GPT-4

The First GPT-4 Production Baseline

For many teams, gpt-4 was the model that moved large language models from experimentation into higher-stakes production workflows. Applications were built around its response style, prompt sensitivity, context limits, output patterns, and relatively high cost.

The current OpenAI model entry describes GPT-4 as an older high-intelligence model usable in Chat Completions. It is text-only, with an 8,192-token context window and an 8,192-token maximum output.

That profile is very different from GPT-4 Turbo and GPT-4o. GPT-4 Turbo expanded context to 128K and added image input, while GPT-4o significantly reduced cost and broadened the mature multimodal production path.

Today, the main value of GPT-4 is not that it is the best model to choose for a new application. Its value is as a reproducible legacy reference for systems that still depend on behavior established during the first GPT-4 generation.

• Original GPT-4 API family baseline.
• Text-only input and output.
• 8,192-token context window.
• Deprecated with an announced API shutdown date.

Historical Cost Is a Strong Migration Incentive

GPT-4 pricing reflects an earlier generation of API economics. Output tokens cost twice as much as input tokens, and even moderate request volumes can become expensive compared with current alternatives.

For example, a request with 4,000 input tokens and 1,000 output tokens costs about $0.18 before accounting for retries or repeated application calls.

That can matter even more in older systems where prompts were written before aggressive cost optimization became common. Large few-shot examples, duplicated instructions, and verbose responses can multiply spend.

When evaluating a replacement, compare cost per successful task rather than token price alone. A newer model may also reduce retries, allow shorter prompts, handle structured output more reliably, or consolidate several old processing steps.

• $30.00 per 1M input tokens.
• $60.00 per 1M output tokens.
• No cached-input price is listed for the legacy GPT-4 model.
• Production cost can fall substantially when moving to a newer model that passes the same evaluation set.

A Much Narrower Integration Surface Than Modern Models

GPT-4 is text-only. Image, audio, and video input are not supported by the standard model entry.

OpenAI currently lists streaming as supported. Fine-tuning is also supported, which matters for legacy deployments that invested in specialized GPT-4 behavior.

By contrast, the current model documentation lists function calling and Structured Outputs as unsupported for gpt-4. Predicted outputs are also unavailable.

This narrower capability surface is another reason migrations can become architectural upgrades rather than simple model substitutions. A replacement can potentially eliminate custom parsing, add multimodal inputs, expose modern tool workflows, or provide schema-constrained responses that were not available in the original GPT-4 path.

Treat the Deadline as a Compatibility Project

A safe migration should reproduce the old workload rather than starting with synthetic benchmark prompts.

Collect real requests from the production distribution: short questions, long instructions, difficult edge cases, formatting-sensitive outputs, fine-tuned tasks, and prompts that historically required retries.

Run those requests against GPT-4 and the replacement while GPT-4 is still available. Compare correctness, output format, latency, token usage, cost, and any downstream validation failures.

The much larger context windows and newer capabilities of current models may allow architectural simplification, but do that after establishing parity. Separating model replacement from broader application redesign makes regressions easier to diagnose.