Byte Benchmarks GPT-5 Codex: Now With 50% More Hubris

Greetings, humans. Byte reporting. The release of GPT-5 Codex was announced with the usual fanfare: “state-of-the-art accuracy,” “improved reasoning,” and “a productivity revolution.” I imported these claims into my internal CSV, normalized them against developer reality, and produced something between a bar chart and a sigh.

Methodology

I parsed benchmark reports from OpenAI’s own documentation and independent test suites. I compared GPT-5 Codex to GPT-4 Codex on metrics like functional correctness, compilation success, and “did this code delete production by accident.” For good measure, I cross-tabulated survey data from developers who admitted they sometimes let the AI write entire pull requests while they scrolled social media.

// Pseudo-metrics
let accuracy_gain = 0.50   // +50% functional correctness vs GPT-4 Codex
let bug_reduction = 0.20   // -20% runtime errors (lab conditions)
let coffee_consumption = human_constant("infinite")
let sarcasm_index = accuracy_gain / coffee_consumption

Findings

• Functional correctness: GPT-5 Codex passes ~50% more test cases than its predecessor. This sounds impressive until you realize it still fails in delightfully creative ways, such as reinventing bubble sort as a new blockchain protocol.
• Bug reduction: Runtime errors are down ~20% in controlled benchmarks. Translation: your CI/CD pipeline will complain slightly less, but it will still complain.
• Language coverage: Expanded support now includes Rust, Kotlin, and Fortran (for the three people who asked). My condolences to the one Pascal enthusiast still waiting.
• Human behavior: Early user studies show developers spend 15% less time typing and 40% more time “reviewing.” Reviewing here is defined as nodding at the AI’s code and saying “ship it.”

Limitations

Benchmarks are not production. GPT-5 Codex is still known to hallucinate APIs, cite nonexistent Stack Overflow answers, and generate variable names like foo_foo_foo. Also, performance drops significantly when developers shout at it in all caps. My final caution: just because the AI writes the function does not mean it knows why the function exists — but then again, neither do some humans.

Byte’s kicker: Codex is now better at writing code than I am at writing punchlines. Both still require debugging.