Over the span of a little more than a month, Google has launched three models—Gemini 4 Carbon, Argon, and Hydrogen—but testing shows that capability, speed, and price still cannot be fully balanced: Carbon suits complex development, Hydrogen prioritizes low-cost, rapid responses, and Claude retains an edge in engineering reliability.

What matters about this update is not simply that several more models have arrived. It signals that frontier-model products are moving from competition based on “leading a single benchmark score” toward competition based on “task cost and real-world reliability.” For enterprises, the answer may no longer be to choose one strongest model, but to route different tasks to different models.

What this is

The Gemini 4 family now has three distinct positions: Carbon targets complex tasks where capability comes first, Argon balances capability against consumption, and Hydrogen emphasizes speed and low cost. All three support a 1-million-Token context window. A Token is the basic unit models use to measure the text, code, and other material they process.

Pricing reflects this tiering. Carbon and Argon currently cost $2 and $10 per million input and output Tokens, respectively, returning to $4 and $20 from December; Hydrogen costs $1.5 and $7.5, with a free usage tier.

Industry view

One evaluator’s experience under intensive use found Carbon slightly more stable than Argon, but it could leave the AI coding tool Antigravity spending ten minutes reading code without making any changes. Hydrogen responds quickly and writes natural Chinese, but it is not proactive about reading existing project materials, making it easy to overlook available context.

Cross-model comparisons also have boundaries. For complex code repositories, branches, and version management, Claude Opus 5.5 was judged more rigorous. Gemini’s advantages lie in Chinese-language communication, copywriting, and subscription allowances. When asked to design a web interface freely, Gemini 4’s default output often feels crowded and only improves noticeably after design specifications and component constraints are added.

These conclusions warrant caution: they come from one developer’s long-term use and are affected by prompts, task types, quota plans, and tooling. They cannot replace systematic evaluation. When selecting models, teams should really compare the time required to complete the same task, the number of rework cycles, and total cost—not positions on public leaderboards.

Impact on regular people

For enterprise IT: procurement will shift from a single model toward portfolio-level orchestration: use a stronger model for review, a lighter model for code reading, research, and writing, while accounting for call volume, waiting time, and rework costs.

For individual professionals: Gemini’s Chinese-language output can help draft proposals, courses, and scripts, but code changes still require human review. For complex projects, clearly define the scope of available materials and acceptance criteria.

For the consumer market: free and low-cost tiers will make everyday writing, file reading, and simple programming assistants more widely available. But fast responses do not guarantee reliable results, and important decisions still cannot be handed directly to a model.