OpenAI Unveils GPT-4 Turbo, GPTs and the Assistants API

On November 6, 2023, at its first DevDay, OpenAI launched GPT-4 Turbo with a 128,000-token context window and lower prices, alongside no-code customizable GPTs, an upcoming GPT Store, and the Assistants API for building agents. Source.

OpenAI held its first DevDay today in San Francisco, a developer conference where Sam Altman unveiled the company's biggest package of updates since the launch of GPT-4 in March. The headline announcement is GPT-4 Turbo, a version of the flagship model that dramatically expands its capacity and cuts its cost, alongside two new tools designed to let anyone build applications on top of it: GPTs and the Assistants API.

GPT-4 Turbo: more context, lower cost

GPT-4 Turbo handles a context window of 128,000 tokens — the equivalent of more than 300 pages of text in a single conversation. Until now, the standard version of GPT-4 was capped at 8,000 tokens, with an extended 32,000-token variant reserved for a limited set of customers. Announcement source.

The model also updates its knowledge cutoff to April 2023, up from the September 2021 limit that earlier versions carried.

The other major change is pricing. According to OpenAI, GPT-4 Turbo costs $0.01 per 1,000 input tokens and $0.03 per 1,000 output tokens, compared with $0.03 and $0.06 respectively for standard GPT-4 — three times cheaper on input requests and half the cost on output. For companies processing large volumes of text, the difference is substantial. Source.

GPT-4 Turbo is rolling out in preview through the API under the name gpt-4-1106-preview, and it includes features aimed squarely at developers: a JSON mode that forces responses into that format, the ability to set a seed for more reproducible outputs, and support for calling multiple functions in a single request.

OpenAI is also folding into a single model capabilities that previously lived separately: vision (GPT-4's ability to interpret images), image generation via DALL-E 3, and text-to-speech conversion, all accessible from the same API.

GPTs: custom chatbots with no code required

The second major announcement is GPTs, custom versions of ChatGPT that any user can build without programming. All it takes is giving instructions in natural language, uploading documents to serve as additional knowledge, and, if needed, connecting actions that call external services. The result is a specialized assistant — for drafting reports in a particular style, answering questions about an internal manual, or helping with a very specific task — that behaves like an app inside ChatGPT.

GPTs are now available to ChatGPT Plus users and ChatGPT Enterprise customers. OpenAI also announced it will launch a GPT Store later this month, a marketplace where creators can publish their GPTs and, according to the company, earn a share of revenue based on the usage they generate.

Assistants API: agents with memory and their own tools

For developers building more complex applications, OpenAI introduced the Assistants API, designed specifically for creating agent-style assistants: programs capable of holding conversations with persistent memory (through "threads" that retain conversation history), running code via a built-in interpreter, querying proprietary documents through retrieval, and calling external functions — all managed by OpenAI on the backend, so developers don't have to build that infrastructure from scratch.

Other measures for developers and businesses

OpenAI used the event to announce a Custom Models program, through which its researchers work directly with companies to train bespoke versions of its models using proprietary data. It also introduced Copyright Shield, a commitment under which OpenAI will cover the legal costs of its enterprise and API customers if they are sued for copyright infringement arising from the use of its tools. In addition, the company announced it will double the tokens-per-minute limits for existing GPT-4 customers.

What changes for developers and businesses

The message from DevDay is clear: OpenAI wants building on its models to be cheaper, simpler, and less dependent on in-house infrastructure. The price cuts and expanded context window open the door to applications that were previously uneconomical, such as analyzing entire long documents in a single call. GPTs, meanwhile, put AI customization in the hands of users with no technical background, while the Assistants API targets developers who want to build agents without reinventing memory management and tooling every time. It remains to be seen how the market responds to the upcoming GPT Store and what role it will play against competitors that already offer their own ecosystems of plugins and assistants.

A large context window is not perfect memory

Accepting 128,000 tokens describes a maximum input size; it does not guarantee equal use of every detail. A document may fit while a condition in its middle is still overlooked. A long-context test should distribute relevant facts near the beginning, middle and end, ask questions that require combining passages and verify citations. Merely loading a file without a size error proves very little. Data source.

The DevDay technical announcement also separates the model, tools and conversation state. The model generates; Retrieval finds material; Code Interpreter runs code; function calling requests a developer-defined function; threads retain messages. Calling the entire assembly an “agent” may be convenient, but it obscures where a failure occurred.

If an answer uses the wrong fact, ask whether the passage entered the context, whether retrieval selected the right document or whether the model misread what it received. If an action fails, inspect the function arguments and the external system's response. That separation turns a vague impression into a reproducible diagnosis.

Price per token is not cost per task

A lower tariff can make each call cheaper while increasing total workflow cost if the model needs retries, produces longer outputs or carries a growing context. The useful calculation adds input, output, tools, storage, searches and repetitions for a completed task with a defined quality threshold. It also measures latency and the human correction rate.

The announced seed was intended to make results more consistent, but OpenAI described it as a beta feature and said it worked “most of the time”. A serious test preserves the exact model, parameters, prompt, tool responses and date. A seed alone does not freeze a remote service that may change.

The transferable skill is to decompose an AI application into context, retrieval, generation, tools and state. Test each component separately and calculate cost per accepted outcome. A dense product announcement then stops being a list of names and becomes an architecture that can be audited.

Tools need one further check: a controlled failure. Make a test function return an error, let a search find nothing and place hostile instructions inside a file. The application should expose the limit without inventing success, prevent the document from replacing developer rules and preserve enough of a log to replay the incident. The ability to fail legibly is part of the product.

The evaluation record should travel with the result. Without it, an excellent answer does not reveal whether it was the first attempt, used an external tool or consumed an exceptional context. With it, a team can compare versions without moving the goalposts.