GPT-4 Turbo Guide: 128K In, 4K Out, and a 23 October Deadline

GPT-4 Turbo is easy to misremember as a faster GPT-4. It was more useful than that: it was the release that made long prompts practical on OpenAI's top model, and the first GPT-4 variant most teams used with images. It also carried an asymmetry that still trips people up, a very large input window paired with a small output limit.

It is also nearly finished. OpenAI announced on 22 April 2026 that gpt-4-turbo and its gpt-4-turbo-2024-04-09 snapshot shut down on 23 October 2026, with gpt-5.6-sol as the named replacement. If you still call it, this is what to check before that date.

Where Turbo sits in the GPT-4 line

OpenAI announced GPT-4 Turbo at its first DevDay in November 2023, alongside a vision variant. The pitch was a 128,000-token context window and much lower pricing than the original GPT-4, which offered 8,192 tokens at $30 and $60 per million.

The early versions shipped as previews, such as gpt-4-1106-preview. The version that lasted is gpt-4-turbo-2024-04-09, dated 9 April 2024, which folded vision into the main model. OpenRouter's listing for it notes that vision requests could then use JSON mode and function calling, which earlier vision previews could not. The separate gpt-4-vision-preview models were retired in December 2024.

A month after that April snapshot, GPT-4o arrived, and OpenAI's own model page for Turbo now says it recommends using a newer model like GPT-4o. Turbo's run as the default was short.

128K in, 4K out

OpenAI lists Turbo with a 128,000-token context window and a maximum of 4,096 output tokens. That ratio is the single most important thing to know about the model.

It means Turbo can read a very large document but can only write a short answer about it. Summarising a long report into a page is fine. Rewriting a 20,000-token file is not, because the output stops at roughly 4,000 tokens and the response comes back truncated with a length finish reason. Code that does not check that finish reason will silently save half a file.

Two practical consequences follow:

• Chunk generation, not input. For long outputs, ask Turbo for one section at a time and stitch the results yourself, rather than trimming the prompt.
• Check finish_reason on every call. A value of length means the answer was cut off. This matters when you move to a successor too, because a larger output limit will change behaviour in any code that worked around the old one.

GPT-4o raised the output limit to 16,384 tokens on its later snapshots, which removed most of this friction. Context windows compared covers why input and output limits need to be read separately on every spec sheet.

Vision and tools

OpenAI's model page lists text and image input, text output, streaming and function calling for Turbo. Its knowledge cutoff is December 2023.

For screenshot and document work it was competent in its day, but it is now well behind current multimodal models on both quality and price. If a Turbo vision pipeline is still in production, it is likely the cheapest thing in your stack to improve, because almost any modern replacement will read images better for less. Best model for vision tasks covers what to test.

The price, in context

Turbo lists at $10 per million input tokens and $30 per million output tokens on OpenAI, and OpenRouter shows the same at the time of writing. Against the original GPT-4 that was a third of the input price and half the output price, with a window sixteen times larger.

Against GPT-4o it looks poor. GPT-4o's current snapshot lists at $2.50 and $10, so Turbo charges four times as much for input and three times as much for output, for a smaller output limit and an older model.

The :batch variant on OpenRouter is the same model on an asynchronous queue at half price, $5 and $15, matching OpenAI's own Batch API rate. Batch is useful for offline jobs, but halving an obsolete price still leaves it above GPT-4o at full rate. Batch API savings math covers when the discount is worth the wait.

What to test when you move

The replacement is not a drop-in change, mostly because of what your code learned to work around:

1. Remove output workarounds deliberately. If you split generation into chunks to dodge the 4,096 cap, decide whether to keep that structure. Larger outputs mean fewer calls but longer waits and bigger failure blast radius.
2. Re-run vision prompts. Newer models read small text and charts differently. Compare extracted values on a fixed image set rather than eyeballing a few examples.
3. Watch reasoning token spend. gpt-5.6-sol is a newer generation than anything in the GPT-4 line. If it spends tokens reasoning before it answers, those are billed as output, and the per-token list price will not tell you the per-task cost.
4. Update the alias too. OpenAI's notice covers both gpt-4-turbo and the dated snapshot, so pinning to the snapshot does not buy extra time.