Google launches Gemini 4 Argon with limited-time half-price API

Gemini 4 Argon is gradually rolling out to global developers, enterprises, and paid subscribers. Google's positioning for this generation is clear: strengthen coding, professional knowledge work, and sustained execution of complex tasks, while pushing visual understanding, spatial reasoning, and tool use further toward the frontier. The official technical blog sums up the core upgrade in one point—stronger long-horizon task execution, meaning the model can understand a goal and then look up information, call tools, modify code, and check results on its own.

Two sets of numbers support the "return to the big three" narrative. In Google's own evaluation, Argon covers areas including office work, agentic coding, machine learning, and scientific reasoning, with overall performance reaching the level of GPT-6 Astra and Opus5.5, and even leading on many items. On the third-party Artificial Analysis Intelligence Index, Argon High ties with GPT-6 Astra Max and Claude Fable5.1 Max at 53 points.

These two sets of numbers are not the same in nature. The former is Google's self-testing, a source_claim that has not been independently verified by third parties; the latter is a third-party index, but the article only gives the tied 53 points and does not provide full evaluation details. In other words, "catching up" is currently more of an official line, while "tied" is the coordinate that can be externally cited.

As eye-catching as the tied 53 points is the price. Launch API pricing is $2 per million input tokens and $10 per million output tokens, with the discount running through November 30; from December 1, prices revert to $4 input and $20 output. In other words, this is a 50% off promotion with a clear deadline, not long-term pricing.

At equivalent performance levels, Argon High's average cost per task is $1.99, about 39% lower than Astra. Based on this, the article roughly calculates that Argon's invocation cost is about half of Opus5.5 and one-fifth of Astra. It should be noted that this "half, one-fifth" comparison is the article's own calculation, an inference, not an officially published pricing conclusion.

Whether the cost advantage translates into real savings depends on whether tasks are done right the first time. The article itself raises this counterpoint: if the model is not smart enough and burns tokens going back and forth on a task, the money saved may all be burned right back. This is exactly what leads the discussion to the actual quality of long-horizon execution.

Google chose to validate coding ability using its own open-source agent framework, ADK. This Python project has about 205,000 lines of code, and the team's task for Argon was: without changing interfaces or existing behavior, reduce the framework's runtime overhead and prove through tests that the optimizations are effective.

The result: Argon worked unattended for 11 consecutive hours, proposed 6 optimization plans touching 32 files, and ran all 16,892 open-source tests. After engineer review, two optimizations were merged, including reducing session history copy overhead and making large modules load on demand. Based on these changes, memory preparation overhead in long ADK sessions dropped by about 40%, and package import time dropped by 21%.

This set of numbers is worth breaking down. Eleven hours unattended, 32 files, 16,892 tests show the model has the ability to keep pushing complex engineering tasks forward; but only 2 of 6 plans being merged also shows there is still a clear discount between "proposing a plan" and "a plan that can be adopted." This is an official self-reported case, a source_claim, and lacks independent reproduction.

In larger-scale engineering tasks, Google says Argon helped complete an architectural refactor in the Chromium browser project involving hundreds of files, compressing work originally expected to take months into weeks. When optimizing the Gemma4E2B inference program, Argon raised throughput under a specific configuration from about 37,900 tokens per second to over 108,000 tokens per second, reaching 2.85 times the original end-to-end speed. Google also says that in recent weeks, Argon in Antigravity has helped the team increase agent code commits by more than 35% and reduce code rollbacks by nearly one-third.

Multimodality and robotics are the parts where the official showcase invests the most. In the official demonstration, after watching a video of a human playing a beanbag toss game, Argon analyzed spatial relationships, calculated full-body joint movements, autonomously wrote simulation code, and transferred the plan to Apptronik's Apollo humanoid robot. Google says this "video-simulation-reality" pipeline can compress robot simulation modeling work that previously took weeks of manual effort into a few hours.

In scientific research scenarios, Google scientists had Argon participate in designing serine esterases; the model completed data organization, model tuning, and candidate screening within hours and proposed 5 new designs, and Google disclosed that all five designs showed strong catalytic activity. Google also announced that it will provide six months of Antigravity's highest usage quota free of charge to 10,000 eligible and selected scientists, opening in batches.

The common point of these cases is that they are all provided by Google and lack independent reproduction. They can prove "what Google believes the model can do," but they cannot be directly equated with "what developers can reliably get in their own projects."

Safety capabilities were singled out this time. In CWE-bench v1, which evaluates vulnerability-fixing ability, Argon scored 68%, sharing the highest score in the official comparison table with GPT-6 Astra; in the Gray Swan indirect prompt injection evaluation, after 15 adaptive attack attempts, Argon's attack success rate was 0.7%.

The article mentions that Google's release page includes a 217-page safety report, and says the assessment by the former Anthropic and now DeepMind-related lead is that this is Google's most detailed to date. As for whether this report is related to the flagship release cadence, the article explicitly states there is no evidence. This is an uncertainty that should be preserved and should not be interpreted as a causal conclusion.

Beyond the official narrative, at least three pieces of counterevidence need to be presented side by side. First, Argon still trails several major competitors in the table on FrontierSWE and Terminal-bench, and whether real project gains can be reliably reproduced in more people's work remains to be seen. Second, some users say their own results are worse than the official showcase, and large-scale usage effects are still unclear. Third, Ethan Mollick, who tried it early, says this generation is a clear leap compared with previous Gemini, with progress in creativity, writing, and user experience, but complex multi-agent tasks still lag, and it is occasionally overconfident and a bit lazy.

In generation quality comparisons, some users also think Opus 5.5 is still stronger, and visual details such as the movement of a whale still look unnatural. The gap between this feedback and the official demo constitutes the most realistic open question for this generation's flagship: self-test data and selected cases can show the upper limit, but what users feel in daily use is the lower limit and stability.