OpenAI models: choose a GPT tier, not just a bigger context

Start with Mini, then test the reason to move up

For rewriting, short classification and routine drafting, compare Mini and Nano using inputs you already understand. Require the same facts, requested format and refusal behavior from both. A cheaper label is not proof of a cheaper finished task: an answer that needs three corrections may consume more tokens and attention than one acceptable first response. Keep the unsuccessful attempts in your cost record instead of comparing only the best answer.

The three GPT-5.6 entries all carry a one-million-token context setting in this app. Their shared capacity cannot tell you whether Terra is sufficient for your code review or Sol fixes a subtle bug more reliably. Use Terra as a candidate for mixed everyday work, Luna for repetitive summaries, and Sol for a demanding test case with executable checks. Those are evaluation starting points from our editorial catalog, not measured speed or accuracy rankings.

Sol pricing has a deadline, not an evergreen promise

OpenAI's published GPT-5.6 Sol standard short-context promotion is $4 per million input tokens and $20 per million output tokens, available at least through November 21, 2026. At those rates, an uncached request with 8,000 input tokens and 2,000 output tokens costs $0.072 before tools or other charges. One hundred such requests cost $7.20. That calculation is a workload example, not an assertion that every document question generates that many tokens.

Do not carry that promotional rate into a yearly forecast without a second scenario. A return to different list pricing, a longer prompt tier, additional reasoning, or repeated generations changes the result. Our Mini figure remains an unconfirmed historical estimate; Terra, Luna and Nano have no verified rate in this snapshot. A missing price is deliberately not zero. Consult the linked OpenAI rate card for the exact model and billing category before deciding that one tier is economical.

GPT prefix caching: the write is part of the bill

OpenAI's current caching guide distinguishes GPT-5.6 and later from earlier models. For that newer family it documents a default thirty-minute cache lifetime with writes billed at 1.25 times ordinary input, and a twelve-hour option in the Responses API with writes at 2.5 times ordinary input. Subsequent reads on most of those models are one tenth of ordinary input. These are provider API rules; this page does not claim that Multimodel Chat exposes every retention option or guarantees a cache hit.

A stable instruction and document prefix is the natural candidate for reuse. Editing material near the beginning can defeat the saving you expected from many repeated questions. Check actual cached-token usage rather than assuming that a long conversation is discounted automatically. Cached tokens still count toward the provider's tokens-per-minute allowance. Cost reduction and throughput capacity are separate questions, so a cheaper repeated prompt can still encounter a rate limit.

Large GPT inputs need an output plan

The table reports the app's configured capacity: one million tokens for the GPT-5.6 trio and 400,000 for Mini and Nano. It does not certify that your account can send every input modality at that size. Long files, conversation history and instructions all affect the next request. For a repository investigation, first name the files and failing behavior that matter; do not fill the window simply because the configured ceiling permits it.

Before switching an important workflow, save a reference input and the expected result. For a patch, run the failing test and an adjacent regression test. For a summary, verify every ticket identifier against the supplied material. Start a fresh comparison with the same source evidence so previous model answers do not become accidental hints. The protocols below describe what to check for each GPT tier; they are not benchmark results or promises that a more expensive tier wins.

Concrete OpenAI model checks

Suggested evaluation protocols, not executed benchmark results.