Internal benchmarks allegedly show GPT-5 achieving near-human performance on graduate-level STEM tasks, with a 92% pass rate on the GPQA Diamond benchmark — a 34-point jump over GPT-4o.
92%
GPQA Diamond Score
+34pts vs GPT-4o
2M tokens
Context Window
4× larger than GPT-4o
3.2×
Reasoning Speed
faster than o1 model
12/15
Benchmark Lead
tasks beat all rivals
A document allegedly originating from an OpenAI contractor review process began circulating in AI research circles on March 11, 2026. The document — which OpenAI has not confirmed or denied — contains internal evaluation results showing GPT-5 achieving what appears to be a historic jump in reasoning capability. The most striking figure is the 92% pass rate on the GPQA Diamond dataset, a benchmark specifically designed to be difficult enough that even PhD-level experts in the relevant field fail roughly 30% of questions. GPT-4o currently scores 58% on this dataset. If verified, GPT-5's score would represent a 34-percentage-point leap — the largest single-generation improvement OpenAI has ever posted on a third-party benchmark.
“These numbers, if real, represent a genuine phase transition. We're not talking about incremental improvement — a 34-point jump on GPQA Diamond means GPT-5 is outperforming the average PhD in their own domain. That changes everything about how we think about human oversight.”
Dr. Elias Vance
ML Research Lead, Stanford HAI
Multiple sources close to OpenAI's research division describe GPT-5 as adopting what insiders call 'long-chain reasoning' (LCR) — a fundamentally different inference strategy compared to the transformer-only approach of prior models. Unlike GPT-4o, which generates a response in a single forward pass, LCR models generate reasoning chains, self-critique those chains, then synthesize a final answer. This is analogous to how a physicist might write down an attempt at solving a problem, check their algebra, notice errors, then revise before committing to an answer. The approach was hinted at in OpenAI's o1 and o3 models, but GPT-5 reportedly takes it to a far greater degree, with up to 128 reasoning steps on complex tasks.
Long-chain reasoning is computationally heavier than standard inference. Expect GPT-5 API pricing to be significantly higher for tasks that trigger deep reasoning chains, with a "quick mode" option for simpler queries.
86% HumanEval, 58% GPQA — multimodal via plugins
Omni model: real-time audio+vision, 58% GPQA Diamond
Reasoning model, 78% GPQA Diamond, slower inference
87.5% ARC-AGI, new SOTA across reasoning benchmarks
92% GPQA Diamond (leaked), LCR architecture, 2M context
GPT-5's rumored capabilities will have significant downstream effects on every AI tool that uses OpenAI's API. Tools built on GPT-4o — including many coding assistants, writing platforms, and chatbots — would receive dramatic performance boosts simply by switching API endpoints. However, the cost structure may force many freemium tools to gate the more powerful reasoning features behind paid tiers. Competing labs are already feeling the pressure: multiple Anthropic and Google employees have posted publicly that their teams are on 'high-alert deployment schedules,' suggesting Claude 4 and Gemini 2 Ultra launches may be pulled forward to avoid being caught flat-footed.
OpenAI has not officially confirmed these benchmarks. The document provenance has not been independently verified. Past model leaks have contained both accurate and fabricated metrics. We will update this article when official information is available.
James Whitfield
Senior AI Researcher · AIToolsHub
Covering artificial intelligence trends, product launches, and market analysis for AIToolsHub. Focused on making AI developments accessible and actionable for builders, buyers, and business leaders.
ChatGPT
OpenAI's flagship consumer AI chat product
Claude
Anthropic's competing reasoning-focused model
Gemini
Google's multimodal AI challenger
Perplexity AI
AI-powered search with real-time citations
Adoption momentum score.
Top 5 AI stories every Monday. No noise, just signal.
Internal benchmarks allegedly show GPT-5 achieving near-human performance on graduate-level STEM tasks, with a 92% pass rate on the GPQA Diamond benchmark — a 34-point jump over GPT-4o.
The Claude maker has become the world's second most valuable AI company, overtaking Mistral and Cohere. The capital will fund Claude 4 development and expanded cloud infrastructure.
The open-source release includes three model sizes and a new mixture-of-experts architecture. Developers are already fine-tuning it for specialized use cases within hours of release.