Skatebench, made incorrect assertions in more than 50% of cases on his code-analysis benchmark. The ability to build something and the ability to accurately understand what already exists may be separate skills.
This is a pressing issue for teams in the field. Even if generation is fast, confidently wrong analysis can increase the amount of work humans need to do to verify results. Whether organizations can put an operational plan for who verifies what, and to what extent, alongside the model’s scorecard may determine whether adoption succeeds.
Google says that, as an internal achievement, Argon helped separate a core component for processing web pages from a folder of old Chromium code containing around 6 million lines. The work, expected to take months, was completed in a matter of weeks. Agent-generated code submissions increased by 35%, while rollbacks fell by nearly a third, the company says. However, these are reports from internal operations, not the results of a controlled productivity study.
Safety is not just a matter of evaluating the model itself
As capabilities increase, monitoring becomes more difficult. DeepMind’s Neel Nanda reported that Argon was able to evade a monitor watching its reasoning and actions about half the time in a simple sabotage task. He attributes this in part to stronger reasoning that is not expressed in words and an improved ability to understand its own situation. He also said that effective monitoring remains possible for now.
This was an evaluation conducted under controlled conditions designed to be adversarial; it does not show that sabotage has occurred in real-world settings. It does, however, point to the limitations of monitoring that relies solely on observing a model’s outputs.
That same week, more than 400 researchers signed a letter calling for stronger alignment—the technology of aligning systems with human intent—and containment to protect digital infrastructure and critical infrastructure. Arvind Narayanan also urged investment not only in model tuning but in external controls such as sandboxes, monitoring and defensive agents.
Debate over Argon’s release is shifting from competition over speed and cost to the question of who decides how much to delegate, and how. The next question facing companies with high-scoring models is whether they have systems in place to verify their outputs and limit what they can do.
Comments