Google AI Pricing (October 2026)

Gemini 4 Argon pricing and paid API access

Google now lists a stable paid API endpoint: gemini-4-argon. Introductory rates are $2 per million input tokens, $0.1 per million cached input tokens and $10 per million output tokens through November 30, 2026. The structured scheduled-rate table above shows Google's December 1 standard rates and Gemini 3.8 Flash's separate January 1 increase. Output billing includes thinking. Caching is implicit only, not explicit.

Free tier and Google AI Pro access are not included; quotas and eligibility apply. The official model specification gives a shared input/output context of 1,048,576 tokens, 65,536 default synchronous output tokens, and up to 1,048,576 with long decode. The larger maximum is not the default.

Brin's recursive self-improvement remarks describe internal AI-assisted R&D, not a new SKU or customer discount. Google's safety report retains human code review and human research direction. Read the RSI claim and research-cost analysis, updated Argon launch guide, or availability API.

A published endpoint is not an independent Labs evaluation. We publish no Argon score without exact settings, raw outputs, task scores, thinking tokens, retries, latency, and billed costs. See Labs coverage notes. Recorded October 16 source review: pricing and model documentation.

WeatherNext pricing and access

WeatherNext research weights and operational forecast-data feeds are separate offerings, not a metered Gemini token endpoint. The reviewed WeatherNext repository, model guide and access guide do not publish a per-token, per-run or per-forecast model rate. This is not a zero-price quote and does not cover third-party services.

WeatherNext 3: forecast-data access

The current Google model guide recommends WeatherNext 3 for new projects. Operational forecast datasets require a Google Account allowlist request; Google says requests are typically approved in 5–7 business days. No existing paid Google Cloud contract is required.

Available surfaces: Google Cloud Storage, BigQuery, Earth Engine. Real-time forecasts relating to less than one hour ago and the future use the GDM Real-Time Weather Forecasting Experimental Data Terms of Use. Data relating to one hour ago or earlier uses CC BY 4.0. Research repository licensing does not replace these data terms.

WeatherNext 2 / Cyclones / Mini: research weights

Code and Colab notebooks use Apache 2.0; other repository materials use CC BY 4.0. These research-weight terms do not establish access to WeatherNext 3 real-time data.

The repository recommends a free Colab v5e-1 runtime for Cyclones Mini. Larger checkpoints require a v5p accelerator or H100-class GPU memory. Runtime availability is not a guaranteed free hosting allowance.

Budget compute, storage, queries, transfer, orchestration and forecast verification separately. Access eligibility does not establish zero Google Cloud usage charges; no numeric infrastructure price is quoted here.

Language-model benchmarks do not measure weather-forecast skill. Evaluate pinned atmospheric inputs, verification data, cyclone-specific metrics, ensemble size and accelerator hours before comparing total cost. The August Cyclones analysis is historical context, not the current operational-data access guide.

WeatherNext cost/access sources verified: 2026-10-25. Research repository · Current model guide · Data access and terms.

Context, caching and workload choice

Use the selected model's current status and source date in the table. Free-tier eligibility, request-length tiers, cache storage, tool fees and agent loops differ by route; a historical provider-wide discount or free allowance is not a budget guarantee.

Evaluate quality, latency and total cost per accepted task. Compare OpenAI and Claude, then use the token calculator and Claude versus Gemini guide. Consumer subscriptions and API billing are separate.