Six Conditions for a Reproducible GPT-5 vs. GPT-4 Benchmark

This article was first published as a practical GPT-5 versus GPT-4 comparison on August 16, 2025. Results from that time cannot remain an undated current ranking after models, interfaces, and evaluation conditions change. The original publication date remains while the article now records the verifiable history and comparison method.

This article defines six conditions for a reproducible GPT-5 versus GPT-4 benchmark after model retirement.

Official launch record

OpenAI published Introducing GPT-5 for developers on August 7, 2025 with API models, parameters, and launch-time evaluations. It is not a current ChatGPT GPT-4 comparison table.

Subsequent change

OpenAI's ChatGPT model retirement notice states that GPT-4o, GPT-4.1 variants, and GPT-5 Instant and Thinking were retired from ChatGPT in February 2026. It also said there was no simultaneous API change. ChatGPT and API availability must therefore remain separate.

The old question “Which should I choose now, GPT-5 or GPT-4?” is no longer directly applicable to the current ChatGPT picker.

Six reproducibility conditions

Product surface: ChatGPT / API
Model identifier:
Run date:
Settings and tools:
Fixed input:
Scoring method:

Record the date because behavior can change. A run with web search or code execution is not the same experiment as one without those tools.

Evaluation tasks

1. answer-keyed information extraction;
2. a small code repair with a test;
3. constrained planning with prohibited actions;
4. a summary checked against its source.

Record correct answers, constraint violations, unsupported claims, time, and cost. Keep prose preference as a separate subjective measure.

Identify current candidates

Use Model Release Notes, the live account, and current API documentation before selecting models for a new comparison. An old screenshot is not evidence of present availability.

Conclusion

A 2025 GPT-5 versus GPT-4 result is a historical observation, not a current guarantee. Record surface, identifier, date, settings, input, and scoring for a reproducible comparison. Select current models from current official documentation.