GPT-4 Turbo
Last reviewed
Sources
17 citations
Review status
Source-backed
Revision
v2 ยท 1,846 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
17 citations
Review status
Source-backed
Revision
v2 ยท 1,846 words
Add missing citations, update stale details, or suggest a clearer explanation.
GPT-4 Turbo is a family of large language models released by OpenAI as a faster, cheaper, and longer-context variant of GPT-4. It was announced on November 6, 2023, at OpenAI's first developer conference, DevDay, with a headline feature of a 128,000-token context window, roughly four times that of the original GPT-4 and equivalent to about 300 pages of text in a single prompt.[1][2] At launch GPT-4 Turbo cost $0.01 per 1,000 input tokens and $0.03 per 1,000 output tokens, which OpenAI described as three times cheaper for input and two times cheaper for output than GPT-4.[1][2] GPT-4 Turbo was distributed through the OpenAI API as a series of preview snapshots before reaching general availability on April 9, 2024, and it became the default GPT-4 class model in ChatGPT for paying users. It occupied the tier between GPT-4 and its multimodal successor GPT-4o, released May 13, 2024.[3]
When GPT-4 launched in March 2023, its public API offered context windows of 8,192 tokens and a 32,768-token variant, with a knowledge cutoff of September 2021. Developers requested longer context, lower prices, and more recent training data. GPT-4 Turbo was OpenAI's response. At DevDay, CEO Sam Altman presented it as "more capable" and "cheaper" than GPT-4, extending the model's knowledge of world events to April 2023 and quadrupling the maximum context length.[1][2]
OpenAI initially shipped GPT-4 Turbo as a preview rather than a production-ready model, describing the early snapshots as not yet suitable for production use and promising a stable version in the following weeks.[1] The preview was made available to all paying developers through the API on the day of the announcement.[2]
GPT-4 Turbo was announced on November 6, 2023, at OpenAI's inaugural DevDay conference and shipped the same day as a preview through the API.[1][2] It reached general availability as the gpt-4-turbo-2024-04-09 snapshot on April 9, 2024, and OpenAI rolled the upgraded model out to ChatGPT's paid Plus, Team, and Enterprise tiers in the week of April 12, 2024.[3][16] The full sequence of dated snapshots is detailed in the model versions section below.
The defining capability of GPT-4 Turbo is its 128,000-token context window, the largest among widely available commercial models at the time of its announcement and large enough to process the equivalent of more than 300 pages of text in one request.[1][2] This enabled use cases such as analyzing long documents and supporting retrieval-augmented generation workflows with substantially more supplied context.
GPT-4 Turbo also carried a more recent knowledge cutoff than the original GPT-4. The first preview was trained on data up to April 2023, and the general availability release pushed the cutoff to December 2023.[2][4] OpenAI described the April 2024 general availability model as "majorly improved," with better performance in areas including writing, mathematics, reasoning, and coding compared with earlier Turbo previews.[5] On standard benchmarks, the GPT-4 Turbo generation scored about 86.5 on the Massive Multitask Language Understanding (MMLU) test, a measure later exceeded by GPT-4o's 88.7.[16][17]
GPT-4 Turbo was released as a sequence of dated snapshots. Reproducible behavior was a stated goal, so OpenAI exposed specific model names rather than only a moving alias.
| Snapshot | Released | Knowledge cutoff | Notes |
|---|---|---|---|
gpt-4-1106-preview | Nov 6, 2023 | April 2023 | First GPT-4 Turbo preview, text input only.[1][2] |
gpt-4-vision-preview | Nov 2023 | April 2023 | Preview adding image input (also gpt-4-1106-vision-preview).[6] |
gpt-4-0125-preview | Jan 25, 2024 | April 2023 | Reduced "laziness" in task completion; fixed a non-English UTF-8 bug.[7][8] |
gpt-4-turbo-preview | Jan 2024 | April 2023 | Alias pointing to the latest preview (then gpt-4-0125-preview).[7] |
gpt-4-turbo-2024-04-09 | Apr 9, 2024 | December 2023 | General availability; vision built in; aliased by gpt-4-turbo.[3][4][5] |
The gpt-4-0125-preview update, released alongside new embedding models and price changes on January 25, 2024, was intended to address user complaints that the November preview had become "lazy," sometimes truncating code or declining to finish tasks. OpenAI said the new snapshot completed tasks such as code generation more thoroughly and fixed a bug affecting non-English UTF-8 outputs.[7][8]
A significant structural change arrived with general availability. The November preview required a separate gpt-4-vision-preview model for image understanding, but the April 9, 2024 release, gpt-4-turbo-2024-04-09, folded vision into the main model. OpenAI stated that vision requests with this model could also use JSON mode and function calling.[3][4] The gpt-4-vision-preview snapshot was subsequently deprecated, with a shutdown date of December 6, 2024.[6]
GPT-4 Turbo was priced substantially below the original GPT-4. At launch the input price was $0.01 per 1,000 tokens and the output price was $0.03 per 1,000 tokens, which corresponds to $10 per million input tokens and $30 per million output tokens.[1][2] OpenAI described this as three times cheaper for input and two times cheaper for output relative to GPT-4.[1][2]
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| GPT-4 (8K context) | $30 | $60 |
| GPT-4 Turbo | $10 | $30 |
| GPT-4o (launch, May 2024) | $5 | $15 |
Image inputs were billed based on resolution; TechCrunch reported a cost of about $0.00765 for processing a 1,080 by 1,080 pixel image.[2] The general availability release in April 2024 retained the $10 input and $30 output per million token pricing.[9]
GPT-4 Turbo with Vision accepts image inputs and produces text outputs, allowing applications to analyze photographs, screenshots, charts, and other visual content. Image understanding first shipped in the gpt-4-vision-preview snapshot in November 2023 and became part of the standard model with the April 9, 2024 general availability release. OpenAI highlighted early adopters using the vision capability, including the AI coding agent Devin from Cognition, the nutrition app Healthify, and the whiteboard tool tldraw.[3][10] OpenAI's broader image, audio, and text input features for GPT-4 had initially launched in September 2023, predating the API-level vision availability.[10] The general availability of vision through the API meant developers could combine image inputs with JSON mode and function calling in a single request.[3][4]
The DevDay release bundled GPT-4 Turbo with several new API capabilities aimed at developers.[1][2]
seed parameter that, when fixed along with other parameters, produces mostly deterministic outputs across calls. At introduction this was in beta and supported only for gpt-4-1106-preview and gpt-3.5-turbo-1106.[11]logprobs option returning the log probabilities of output tokens, useful for confidence estimation and classification.[12]DevDay also introduced products beyond GPT-4 Turbo itself, including the Assistants API for building agent-like applications, API access to the DALL-E 3 image model and to text-to-speech voices, and an experimental fine-tuning program for GPT-4.[1][2] These were separate offerings announced at the same event rather than features of the GPT-4 Turbo model.
GPT-4 Turbo was received as a meaningful upgrade on price and context length, with the 128,000-token window noted as exceeding contemporaries such as Anthropic's Claude 2.[2] Developer feedback was not uniformly positive. In late 2023 users reported that the November preview had grown "lazy," refusing to complete code or returning placeholders, which prompted the January 2024 gpt-4-0125-preview snapshot specifically aimed at the issue.[7][8] Independent benchmarking found that the January snapshot performed similarly to the November one on capability measures, consistent with OpenAI framing it as a reliability fix rather than a step change in intelligence.[8]
The April 2024 general availability release drew attention for its "majorly improved" claim and for finally bringing vision into the production API. OpenAI said the model "performed better than its predecessor across all six question sets" in its Simple Evals suite, with the largest gain on Google's graduate-level GPQA benchmark, where gpt-4-turbo-2024-04-09 answered 49.1 percent of questions correctly, described as a nearly 10 percent improvement over the prior version.[16] OpenAI summarized the release by stating, "We've improved capabilities in writing, math, logical reasoning, and coding."[16] Coverage noted the improved model would also roll out to ChatGPT.[3][5]
GPT-4 Turbo's position as OpenAI's flagship was brief. On May 13, 2024, about five weeks after Turbo reached general availability, OpenAI introduced GPT-4o, a natively multimodal model handling text, audio, and image in a single neural network. OpenAI said GPT-4o matched GPT-4 Turbo's intelligence on text and code while being roughly twice as fast, 50 percent cheaper, and offering five times higher rate limits, launching at $5 per million input tokens and $15 per million output tokens.[3][17] On the MMLU benchmark, GPT-4o scored 88.7 against 86.5 for GPT-4 Turbo.[17] The principal architectural difference is that GPT-4 Turbo bolts image input onto a primarily text model and routes audio through separate speech-to-text and text-to-speech systems, whereas GPT-4o processes text, audio, and image natively end to end. Later models in the GPT-4.1 and GPT-5 families further displaced Turbo.
In ChatGPT, OpenAI announced in April 2025 that the GPT-4 class model would be replaced by GPT-4o, while GPT-4 Turbo remained accessible through the API.[14] OpenAI's deprecation schedule subsequently set the preview alias gpt-4-0125-preview (and gpt-4-turbo-preview) for shutdown on March 26, 2026, and the general availability gpt-4-turbo and gpt-4-1106-preview snapshots for shutdown on October 23, 2026, with newer models recommended as replacements.[15] As of mid-2026, the general availability GPT-4 Turbo model therefore remained available in the API ahead of its scheduled retirement.