OpenAI has released GPT-5, and the AI community's reaction has been a mixture of awe and recalibration. The model delivers on the most important promise: a qualitative improvement in reasoning that makes it genuinely more useful for complex, multi-step tasks. But it also raises new questions about where the frontier of AI capability is heading and what it means for the companies and individuals who depend on these systems.
We've spent two weeks testing GPT-5 across a range of tasks — coding, mathematical reasoning, creative writing, research synthesis, and multimodal analysis — and spoken with researchers and developers who have been using it in production. Here is our comprehensive assessment.
Architecture: The Reasoning Leap
GPT-5 uses an improved chain-of-thought reasoning architecture that builds on the o1 and o3 series. Unlike previous GPT models, which generate responses by predicting likely next tokens, GPT-5 automatically decomposes complex problems into sub-tasks before responding. This "extended thinking" mode is applied selectively — the model determines when a problem requires deep reasoning and allocates additional compute accordingly.
The result is a model that handles multi-step problems with a consistency that previous versions lacked. In our testing, GPT-5 solved 94% of problems on the MATH benchmark — a dataset of competition mathematics problems that requires genuine mathematical reasoning, not pattern matching. GPT-4 Turbo scored 72% on the same benchmark. The improvement is not marginal; it represents a different class of capability.
The context window has expanded to 2 million tokens, enabling GPT-5 to process entire codebases, legal documents, or research corpora in a single prompt. In practice, this means you can paste an entire software project and ask GPT-5 to identify security vulnerabilities, refactor the architecture, or explain how a specific feature works — tasks that previously required careful chunking and multiple API calls.
Benchmark Results: Where GPT-5 Stands
On MMLU (Massive Multitask Language Understanding), GPT-5 scores 92.3% — a 4-point improvement over GPT-4 Turbo and slightly ahead of Google Gemini Ultra 2 (91.8%) and Anthropic Claude 3.5 Opus (91.2%). The differences at this level are small enough that real-world performance differences depend more on specific use cases than benchmark scores.
On HumanEval (code generation), GPT-5 scores 97.1% — effectively solving nearly every problem in the benchmark. This is less impressive than it sounds, because HumanEval has become saturated: the top models all score above 95%, and the benchmark no longer meaningfully differentiates them. More revealing is SWE-bench, which tests the ability to resolve real GitHub issues in production codebases. GPT-5 scores 71% on SWE-bench — a significant improvement over GPT-4 Turbo's 49% and the best result of any model to date.
On GPQA (Graduate-Level Google-Proof Q&A), which tests expert-level knowledge in biology, chemistry, and physics, GPT-5 scores 78% — above the average score of human PhD students (65%) but below the performance of domain experts (87%). This is a genuinely impressive result that suggests GPT-5 has internalized scientific knowledge at a level that exceeds most human generalists.
Multimodal Capabilities
GPT-5 is natively multimodal, processing text, images, audio, and video in a single model. The image understanding capabilities are substantially improved — GPT-5 can read and interpret complex charts, diagrams, and technical drawings with a level of accuracy that makes it genuinely useful for engineering and scientific applications.
The audio capabilities are new: GPT-5 can transcribe, translate, and reason about audio content, including identifying speakers, detecting emotional tone, and summarising long recordings. In our testing, it transcribed a 90-minute podcast with 98.7% accuracy and produced a structured summary that captured the key arguments and evidence better than our human note-taker.
Pricing and Access
GPT-5 is available through ChatGPT Plus ($20/month) and ChatGPT Pro ($200/month), with Pro users getting priority access and higher rate limits. API pricing is $15 per million input tokens and $60 per million output tokens for the standard model — approximately 3x the cost of GPT-4 Turbo. A "GPT-5 Mini" variant, optimised for cost and speed, is priced at $0.15 per million input tokens and $0.60 per million output tokens.
For most applications, GPT-5 Mini will be the right choice. It delivers approximately 85% of GPT-5's capability at 1% of the cost — a ratio that makes the full model hard to justify for anything other than the most demanding tasks. The Mini model is particularly strong on coding and structured data tasks, where the capability gap between it and the full model is smallest.
The Competitive Landscape
GPT-5's release has intensified the competition at the frontier of AI capability. Google responded within weeks with Gemini Ultra 2.5, which matches GPT-5 on most benchmarks and exceeds it on multimodal tasks involving video. Anthropic's Claude 4 Opus remains the preferred model for many developers due to its longer context window and more consistent behaviour on complex reasoning tasks.
The competitive dynamic has shifted from "which model is best" to "which model is best for my specific use case." GPT-5 is the strongest general-purpose model. Claude 4 Opus is the best for long-context reasoning and document analysis. Gemini Ultra 2.5 is the best for multimodal tasks involving video and real-time information. The era of a single dominant model appears to be over.
What GPT-5 Means for Work
The practical implications of GPT-5 for knowledge workers are significant. In our testing, GPT-5 can produce first drafts of complex documents — legal briefs, technical specifications, research reports — that require substantially less editing than outputs from previous models. It can debug complex code issues that stumped GPT-4. It can synthesise research across dozens of papers and identify genuine insights rather than just summarising.
The productivity gains are real, but they are not evenly distributed. Workers who use AI tools effectively will see dramatic productivity improvements; those who don't will fall further behind. The gap between AI-augmented and non-AI-augmented knowledge workers is widening with every model generation.