GPT-4o (2024-05-13)

The Original API Version of GPT-4o

The model ID gpt-4o-2024-05-13 captures the launch-era behavior of GPT-4o rather than a later revision of the family. For developers maintaining older applications, that distinction can matter because later GPT-4o snapshots changed pricing, output limits, supported response formats, and model behavior.

OpenAI initially exposed GPT-4o to developers as a text-and-vision model. The API snapshot accepts text and image input and produces text output. Although the broader GPT-4o research announcement demonstrated native audio and video capabilities, those modalities were not exposed through this standard May 2024 API snapshot.

The original snapshot supports a 128,000-token context window and up to 4,096 output tokens. It launched at $5 per million input tokens and $15 per million output tokens.

• API model ID: gpt-4o-2024-05-13.
• Released with GPT-4o on May 13, 2024.
• Text and image input with text output.
• Useful primarily for legacy compatibility and historical regression testing today.

Preserve Old Visual Test Cases Before Migrating

Applications that adopted GPT-4o early often used it for screenshots, photographed documents, charts, UI states, diagrams, and other image-assisted workflows.

Those workloads should be part of any migration evaluation. A replacement model may have stronger vision overall, but production safety depends on the specific images your system processes: small interface text, dense tables, unusual crops, low-resolution scans, or visual layouts that historically caused failures.

The May snapshot gives you a launch-era reference point. Keep the image, prompt, and expected result fixed, then compare the same input against the replacement model.

The API version of this snapshot does not expose native audio or video input. The May 2024 product announcement demonstrated broader GPT-4o multimodality, but developers initially received GPT-4o in the API as a text-and-vision model.

Output Length Is a Useful Migration Test

The large input window made the launch snapshot practical for substantial conversations, documents, and image-assisted prompts. But the 4,096-token output limit can constrain long reports, code generation, or verbose structured responses.

Later GPT-4o snapshots increased the output ceiling. That means a migration can change not only quality and price, but also how much content the model can return in a single call.

When evaluating a replacement, include cases that historically hit output truncation. A newer model may simplify application logic by reducing continuation calls, partial-response handling, or manual output splitting.

• Context window: 128,000 tokens.
• Maximum output: 4,096 tokens.
• Large input capacity for a 2024 multimodal model.
• Much smaller output ceiling than later GPT-4o revisions.

This Snapshot Should Now Be Treated as a Migration Dependency

As of October 4, 2026, the shutdown date is close. OpenAI lists GPT-5.6 Sol as the recommended replacement for this snapshot.

That changes the purpose of evaluating the model. New applications should not select the May 2024 snapshot as a fresh production dependency. Existing users should instead document the behavior they need to preserve, build an evaluation set, test the replacement, and remove the deprecated model from production configuration.

A useful migration suite should include normal traffic and failure cases: image inputs, long prompts, tool calls, JSON parsing, output-length edge cases, and any requests where the old model required special handling.