According to multiple enterprise users who have gained access, OpenAI has begun opening a GPT-5 preview to select partners. While GPT-5's overall capability gains over GPT-4o are substantial, the feature generating the most discussion in the industry isn't the benchmark numbers — it's a new capability called "Visible Chain of Thought."
Model Reasoning Becomes Auditable
In every major large language model to date, the model's internal reasoning process has been a black box. You see the final output, but not how the model arrived at it, what trade-offs it considered, or where it was uncertain. OpenAI's o1 series introduced chain-of-thought (CoT) reasoning, but the reasoning summaries presented to users were still compressed and filtered.
GPT-5's Visible Chain of Thought goes further, exposing more complete intermediate reasoning steps. According to early users, when handling complex compliance reviews or medical diagnostic support tasks, it's now possible to see the model progressively list its considerations, flag sources of uncertainty, and explicitly compare weights across multiple candidate conclusions — before delivering a final answer annotated with a confidence score.
This feature carries significant implications for high-stakes industries like finance, healthcare, and law. One of the most persistent concerns about enterprise AI adoption has been the inability to explain or audit model decision logic to regulators and compliance teams. Visible reasoning directly addresses this.
Benchmark Numbers
According to OpenAI's published technical report summary, GPT-5 shows clear improvements across several dimensions:
- MATH competition problems: 92.4% accuracy, surpassing o3's 88.7%
- SWE-bench (software engineering tasks): 61.3% pass rate, a new state-of-the-art record
- GPQA (graduate-level science Q&A): 76.1%, approaching human expert level
- Multilingual reasoning: approximately 12 percentage points above GPT-4o
GPT-5 vs. Claude Opus 4
Comparing GPT-5 with Claude Opus 4 — released around the same time — is one of the most active discussions in the AI community right now. Based on what's currently known, the two models appear to have distinct strengths:
GPT-5 holds an edge on pure reasoning tasks (math, logic, science Q&A), and the Visible Chain of Thought feature is a clear differentiator. Claude Opus 4 performs better on long-document processing, writing quality, and instruction-following precision, and is generally considered more conservative and consistent in its safety behavior.
For most enterprise users, the choice between the two models will come down to specific use-case requirements rather than a single performance ranking.
Caveats and Open Questions
Not all early users have offered unqualified praise. Some report that Visible Chain of Thought significantly slows response times on certain tasks, and that when reasoning chains become very long, users actually need to spend more effort validating the intermediate steps than simply reviewing the final output.
More fundamentally, several researchers have questioned whether the "visible reasoning" actually reflects the model's internal computation, or whether it's simply explanatory text generated after the fact during the output phase — a distinction that matters enormously for the auditing use case it's intended to support.
GPT-5's full release is expected to roll out to all ChatGPT Plus users and API developers within the coming months.