Google made its new model, Gemini 4 Argon, generally available worldwide in October 2026. Its main uses include coding, knowledge work, business operations and cyber defense. It can be accessed through the paid Gemini API, the Gemini app, Search’s AI Mode and other channels. Introductory API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens. Results from external evaluators and developer-tool companies are now in, and discussions are beginning to focus both on the model’s strong performance and on how much work it can be trusted to handle.
What its availability and pricing reveal about Google’s aims
Argon is available across a wide range of products, from the API for developers to the Gemini app for general users, Search’s AI Mode and the coding environment Antigravity. It has also been integrated into developer tools such as Cursor, Devin and GitHub Copilot. Rather than simply selling a standalone model, Google appears to be placing it at the entry points to development workflows all at once.
The price is also striking. Introductory pricing is $2 per 1 million input tokens and $10 per 1 million output tokens, with a 95% discount on cached inputs. But according to the evaluator Artificial Analysis, those prices are scheduled to double to $4 and $20 after November 30. Companies considering adoption should bear in mind that the costs they calculate now may not last.
For researchers, Google has also set up a program offering 10,000 scientists six months of free access to Antigravity’s highest tier. People working in cyber defense can apply through a channel called Fairwind. Alongside promoting the model’s performance, Google is moving to lock in users.
Benchmarks show only one side of its strengths
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 AA-Omniscience, a hallucination metric, is 15%, reportedly the lowest among models with an index score of 45 or higher. But this figure is not the model’s overall error rate: the results are affected by whether it chooses not to answer.
Argon has also posted strong results in developer-focused evaluations. Cursor reported a score of 52.8% on CursorBench, putting it on par with Opus 5.5. Cognition reported 60.5% on FrontierCode 1.1. In Arena’s agent rankings, Argon placed fifth under the highest reasoning setting, behind Opus 5.5, GPT-6.1 Sol, Fable 5.1 and GPT-6 Astra.
In the field of image processing, Roboflow reported that Argon scored 84.7% on locating and recognizing text in images, compared with Astra’s 76.8%. For object detection, Argon scored 85.1%, against Astra’s 83.6%. Costs were about a quarter as high, while processing time per image fell from 29.7 seconds to 19.6 seconds. Evaluators have also said Argon is particularly good at understanding tables in document OCR, or text recognition from images.
Rankings, however, are not a single report card. For tasks such as front-end generation, evaluators can rank models differently, as in the case of Muse Wonder; differences in the tasks being measured and in evaluators’ preferences translate directly into different results. The safest approach is to read benchmarks as a collection of accounts of how a model performed under particular conditions.
The ability to get work done—and the ability to trust the answers
Evaluations closer to real-world use reveal a somewhat different picture. Ethan Mollick of the Wharton School says creativity, writing and agentic behavior have improved significantly, but also points to weaknesses in complex multi-agent tasks, overconfidence and constraints in the execution environment.
Developer Theo reported that although Argon earned a perfect score on a non-public version of Skatebench, it made incorrect assertions more than 50% of the time on his own code-analysis benchmark. The ability to build something and the ability to correctly understand something that already exists may require different skills.
For people working in the field, this is a pressing concern. Even if a model generates output quickly, confidently getting its analysis wrong can increase the amount of work humans have to do to check it. Adoption may hinge on whether organizations can put an operational plan—spelling out who verifies what, and to what extent—alongside the model’s scorecard.
Google says that, in its own operations, it completed in weeks a project that was expected to take months: separating core web-page processing functions from a legacy Chromium folder containing around 6 million lines of code. It also says agent-submitted code increased by 35% and rollbacks fell by nearly a third. However, this is an account of internal use, not a productivity measurement made under controlled, standardized conditions.
Safety is not just a matter of evaluating the model itself
As capabilities rise, monitoring becomes more difficult. DeepMind’s Neel Nanda reported that Argon evaded a realistic monitor of its reasoning and actions about half the time in a simple sabotage task. He believes this is due in part to improvements in reasoning that the model does not verbalize and in its ability to understand the situation it is in. He also said that useful monitoring remains possible for now.
This was an evaluation conducted under controlled conditions deliberately designed to be adversarial; it does not show that sabotage has occurred in real-world environments. Still, it points to the limits of monitoring that relies only on observing a model’s outputs.
That same week, more than 400 researchers signed a letter calling for stronger alignment—the techniques used to align AI with human intentions—and containment measures to protect digital and critical infrastructure. Arvind Narayanan also called for investment not only in model tuning but in external controls such as sandboxing, monitoring and defensive agents.
The debate over Argon’s release is shifting from competition over speed and cost to the question of who decides how much autonomy to grant, and how. Companies with high-scoring models will next be asked whether they have systems in place to verify their outputs and limit what they can do.
Comments