Google drops Gemini 4 Argon for high-stakes enterprise work

For months, major AI labs have been engaged in a race to see who can build the cheapest lightweight model. But when it comes to refactoring millions of lines of legacy code, miniature models fall flat.

Google officially launched its latest frontier model, Gemini 4 Argon. It marks Google's first flagship-tier model above Flash in over seven months, and it is now available to developers, enterprises, and Google AI Pro and Ultra subscribers.

Google priced the model aggressively at $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95%. It also took the top spot on the Vals Index, a benchmark measuring model performance across finance, coding, legal, and tax work.

Refactoring Chrome and designing enzymes

Instead of leaning solely on generic synthetic benchmarks, Google highlighted serious enterprise and internal engineering workloads. Google reported that Argon inside its Antigravity platform helped internal teams boost agentic code commits by over 35% while cutting code rollbacks by nearly a third.

The Chrome engineering team put Argon to work inside Chromium's massive 40-million-line codebase. The model separated core webpage handling from a sprawling 6-million-line legacy directory across hundreds of files, finishing months of architectural restructuring in a matter of weeks.

Beyond software, the model made surprise inroads in wet labs and physical environments. Researchers used Argon to optimize the open-source quantum chemistry package xTB, accelerating molecular structure prediction by up to 2.3 times. It also helped design five novel serine esterases that successfully displayed catalytic activity in physical laboratory trials.

In robotics, Argon demonstrated what Google calls a video-to-sim-to-real pipeline. After watching ordinary video of a person playing cornhole, Argon mapped the environment coordinates, derived the joint physics, generated simulation code, and guided Apptronik's Apollo humanoid robot to sink seven consecutive shots.

The million-token catch and independent benchmarks

Argon's headline feature is a unified 1,048,576-token context window that can technically output up to one million tokens. But there is an architectural caveat users need to know.

The input prompt, the model's internal thinking tokens, and the visible output all share that same single pool. If Argon burns 500,000 tokens on deep reasoning to solve an intricate puzzle, you are left with at most 500,000 tokens of actual visible output.

Furthermore, API requests default to 65,536 output tokens per call. Unlocking the full million-token output requires chaining calls using continuation tokens. Subscribers in the consumer Gemini app will not have to deal with that plumbing, but developers will.

Independent evaluations from Artificial Analysis paint a nuanced picture. On its Intelligence Index, Argon scored 52.56 under high-reasoning conditions, basically matching OpenAI's GPT-6 Astra (52.67), while trailing Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5.

However, Argon pulled off an exceptional 15% hallucination rate on the AA-Omniscience test, compared to over 50% for GPT-6 Astra. The tradeoff is strict conservatism: its answer accuracy sat at 50%, behind Astra's 63%.

Why it matters

For enterprise developers grappling with legacy technical debt or researchers screening molecules, Argon is well worth testing today. Google is even offering 10,000 scientists six months of top-tier Antigravity access for free, clearly aiming to make Argon the default engine for high-stakes science and engineering.

Across the industry, this release shifts the frontier rivalry from chatbot chatter to autonomous, multi-step problem solving. By pricing input at just $2 per million tokens with huge caching discounts, Google puts immense economic pressure on Anthropic and OpenAI to cut costs on deep reasoning workloads.

We think Google made a clever strategic bet by prioritizing an ultra-low hallucination rate over raw bravado. In enterprise settings, an agent that admits what it does not know is far more valuable than one that invents plausible-sounding bugs. Still, the shared thinking-output window means long reasoning chains can eat into your output quota without warning, and a 50% accuracy score shows it still stumbles. It is not an omniscient brain, but it is an exceptionally capable workhorse.