Google Releases Gemini 4 Argon, Limited-Time Half-Price API

If you ask me, Google is a seasonal climate model company—when the weather cools down, its flagship matures.

According to Google's positioning, this generation mainly strengthens coding, professional knowledge work, and sustained execution of complex tasks, while also pushing its previously strong capabilities in visual understanding, spatial reasoning, and tool calling further to the frontier.

In the official demo, it's clear that what Gemini 4 Argon is most confident about this time is still multimodality.

Enter a text prompt simulating a kite, and the generated result simply takes off.

And on the question everyone cares about more—whether it can return to the top three and give users one more model choice—Gemini 4 Argon performed quite admirably in the evaluations this time.

In Google's own evaluations, Gemini 4 Argon covered multiple areas including office work, agentic coding, machine learning, and scientific reasoning, with overall performance already reaching the level of GPT-6 Astra and Opus 5.5, and even leading in quite a few items.

On the third-party Artificial Analysis Intelligence Index, Argon High also tied with GPT-6 Astra Max and Claude Fable 5.1 Max at 53 points.

More importantly, while reaching the same level, Argon High's average cost per task is only $1.99, about 39% lower than Astra.

This cost advantage is reflected in the launch pricing: the API costs $2 per million input tokens and $10 for output. A rough calculation shows that Argon's invocation cost is about half that of Opus 5.5 and one-fifth that of Astra.

Worth noting, the discount runs until November 30; starting December 1, input and output prices will return to $4 and $20 respectively. In other words, 50% off first—everyone is welcome to try.

Of course, if the model isn't smart enough and burns tokens back and forth on a task, the money saved might all get burned right back.

Ethan Mollick, who tried Gemini 4 Argon in advance, also gave a fairly measured assessment:

Compared with previous Gemini models, this generation is a clear leap, with progress in creativity, writing, and user experience, but there is still a gap on complex multi-agent tasks, and it occasionally is overconfident and cuts corners.

As for why the new flagship made everyone wait so long again, I looked through Google's new release page and saw a 217-page safety report~

(Google is still too afraid of something going wrong)

From the Pro version's repeated delays, to the announcement of an initial limited release on September 30 this year, to who knows how many "see you next week"—

Gemini 4 Argon is finally here.

In the official technical blog, Google summarizes Argon's core upgrade in one point: stronger long-horizon task execution capability.

Simply put, the model no longer just answers questions; after understanding the goal, it can look up information, call tools, modify code, and check results on its own, continuously advancing across multiple steps to truly finish a complex task.

However, compared with examples in enterprise and engineering scenarios, what Google was most eager to show off and PR this time was actually multimodal capability.

For example, Sundar Pichai posted a wave of demos on Twitter of "our own model generating our own phone"—

Argon directly generated 3D scenes related to the Pixel phone.

In another case, Argon even generated a complete 3D dynamic typography effect entirely through code, including the music in it.

The whole process did not use pre-recorded materials or video editing software.

When reposting the wizard chess demo, Hassabis also specifically mentioned the model's understanding of physical laws and 3D modeling.

This scene even reminded him of his own experience playing Battle Chess on the Amiga back in the day.

Of course, netizens were not idle either, and soon put Argon together with GPT, Claude, Kimi, and other models for generation comparisons.

Overall, Google's visual quality this time is indeed very striking, although in details, such as the movement of the whale, some unnaturalness can still be seen.

However, some netizens' tests also showed that Opus 5.5 remains absolutely divine.

Beyond the good-looking demos, this spatial understanding capability also extends further to real robots.

In the official demonstration, Argon watched a video of a human playing a beanbag tossing game.

It then analyzed the spatial relationships in the environment, calculated full-body joint movements, autonomously wrote simulation code, and transferred the solution to Apptronik's Apollo humanoid robot.

It's fair to say that, seeing GPT and Claude aggressively move into robotics these past two days, Google finally couldn't sit still this time.

Google says this "video—simulation—reality" pipeline can compress robot simulation modeling work that used to take weeks of manual effort into a few hours.

The significance behind this is that human demonstrations are becoming easier to convert into tasks that robots can learn, practice, and verify.

In the past, watching videos to understand actions, building simulation environments, and writing control code were often three disconnected steps.

Now, multimodal models are beginning to try to connect this chain.

On the coding side, the core improvement of Gemini 4 Argon this time is its ability to handle the long-term maintenance and optimization of large real-world projects.

In the past, it could only complete individual coding tasks; the new generation of models is better at understanding the structure of an entire codebase, locating performance bottlenecks, and making cross-file changes, running tests, and pushing a complex engineering task forward continuously.

Google first validated this capability on its own open-source agent framework, ADK.

This Python project has about 205,000 lines of code. The team gave Argon the task of reducing the framework's runtime overhead without changing interfaces or existing behavior, and proving through tests that the optimization was effective.

As a result, Argon worked continuously for 11 hours unattended, proposed six optimization plans involving 32 files, and ran all 16,892 open-source tests.

In the end, after engineer review, two of the optimizations were merged, including reducing the overhead of copying session history 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 fell by 21%.

Google has also applied similar capabilities to larger-scale engineering tasks.

In the Chromium browser project, Argon helped complete an architectural refactor involving hundreds of files, compressing work originally expected to take months into a few weeks.

In addition, while optimizing the Gemma4E2B inference program, Argon raised throughput under a specific configuration from about 37,900 tokens per second to more than 108,000 tokens per second, reaching 2.85 times the original end-to-end speed.

Google says such optimizations used to take months, but this time took only a few days.

Looking at Google as a whole, over the past few weeks Argon in Antigravity has helped teams increase agent code commits by more than 35% and reduce code rollbacks by nearly one-third.

Of course, some netizens said their own experience was not as good. The real performance will still depend on the results of large-scale use later.

This also corroborates the earlier leaderboards: Argon is strong on DeepSWE and Vibe Code Bench, but on FrontierSWE and Terminal-bench it still trails several major competitors in the table.

The benefits in real projects are already there; whether they can be reliably reproduced in more people's work remains to be seen.

In professional knowledge work, the improvement of Gemini 4 Argon this time is mainly reflected in handling multiple materials, linking information, and organizing the results into deliverables that can be used directly.

The official demo showed a very typical office scenario:

The user gives Gemini a folder full of scattered project materials and asks it to organize them into a client proposal.

Gemini then generates a presentation based on custom skills, drafts emails, sends materials, and sets up automated workflows for subsequent client requests.

In addition, the Gemini app has also added a task entry point. Users can switch from chat to task execution mode, letting the model continue working after the user leaves the computer.

At the same time, Antigravity's coding interface has also been integrated into the Gemini macOS app.

In scientific research scenarios, this continuous execution capability extends further to experimental validation.

Google scientists had Argon participate in designing serine esterases. The model completed data organization, model tuning, and candidate screening within hours, ultimately proposing five new designs.

Researchers then sent these designs to the lab to verify whether they could actually work.

The assay molecule used in the experiment produces a product that fluoresces under ultraviolet light after its ester bond is cleaved. According to the results disclosed by Google, all five designs showed strong catalytic activity.

The model's proposals have already reached the stage of experimental validation.

Google also announced that it will provide 10,000 eligible and selected scientists with six months of Antigravity's highest usage quota for free, opened in batches.

Although I joked earlier that Google was afraid of something going wrong, this time its safety capabilities really deserve a separate mention.

In CWE-bench v1, which evaluates vulnerability-fixing ability, Argon scored 68%, tied for the highest score in the official comparison table with GPT-6 Astra.

The former Anthropic and current DeepMind Mechanistic Interpretability lead said that Argon's safety report is Google's most detailed to date.

217 pages—how could it not be detailed?

In the Gray Swan indirect prompt injection evaluation, after 15 adaptive attack attempts, Argon's attack success rate was 0.7%.