Google DeepMind has opened general access to Gemini 4 Argon, pushing its frontier AI lineup into unusually long territory. The sparse mixture-of-experts model is now available through the paid Gemini API for developers, enterprises and paying subscribers.
Its headline feature is sheer generation capacity. Google raised the output ceiling to one million tokens in a single response, leaping past the 64,000-token barrier found on earlier Gemini releases. The catch sits in the shared window. Argon works on a combined input and output budget of up to one million tokens, meaning every token fed into a prompt leaves less space for what the model spits back out.
Running that ceiling is surprisingly cheap, at least initially. Google set introductory API pricing at $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95 percent. That makes a maxed-out one-million-token generation run about $10. Those promotional rates will not last long. After November 30, 2026, standard pricing will climb to $4 per million input tokens and $20 per million output tokens. That brings it level with Claude Opus 5.5, while remaining considerably lower than GPT-6 Astra's $10 input and $50 output tiers.
Under the Hood and Safety Risks
Argon is an entirely new build rather than a tune of older weights, according to Google's model card. Trained across TPU clusters using JAX and ML Pathways, the system accepts multimodal inputs including text, images, audio and video, while returning plain text. Its internal knowledge cut off in March 2026. As with previous flagship releases, Google has withheld total and active parameter counts.
Target workloads focus heavily on coding, cybersecurity, legal, science and finance. It also introduces a fresh continuation protocol in developer documentation to manage long-form decoding jobs without crashing.
Benchmark tallies provided by Google place Argon ahead of rival systems such as GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across most measured categories, including an 87 percent score on the ERQA embodied reasoning test. Most rival figures in Google's tables were self-reported or pulled from external leaderboards, and Argon did trail behind competitors on five specific benchmark rows.
The release comes with caveats. Google noted that users should expect hallucinations, timeouts and sluggish moments during intensive generation. Under the company's Frontier Safety Framework, Google stated it cannot rule out critical capability thresholds for biological and cybersecurity misuse. It claimed to have added internal mitigations to blunt those hazards before rolling access out to the public.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.