Tech • AI • Robotics • Game
Did Google just dethrone OpenAI and Anthropic?
9.4/10TL;DR
Google has launched Gemini 4 Argon, a new multimodal flagship AI model that posts top benchmark scores and aggressive pricing, but early hands-on testing suggests its coding and media-generation performance remains uneven against rivals from OpenAI and Anthropic.
KEY POINTS
A new flagship built from scratch
Gemini 4 Argon is presented as a fresh model rather than an update to earlier Gemini releases. It uses a sparse mixture-of-experts transformer architecture and was trained on Google TPUs, with a knowledge cutoff of March 2026, implying roughly six months of training before launch. The rollout began in late September through a restricted Fairwind access program before broader availability.
Benchmarks show broad strength
Google says Argon leads on the Deep Software Engineering Benchmark, aimed at long-horizon coding tasks, and tops the VAL index, which weighs areas such as finance, legal work and software by their share of the US economy. It also posts 91.7% on LVBench for long-video understanding and is positioned as a leading model for legal analysis.
Not every score points to dominance
Some published results show rivals still ahead in important areas. On Terminal Bench 4.0, Claude Opus 5.5 beats Argon by 66 to 57, and on Frontier S, GPT-6 Astra leads by more than 10 points. On Arena AI, Argon ranks first in subjective text comparisons, but falls to ninth on a web-development coding arena, raising questions about whether its practical coding ability matches its broader benchmark profile.
Lower prices are part of the strategy
Through November 30, Argon is priced at $2 per million input tokens and $10 per million output tokens, undercutting many premium frontier-model rates. That makes the launch as much a pricing move as a technical one, especially as competition intensifies among top AI providers.
A messy product rollout frustrated users
The launch also drew criticism over access. Google replaced an earlier Gemini 3.1 offer with Argon, but initially limited it to the AI Ultra subscription priced at $100 per month. Broader users gained access only days later and with lower usage quotas, leaving developers complaining that the release was unnecessarily confusing and restrictive.
Multimodal output remains a differentiator
Argon is marketed as an omni model, able to accept and generate multiple media types including text, images, audio and video. That gives Google a capability edge in raw video output over some competitors, though practical quality appears more mixed than the feature list suggests.
Early video generation looked underwhelming
In one test, Argon was asked to create a promotional video for an app. The result was visually polished but still recognizably AI-generated and not convincing enough for high-end commercial use. By contrast, a competing workflow using Claude to generate motion graphics through code produced an output that looked better and was easier to inspect and customize.
3D coding tests exposed weaknesses
A side-by-side Three.js prompt to build a realistic interactive 3D horse highlighted the model’s limitations in coding-heavy creative tasks. Outputs from Claude and GPT-6 Astra were described as flawed but salvageable with follow-up prompts. Argon’s result delivered a cleaner interface but a lower-fidelity animal model that did not meet the request for realism.
Safety and alignment questions emerged
The same 3D test also surfaced an unwanted anatomical detail in Argon’s horse model, despite guardrails being active for general users. That result suggests Google’s effort to reduce refusals may still leave gaps in visual-output alignment, especially when prompts involve realistic biological forms.
CONCLUSION
Gemini 4 Argon gives Google a stronger position in the frontier AI race with high benchmark scores, multimodal breadth and aggressive pricing. But early use suggests that standout metrics do not yet guarantee better real-world coding or generation performance than the top models from OpenAI and Anthropic.
Ask a question