Gemini 4 Argon's Price Doubles December 1: Every AI API Deadline Through February

Google will double the API price of Gemini 4 Argon on December 1, 2026, per its pricing page. Input goes from $2 to $4 per million tokens and output from $10 to $20, with cached input going from $0.10 to $0.20. The batch and flex tiers double too. It is the largest of a dozen price changes, retirements and policy updates that land between this week and late February.

Every date below comes from the provider's own pricing or deprecation page, checked on October 20. The providers can move them, and Google calls its dates the earliest possible shutdowns.

The Gemini change is the one with a bill attached. A workload that sends 100 million input tokens and 20 million output tokens a month pays $400 at today's rate and $800 from December 1. Batch keeps its half-price discount, so a job that can wait hours goes from $200 to $400 a month. Google gave no reason on the pricing page.

OpenAI's GPT-5.6 Sol is the opposite case. Its price dropped on August 21 from $5 to $4 per million input tokens and from $30 to $20 for output, per WinBuzzer. OpenAI's pricing page says the promotional price "is available at least through November 21, 2026." It says nothing about what comes after. A team that switched to Sol for the discount should price its December on both numbers.

The shutdowns matter more than the prices for most teams. OpenAI is retiring almost every model a two-year-old codebase might still call, per its deprecations page. The October 23 batch alone covers gpt-4-turbo, gpt-4.1-nano, o1, o1-pro, o3-mini and o4-mini. The January 14 batch takes gpt-4o and gpt-4.1. Requests to a retired model fail. OpenAI also stops creating new fine-tuning jobs for existing customers on January 6, 2027, so a team that retrains a fine-tune on fresh data has until then to move it.

Finding the calls is the slow part. Anthropic's console can export usage as a CSV broken down by API key and model, which shows which keys still call Sonnet 4.5. For every provider the more reliable check is a search of code and configuration for the model strings themselves. A retired model fails at request time, often in a code path nobody has run this quarter, like a nightly report or a fallback route.

Anthropic's list is shorter. Claude Sonnet 4.5 retires on November 30, per its deprecation page, and the recommended replacement is Claude Sonnet 5.5. Two other models carry "not sooner than" dates that have arrived or are close: Claude Haiku 4.5 (October 15) and Claude Opus 4.5 (November 24). Neither has been deprecated, and Anthropic promises at least 60 days' notice before retiring a public model.

The Anthropic policy change is easy to miss because it is not a model. The new Usage Policy takes effect November 12. It bars using Claude's outputs to train an AI model "without prior authorization from Anthropic," makes customers responsible for what their agents do through tools and browsers, and asks for qualified human review where Claude supports recommendations in areas like legal, medical, finance and employment.

What about the image models this site uses? OpenAI's gpt-image-1 goes on October 23, and its replacements, gpt-image-2.5-flare and sunburst, cost $30 per million output tokens. A medium-quality card image here uses about 300 output tokens, or about a penny.

Two earlier pieces fill in the context: Gemini 4 Argon's launch, which we covered on October 2, and Meta's competing Muse Wonder pricing.