As of early 2025, OpenAI has not officially released GPT-5 or its coding benchmarks. However, based on leaked reports and industry speculation, GPT-5 is expected to achieve approximately 85-90% on HumanEval (compared to GPT-4’s 67%) and 75-80% on CodeContests, representing a significant leap in code generation capabilities.

What Coding Benchmarks Will Measure GPT-5’s Performance?

GPT-5 will likely be evaluated on established coding benchmarks including HumanEval (function-level Python problems), MBPP (Mostly Basic Python Problems), CodeContests (competitive programming), and potentially newer multi-language assessments. These benchmarks test everything from basic syntax to complex algorithmic reasoning. Early insider reports suggest GPT-5 could approach or exceed 85% on HumanEval, a substantial improvement over GPT-4’s 67% baseline score.

How Will GPT-5 Compare to GPT-4 in Coding Tasks?

The expected improvements center on three areas: better handling of complex multi-file codebases, reduced hallucination in API usage, and stronger debugging capabilities. If leaked benchmarks prove accurate, GPT-5 could represent a 20-30% relative improvement over GPT-4 on coding tasks, with particularly strong gains in languages beyond Python like Rust and C++.

When Will Official GPT-5 Coding Benchmarks Be Released?

OpenAI has not confirmed a GPT-5 release date. Most industry analysts predict a late 2025 or early 2026 launch, at which point official benchmark results will be published. Until then, all performance claims should be treated as speculative.