Google 发布 Gemini 4 Argon 并面向全球开放

Google is rolling out Gemini 4 Argon globally, making its new flagship AI model available to developers, businesses, and consumers starting today.

Gemini 4 Argon is accessible through the API in AI Studio and the Gemini platform, according to Google. Google AI Ultra subscribers can use the model in Google Antigravity, while AI Ultra and Pro subscribers get access through the Gemini app and Google Search's AI mode.

Not everyone benefits equally, though. Free Gemini users and paying Plus subscribers now only have access to the weaker Flash models under a new model tier structure. Argon replaces the Gemini Pro class as Google's new frontier series. Google says more Gemini 4 models will follow, which likely means updated Flash variants that would eventually trickle down to free and Plus users as well. Internally, Google is reportedly already testing a potential successor codenamed "Carbon," which is supposed to outperform Argon at coding in particular.

Pricing and benchmark results remain as previously reported: $2 per million input tokens and $10 per million output tokens at the introductory rate, with a 95 percent discount on cached inputs. Argon holds the top spot on the Vals Index, but it still trails Anthropic's Claude Opus 5.5 in independent overall rankings. And because Google waited roughly two weeks to ship, it now also sits behind Meta's Muse Wonder Full, which launched just yesterday.

*Google doesn't state this figure explicitly but says the cache is 95 percent cheaper than the regular input token price.

Google says Argon turned months of internal engineering into weeks

Google shared some internal numbers with the launch, though they're self-reported. Software engineers using Argon increased agent-driven code submissions by more than 35 percent, while code rollbacks dropped by nearly a third, the company says.

In the Chromium project, Argon helped the Chrome team complete an architectural refactor across hundreds of files within the 40-million-line codebase. Work that would have taken months was finished in a few weeks. For users, the payoff is that web pages can move between tabs, sidebars, and floating windows without needing to reload.

Working with Google's open-source Agent Development Kit, Argon ran autonomously overnight and found optimizations that cut session preparation time by 40 percent and sped up cold imports by 21 percent. Inference speed for the open Gemma models on an A100 via vLLM jumped by a factor of 2.85, from roughly 37,900 to over 108,000 tokens per second. The changes are visible in upstream pull requests.

Argon can also generate complex 3D models, games, and animations in pure code that run directly in the browser. One demo features a raytracing engine written entirely in Python with no external libraries. Google published the source code.

In the demo below, Argon starts from a single image in Antigravity and controls Blender via MCP in an autonomous loop. It modeled gothic architecture and set up audio-reactive lighting to produce a finished 3D scene, Google says.

3D remains a weak spot for Meta's Muse models, which according to Meta AI chief Alexandr Wang were not specifically trained for 3D demo games.

Google offers 10,000 researchers free access and claims its strongest safety record yet

Google is also offering 10,000 researchers the highest usage tier of Google Antigravity free for six months. Spots are awarded on a rolling basis depending on eligibility, capacity, and selection. Applications are open now.

Cybersecurity defenders can apply separately for the Fairwind program, which grants access to Argon without cyber safeguards. Google plans to expand the program to hundreds of additional partners.

On the safety side, Google calls Argon's architecture its strongest to date. External red teaming in the bio domain bypassed fewer than 1 percent of tested safety checks, and on an adversarial set, violations dropped by nearly half compared to earlier Gemini models.

On agent alignment, attempts to take unauthorized shortcuts fell to 2 percent. An automated misalignment monitor analyzes the agent's reasoning and actions in real time.

Argon also scored the lowest attack rate on the Gray Swan Indirect Prompt Injection benchmark, with a successful attack rate of just 0.7 percent across 15 adaptive attempts. In the previous Gray Swan test, GPT-6 Astra hit 8.5 percent and Claude Opus 5 came in at 4.8 percent. Google's write-up doesn't make clear whether the exact same tests were used, though. More safety details are available in the model card.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.