GPT-4 in 2026: What It Was, What It Costs and Moving Off It

Plenty of production code still sends model: "gpt-4". Usually nobody chose it recently. It was the best option when the integration was written, it kept working, and the line never came up in review again. In October 2026 that line costs more per token than almost anything else in OpenAI's catalogue and has days left to run.

OpenAI announced on 22 April 2026 that gpt-4 and its gpt-4-0613 snapshot shut down on 23 October 2026, with gpt-5.6-sol named as the replacement. This guide is for the person who has just found that line and needs to know what it was doing.

What GPT-4 changed in March 2023

OpenAI released GPT-4 on 14 March 2023 and published a technical report the next day. The report describes a large multimodal Transformer, pre-trained to predict the next token and then aligned with post-training. It deliberately left out the architecture, parameter count and training details, and that has never changed. Any figure you see for GPT-4's size comes from press reporting, not from OpenAI.

The headline claim was human-level performance on professional exams, including a simulated bar exam score around the top 10% of test takers. That figure was later challenged by an independent re-analysis that argued the percentile was overstated depending on which cohort you compare against. It is a useful early example of why exam-style results need reading with care, which benchmark contamination explained covers in more depth.

What GPT-4 actually changed in practice was reliability on multi-step instructions. It was the first model many teams trusted to follow a long system prompt, produce valid structured text most of the time, and reason through a problem without derailing. A lot of prompt engineering habits still in use were formed against it.

The spec sheet in 2026 terms

OpenAI's model page now describes GPT-4 simply as an older high-intelligence model. The numbers explain why:

• Context: 8,192 tokens, shared between prompt and completion.
• Modalities: text in, text out through the API.
• Price: $30 per million input tokens and $60 per million output tokens, or $15 and $30 through the Batch API.
• Snapshots: gpt-4-0613 is the default. The older gpt-4-0314 was shut down on 26 March 2026.

Put those against GPT-4o at $2.50 and $10 with a 128,000-token window. GPT-4 costs twelve times as much per input token and six times as much per output token for a window one-sixteenth the size. There is no workload where that trade made sense recently, other than not wanting to touch working code.

Knowledge cutoff is murkier than it should be. OpenRouter's catalogue lists September 2021, while OpenAI's current model page shows December 2023. Both cannot describe the same frozen weights, so treat the date as unknown and do not rely on GPT-4 for anything that depends on recent events without testing it.

The 32K version, gpt-4-32k, is already gone. OpenAI retired it on 6 June 2025 and pointed users at GPT-4o.

Why it still shows up in code

Three patterns account for most surviving GPT-4 calls:

1. Hard-coded defaults in libraries. Older SDK wrappers, agents and evaluation tools shipped with gpt-4 as the default model string. Upgrading the dependency often changes it; pinning an old version keeps it.
2. Grader and judge prompts. Teams that built evaluation suites in 2023 used GPT-4 as the judge and never moved it, because changing the judge changes every historical score.
3. Fine-tuned models. Fine-tuned GPT-4 models, listed as ft-gpt-4, are on the same 23 October shutdown.

Search for the string across every repository and configuration store, not just the main application. Judge prompts and batch scripts are where it hides.

Replacing a GPT-4 judge is the hard case

Replacing GPT-4 as a generator is mostly an upgrade: almost any current model will be cheaper and better. Replacing it as a grader is not, because your historical scores were produced by GPT-4's particular leniencies.

Run the new judge over a sample of outputs you have already scored with GPT-4, and look at agreement rather than average score. If the two judges disagree on more than a small fraction of items, you have a calibration break, and trend lines across the switch are no longer comparable. Note the date of the change next to your metrics so nobody reads a jump in scores as a jump in quality.

A short migration checklist

• Change the model string, then the token limits. Code written for an 8K window often truncates input aggressively. A modern window may let you remove that truncation, which can change outputs more than the model swap does.
• Re-check temperature and length. Newer models tend to write longer answers by default. If downstream code parses fixed formats, test it.
• Consider a reasoning model only where it is needed. gpt-5.6-sol is OpenAI's named replacement, but for simple rewriting or extraction a smaller model may do the job for far less.
• Compare cost per task. Almost any replacement will be cheaper than $30 and $60, but reasoning models bill thinking tokens as output, so measure rather than assume.

Model deprecation and migration covers the general process, and the GPT-4 Turbo guide covers the sibling model shutting down on the same day.