---
title: The Smart tier
description: The Smart tier runs gpt-5.6-terra by default — $3.00 per million input tokens and $18.00 per million output tokens. What lands here, what it costs, and what happens when it cannot run.
---

{/*
  ============================================================================
  DO NOT EDIT THIS FILE BY HAND. IT IS GENERATED, AND YOUR EDIT WILL BE LOST.

  Generated by : apps/platform/scripts/build-docs-model-guides.ts
                 (in the chatinone-platform repo)
  Generated from: packages/core/data/models.json  ->  pickerModels(CATALOG_MODELS)
                 apps/platform/src/model-routing.ts  (DEFAULT_TIER_BINDINGS)
  This page     : Smart tier

  To change anything below, change the source in that repo and re-run:
      pnpm --filter @chatinone/platform gen:docs-model-guides -- --write

  Committed snapshots live in that repo under
  apps/platform/test/fixtures/docs-model-guides/, and
  apps/platform/test/docs-price-table-drift.test.ts fails if a snapshot, or the
  SET of pages, disagrees with the regenerated output.

  Prices here are RETAIL — what you are actually charged. They come from the
  same projection the in-product engine picker is built from, so this page and
  the product cannot disagree.
  ============================================================================
*/}

**Smart** is one of four options in the composer's engine menu, alongside **Auto**, and the other two tiers. It exists to be the balanced default: good answers, long context, no drama.

## What it runs

The Smart tier runs `openai/gpt-5.6-terra` by default, made by OpenAI.

| Runs as | Input / 1M tokens | Output / 1M tokens |
| --- | --- | --- |
| **Smart** tier default | $3.00 | $18.00 |
| **Smart** tier default, repeated context | $0.30 | — |

The second row is the **repeated-context rate**. When a request re-sends context the model has already seen, those tokens are billed at that lower rate instead of the full input rate — it is applied automatically by the biller, not something you switch on. See [How billing works](/docs/how-billing-works).

You can point this tier at a different model in Settings → *How it works* → **Set an engine per speed tier**, and you can override it for a single request from **Pick an exact engine**.

## What lands here

Send work here when there is real material to read and a real answer to write: summarise a long document, draft something from sources you attach, work through a task that has a few steps, answer a question that needs the whole thread rather than the last message.

This is the tier to reach for when you have no particular reason to choose. If the work turns out to be a long chain of interdependent steps, **Max** is the one built for that.

Two things this page deliberately does not tell you: how **Auto** decides, and how a request moves between models. Those are internal and we do not publish them. What we do publish is the above — pick a tier and you know which model runs and what it costs.

## What it costs on a real run

Both examples below run as the **Smart** tier default, at the two rates above:

| Runs as | Run shape | Cost |
| --- | --- | --- |
| **Smart** tier default | 4,000 in, 800 out (a short question with a little context) | **$0.0264** |
| **Smart** tier default | 100,000 in, 10,000 out (a long document and a long answer) | **$0.48** |

That is the whole arithmetic: tokens × the two rates. Nothing is rounded up, there is no per-run fee, and the receipt itemises every line. The ceiling on one request is this model's context window, 1.1M tokens (about 788,000 words, roughly 1.6k pages) — instructions, attachments and the answer all share it.

## What happens when it cannot run

If `gpt-5.6-terra` loses its price or its capacity, this tier does **not** silently run something else:

- If you had re-bound **Smart** to your own choice and that choice lost supply, the tier falls back to its own default and **tells you it did** — you get a notice naming what you asked for and what actually ran.
- If the default itself is unavailable too, the tier reports itself unresolved rather than borrowing another tier's model. There is no quiet cross-tier substitution.

This is the only failure behaviour we publish, because it is the only one we can point at in the code that implements it.

## What you will and will not notice

You will notice: which tier you are on (the engine menu shows it), a fallback when one happens (it is reported, not hidden), and every charge (itemised on the receipt — see [Reading a receipt](/docs/reading-a-receipt)).

You will not notice, because we do not expose it: how **Auto** chooses, or anything about where capacity comes from.

## What we have not verified

- We publish no measured speed for this tier. Tiers are named for what they are **for**, not for how fast they run: "Smart" is a meaning, not a benchmarked latency, and we have no timings we would stand behind.
- We publish no measured quality comparison between tiers. Nothing here claims Smart answers better or worse than another tier on your work.
- Image input on `gpt-5.6-terra` has **not** been measured. If you need to hand over a picture, use a model from [Models that can read your images](/docs/models-that/read-your-images) instead of assuming this tier can.

## Next

- [Model prices](/docs/model-prices) — every model and every price in one table
- The other tiers: [Fast](/docs/tiers/fast) and [Max](/docs/tiers/max)
- Start a run at [elsedoes.com/home](https://elsedoes.com/home)

**Prices last changed 2026-07-29** (source `ced6a4f`). Every figure here is generated from the same pricing code that bills your runs.