# What is GPT-4 Turbo, and what changes compared with GPT-4?

Date: 2023-11-07 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-gpt-4-turbo/

**Answer:** GPT-4 Turbo is a new version of OpenAI's GPT-4 model, released in preview for developers on November 6, 2023, that knows about world events up to April 2023 and can read 128,000 tokens, roughly 300 pages, in a single prompt. It is also much cheaper to use than GPT-4, with input text priced at a third and output text at half of GPT-4's rates, and it follows formatting instructions more reliably.

Yesterday, at its first developer conference, OpenAI previewed GPT-4 Turbo, the next version of its most capable model. GPT-4 itself arrived in March and became available to all developers in July. The new version reads far more text at once, knows about more recent events, and costs considerably less to run.

For now, GPT-4 Turbo is a preview for developers who pay for OpenAI's API, the programming interface apps use to send requests to the model. They can try it today under the name `gpt-4-1106-preview`. OpenAI says a stable version, ready for production use, will follow in the coming weeks.

## What GPT-4 Turbo is

Like every large language model, GPT-4 Turbo works by predicting text. Trained on enormous amounts of writing, mostly from the web, it has learned which words tend to follow which in a given context. Give it the start of an email that ends "Looking forward…" and it will likely continue with something like "to hearing from you."

Turbo is not a brand-new kind of model. OpenAI presents it as GPT-4 improved and optimized: more capable, according to the company, and cheaper to serve. It comes in two forms: a text-only version available now, and a version that can also look at images, offered separately as `gpt-4-vision-preview`.

## What changes compared with GPT-4

| | GPT-4 | GPT-4 Turbo (preview) |
|---|---|---|
| Knowledge of world events | Up to September 2021 | Up to April 2023 |
| Context window | 8,000 or 32,000 tokens | 128,000 tokens (about 300 pages) |
| Price per 1,000 input tokens | $0.03 (8K) or $0.06 (32K) | $0.01 |
| Price per 1,000 output tokens | $0.06 (8K) or $0.12 (32K) | $0.03 |
| JSON mode | No | Yes |

A few terms explain the table. A **token** is a chunk of text, often a short word or a piece of a longer one; the word "fantastic," for instance, might be split into three tokens. A thousand tokens is roughly 750 English words. The **context window** is how much text the model can consider in one request, counting both what you send and what it writes back. **Input tokens** are what you send; **output tokens** are what the model generates.

At 128,000 tokens, GPT-4 Turbo's window is four times the largest GPT-4 version, and bigger than the 100,000 tokens of Anthropic's Claude 2, until now the largest available in a commercial model. That is enough for a long report, a contract with all its appendices, or a sizable chunk of a software project in one go.

## Better at following instructions

OpenAI also says GPT-4 Turbo does a better job when a task demands careful obedience to instructions, such as replying in an exact format. Three developer features support this:

- **JSON mode.** JSON is a common text format apps use to exchange data. With a new setting called `response_format`, the model is constrained to produce a valid JSON object, so an app can read the reply without breaking.
- **Better function calling.** Developers can describe functions in their app, and the model chooses which to call and with what details. It can now request several calls in one message, and OpenAI says it gets the details right more often.
- **Reproducible outputs.** A new `seed` setting makes the model return the same answer to the same request most of the time, which helps with testing and debugging. OpenAI also says it will soon return the probabilities of the most likely words, useful for features like autocomplete.

## A worked example: what a long document costs

Suppose an app summarizes a 20,000-token document (around 15,000 words) and the summary is 1,000 tokens long.

| Model | Fits? | Input cost | Output cost | Total |
|---|---|---|---|---|
| GPT-4 (8K) | No, the document is too long | | | |
| GPT-4 (32K) | Yes | 20 × $0.06 = $1.20 | 1 × $0.12 = $0.12 | $1.32 |
| GPT-4 Turbo | Yes, with lots of room to spare | 20 × $0.01 = $0.20 | 1 × $0.03 = $0.03 | $0.23 |

The same job costs less than a fifth as much, and GPT-4 Turbo could take a document about six times longer before running out of room. For an app that runs this task thousands of times a day, that difference decides whether a feature is affordable.

> **Try it:** Take a document you know well, such as a report you wrote. Estimate its length in tokens by dividing its word count by 0.75. Then work out what one summary would cost with GPT-4 (32K) and with GPT-4 Turbo using the prices in the table above.

## Why it matters

Price and window size set the limits of what developers can build. Many projects that were technically possible with GPT-4 were too expensive to offer widely, or required chopping documents into pieces and stitching answers together. GPT-4 Turbo loosens both limits at once.

The newer knowledge cutoff matters too. A model trained on data up to April 2023 can answer questions about events from 2022 and early 2023 that GPT-4 had never seen, though it still knows nothing after its cutoff unless given that information.

Some caution is warranted. The model is a preview, and OpenAI's claims of greater capability come from its own testing. Developers will be comparing it with GPT-4 on their own tasks over the coming weeks.

## Should you switch from GPT-4 to GPT-4 Turbo now?

If you are experimenting or building a prototype, there is little reason not to try it. The lower price and larger window make testing cheaper, and changing the model name is usually a one-line edit. Run the same set of prompts through both models and compare quality, speed and cost on your own examples.

If you run an app in production, be more careful. Previews can change, and OpenAI says its stable, production-ready version is still a few weeks away. A sensible path is to test now, note where the answers differ, and switch fully when the stable version arrives.

If you use ChatGPT rather than the API, you do not choose a model name directly. OpenAI's DevDay announcements for ChatGPT are separate, so check what your plan offers inside the app itself.

## Sources

- New models and developer products announced at DevDay: https://openai.com/index/new-models-and-developer-products-announced-at-devday/
- OpenAI debuts GPT-4 Turbo and fine-tuning program for GPT-4: https://techcrunch.com/2023/11/06/openai-launches-gpt-4-turbo-and-launches-fine-tuning-program-for-gpt-4/