Gemini prices double. GPT-4o retires 14 Jan 2027.

Choosing a model used to be a one-time decision. In 2026 it is a calendar. Of the 26 models on the Good AI Prism model cards, several change price or disappear before the first quarter of 2027 ends, and most of the others carry no published end date at all. The useful question is no longer "which model is best?" but "which model, for which job, until when?".

Start with the job, then the price unit

Model prices are quoted per million tokens, input and output separately. That unit hides a wide range. At the top, GPT-6 Astra and Claude Fable 5.1 list at $10 input and $50 output. At the bottom, GPT-6 Luna and Claude Haiku 5.5 start at $0.10 input and $0.50 output. For the same million input tokens, that is a hundredfold span.

100×. The spread in list prices per million input tokens across the 26 models, from $0.10 to $10 (provider price pages, read 25 Oct 2026).

Most business jobs do not need the top of that range. Classification, routing, extraction and first-line answers run well on the low-cost tier; long reasoning, complex code and multi-step agents may justify the top. A sensible setup routes each task to the cheapest model that passes your own test set. That is also why 81% of large companies already use three or more model families (a16z, Jan 2026).

The middle of the market has converged on one price point. Claude Sonnet 5.5, GPT-6 Sol and Gemini 4 Argon all list at $2 input and $10 output per million tokens today. You can compare these three side by side,gemini-4-argon) on price, context, licence and retirement in the same units.

Then check the date

Retirement is where buyers get caught. OpenAI shut down two older snapshots on 23 October 2026 and retires five more GPT-4o and GPT-4.1 snapshots on 14 January 2027, according to its deprecations page. Any tool built on them changes with them, whether or not its vendor tells you. Anthropic retires Claude Sonnet 4.5 on 30 November 2026.

Prices move too. Google doubles the list price of Gemini 4 Argon on 1 December 2026, from $2 / $10 to $4 / $20, and Gemini 3.8 Flash doubles on 1 January 2027, per its published price list. A workload that is cheap in November can cost twice as much in January without a single line of code changing.

Some providers now publish the opposite: a floor. Anthropic states it will not retire Claude Opus 5.5 before 22 September 2027, Sonnet 5.5 before 28 September 2027 and Haiku 5.5 before 7 October 2027. A published minimum life is a buying criterion in its own right, because it tells you how long your evaluation stays valid.

Open weights buy control, not a free lunch

Eight of the 26 models offer open weights you can host yourself today, from gpt-oss-120b under Apache 2.0 to Llama 4 Scout with a 10-million-token context window. They carry no list price, but they do carry hosting, operations and licence terms. The Llama 4 licence is a custom commercial licence. Pharia-1 from Aleph Alpha is open for research and education only. Mistral Medium 3.5 uses a modified MIT licence. Read the licence before you read the benchmark.

Location matters as much as licence. Mistral stores API data in the EU by default. DeepSeek states that API data is stored in the People's Republic of China, under PRC law. OpenAI offers data residency in named regions through its sales team, and Anthropic's models also run on Amazon Bedrock, Google Vertex AI and Microsoft Foundry. If your data cannot leave the EU, that one fact removes options faster than any price.

The EU sets the frame

Since 2 August 2026, providers of general-purpose AI models face enforceable EU duties under the AI Act as amended in 2026. For buyers, the practical effect is documentation: model providers must supply technical information downstream. Ask for it. The AI Models technology brief explains which documents to request and how to use them.

Five checks before you commit

1. Name the job and build a small test set from your own cases.
2. Price the task, not the token: tokens per task × list price × volume, for input and output.
3. Write down the retirement date or the published minimum life, and put it in the contract.
4. Check licence and location for every model inside the tools you buy, not just the ones you call.
5. Keep a second model ready that passes the same test set, so a retirement or price change is a switch, not a project.

The model cards are refreshed as providers change their pages. Each fact shows its source and the date we read it, so you can see when it was last true. Use the Compare tool to put two or three side by side, and the ETHORITY Intelligence signals to see the next date coming.