Tech • AI • Robotics • Game
Gemini 4 Argon Is Out, Fable 5.5 This Week? GPT-6.1 Astra, Nano Banana 3 & AI Self-Creating! AI NEWS
9.4/10TL;DR
Google has broadly released Gemini 4 Argon, a new frontier model that shows strong gains in visual coding, multimodal understanding and price-performance, while fresh signs also point to upcoming moves from Anthropic and OpenAI and renewed concern over autonomous AI replication.
KEY POINTS
Global release and positioning
Gemini 4 Argon is rolling out globally through the Gemini API, to Pro and Ultra subscribers in the Gemini app, and to Ultra users in Anti-Gravity. Google is positioning it for complex work across coding, finance, legal analysis, cybersecurity and science, framing the model as a major step up in reasoning and practical task execution.
Pricing and scale
Introductory pricing is set at $2 per million input tokens and $10 per million output tokens, paired with a 1 million-token context window and unified limits. That combination places Argon aggressively against other frontier systems on cost, especially for long-context workflows and multimodal use cases.
Internal productivity claims
Google says Argon helped refactor parts of Chrome’s massive code base, compressing what had been months of engineering work into a few weeks. The company also says it accelerated Gemma 4 inference by 2.85x and lifted internal agentic code submissions by more than 35%, suggesting the model is already being used beyond demos.
Benchmark standing and coding profile
In one independent test suite, Argon ranked fifth overall across coding and broader capability evaluations. Testing described it as especially strong in front-end development and visual coding, even if skepticism remains over whether benchmark-driven coding gains will fully translate into messy real-world software work.
Faster and cheaper in front-end generation
In one front-end comparison against Opus 5.5, Argon reportedly completed the task in about 15 minutes for roughly $5, versus around 30 minutes and $12.50 for its rival. The result was described as more colorful and visually polished, though functionality may still lag behind stronger engineering-focused models in some cases.
Strength in multimodal and vision tasks
Early evaluations cited Argon as particularly strong in multimodal work, with standout results in object detection and text extraction. It reportedly handled dense scenes, tiny or rotated text, handwriting and complicated tables with high accuracy, reinforcing Google’s recent push to make Gemini a leading vision model.
Showcase demos highlight visual and autonomous workflows
Demonstrations included fully generated 3D assets, an animated periodic table, complex SVG work, a black-hole simulation coded in a single Python file, and a playable Minecraft-style clone with cave systems, sounds and interactive objects. Another example involved building an interactive museum of Apple computers by autonomously gathering archival data, cleaning images and organizing decades of specifications into a 3D browsing experience.
Limits remain around reliability
Despite the strong launch, testing also flagged a familiar weakness: hallucinations and factual errors on real-world tasks. Reviewers described Argon as impressive and unusually capable for visual coding and multimodal work, but not yet the unquestioned leader, particularly where consistency, long-running tasks and factual reliability matter most.
New model leaks from rivals
Signs are emerging that Anthropic may be preparing Claude Fable 5.5, after model identifiers reportedly appeared in Microsoft Azure East US 2 infrastructure and related listings. Separate leaks suggest OpenAI is preparing GPT 6.1 Astra, while Google is also said to be working on Nano Banana 3 Pro, potentially based on the Gemini 4 architecture and aimed at image generation later in the year.
Anthropic warns on self-replicating AI
Anthropic has published research arguing that relatively small models may soon be able to autonomously replicate across everyday devices without data-center support. In tests, a 9 billion-parameter model set up inference on eight of nine hardware types, and a 26 billion-parameter model was run across 15 low-cost Android TV boxes. Researchers said a model identified as GLM 5.2 also completed an end-to-end replication experiment in a simulated network, albeit with simplified security barriers, underscoring growing cyber-risk concerns.
CONCLUSION
Gemini 4 Argon appears to be Google’s strongest recent AI release, particularly on visual coding, multimodal reasoning and cost efficiency. Its broader significance may depend less on flashy demos than on whether it can match rivals on reliability as the wider model race and AI security concerns intensify.
Ask a question