OpenAI released GPT-5 on August 7, 2025. The company described it as "our best AI system yet" — a claim backed by official benchmark data that is publicly available and worth examining carefully. This article looks at what OpenAI actually published, what those numbers mean in practice, and where honest limitations remain.
What OpenAI Published
According to OpenAI's official GPT-5 announcement, the model achieves the following on publicly documented benchmarks:
- SWE-bench Verified (real-world software engineering tasks): 74.9%
- AIME 2025 (mathematics competition): 94.6% without tools
- MMMU (multimodal understanding): 84.2%
- HealthBench Hard: 46.2%
- Aider Polyglot (coding across languages): 88%
On hallucination specifically — a critical metric for production use — OpenAI states that GPT-5 with thinking produces approximately 80% fewer factual errors than o3 on LongFact and FActScore benchmarks. On real-world production traffic prompts with web search enabled, responses are approximately 45% less likely to contain a factual error than GPT-4o.
- 74.9% on SWE-bench Verified (real-world coding tasks)
- 94.6% on AIME 2025 mathematics benchmark (without tools)
- ~80% fewer factual errors than o3 on open-ended fact-seeking benchmarks
- ~45% fewer factual errors than GPT-4o on production traffic with web search
- 400,000 token context window (272K input + 128K output)
The Architecture Behind the Results
GPT-5 introduces what OpenAI calls a "real-time router" — a system that analyses the complexity and type of each query to determine whether to use a fast-response model or engage deeper chain-of-thought reasoning ("GPT-5 thinking"). According to OpenAI's technical documentation, this routing is trained continuously on real signals including user model switches and measured correctness.
The practical implication is that users no longer need to explicitly select between speed and quality modes. The model makes that decision automatically — though users can override it by specifying reasoning effort in the API via the reasoning_effort parameter.
What the Benchmarks Don't Capture
Benchmarks are an imperfect proxy for real-world utility. SWE-bench, for example, tests on a fixed subset of 477 verified tasks from open-source repositories — a curated environment that differs from the messiness of production codebases with proprietary dependencies and undocumented assumptions.
OpenAI itself notes in the GPT-5 system card that "as with all language models, we recommend you verify GPT-5's work when the stakes are high." The hallucination reduction is meaningful progress, but it does not eliminate the fundamental need for human review of AI-generated content in consequential contexts.
The Latency Consideration
The "thinking" mode that drives GPT-5's strongest benchmark results involves extended chain-of-thought reasoning — the model generating a longer internal reasoning trace before producing a final answer. This introduces meaningful latency. OpenAI offers a minimal reasoning effort setting that minimises thinking time for applications where response speed is the priority.
For interactive applications, this trade-off between capability and latency is a genuine design consideration. For batch processing, research, and non-time-sensitive workflows, the performance gains are available without user-facing friction.
The Succession Question
GPT-5 has already been followed by GPT-5.2 (December 2025) and GPT-5.4 (early 2026), each iteration improving on specific dimensions. OpenAI has stated it has no current plans to deprecate GPT-5 or GPT-5.1 from the API. For developers and enterprises evaluating which model to standardise on, the rapid iteration cadence is itself a planning consideration — capabilities available today may be superseded within months.