Veteran developer and independent researcher Simon Willison has published a detailed technical explainer. Dissecting the inner workings of ChatGPT. Including how the model samples tokens, exposes logprobs to developers, and dispatches tool calls. The August 30, 2026 post on his long-running blog, titled Understanding How ChatGPT Works, is built entirely from observable API behavior rather than OpenAI's internal documentation. Making it one of the clearest practitioner-focused guides to date. This guide covers Understanding ChatGPT Work in detail. This guide covers Understanding ChatGPT Work in detail.
Understanding ChatGPT Work: What did Willison actually discover about ChatGPT's internals?
Willison's methodology is unusual: rather than relying on leaked specifications or press releases. He reverse-engineered ChatGPT by deliberately confusing the. Model and inspecting the structured outputs. According to the post, when a user uploads a file the system cannot. Read, ChatGPT returns a structured 'tools' function-call block naming the offending attachment. Which Willison argues demonstrates that file analysis is implemented as a discrete tool. Rather than baked into the base model. He also documented that ChatGPT surfaces token-level logprobs (logarithmic probabilities) through its Responses. API, allowing developers to score alternative completions programmatically.
In the same post, Willison detailed the 'work' tool pattern, noting that ChatGPT invokes internal tools such as 'file_search', 'web_search', 'code_interpreter', and 'image_generation' via JSON function-call envelopes identical to those used by third-party custom GPTs. This convergence, he wrote, suggests OpenAI treats first-party and external tools as architecturally interchangeable. He also highlighted how sampling temperature, top_p, and presence penalties are passed as named parameters to the same underlying inference endpoint, an implementation detail OpenAI's official docs only describe abstractly.
How do token sampling and logprobs shape real ChatGPT outputs?
Willison explained that ChatGPT does not generate the single most probable next word. Instead, it scores every token in its vocabulary and samples one according to the configured temperature and top_p values. Low temperature concentrates probability mass on a handful of high-likelihood completions, producing the dry factual summaries ChatGPT is known for; raising temperature flattens the distribution and produces the divergent creative writing users report. He demonstrated this by querying identical prompts at temperature 0 and temperature 1.2 and comparing the resulting logprobs arrays returned by the API.
Logprobs, the post notes, are the developer escape hatch most users never see. By requesting the top_logprobs parameter, integrators receive the raw probability score of each candidate token, enabling tasks such as confidence-weighted retrieval, hallucination detection, and grammar-aware reranking. Willison framed this as the most underrated capability in the Responses API, arguing that any production system serious about reliability should log and monitor logprob drift the same way it monitors latency.
What This Means for Developers and AI Builders
For engineering teams building on ChatGPT, Willison's analysis carries concrete recommendations. First, treat every ChatGPT capability as a tool call rather than assuming monolithic intelligence; this affects how you design prompts and how you handle refusals. Second, instrument logprob distributions to detect silent regressions when OpenAI swaps underlying models. Third, when reasoning over long contexts, use the file_search tool rather than pasting full documents, because file_search carries its own retrieval step and avoids context-window degradation.
The post implicitly challenges a common misconception:. That ChatGPT's web search and code execution are 'built into' the chat model. They are not. They are discrete tools called through the same function-call interface any developer can target. Which means any team can replicate a slice of ChatGPT's apparent intelligence using. The Responses API plus a retrieval backend. And a sandboxed Python interpreter. The architectural transparency is a competitive advantage for builders willing to do the integration work.
What's Next: Why Observable AI Architecture Matters
Willison's approach signals a maturing practice: treating frontier AI as an observable system. Rather than an opaque oracle. Expect more independent researchers to publish reproducible probes of Claude, Gemini. And Llama endpoints using the same methodology, especially as providers continue to ship. Logprob and tool-call transparency. The competitive differentiator for 2026 is no longer raw model size. It is the quality of the developer ergonomics wrapped around it.
Key Points
Simon Willison published 'Understanding How ChatGPT Works' on August 30, 2026, re Whether you are new to Understanding ChatGPT Work or already experienced, the sections below have you covered.verse-engineering ChatGPT. From observable API behavior.
ChatGPT invokes built-in capabilities such as web_search and code_interpreter via the same JSON tool-call interface used by third-party GPTs.
The Responses API exposes per-token logprobs, enabling confidence scoring, hallucination detection, and production monitoring.
The Bottom Line
ChatGPT is less a single model and more a router in front of a tool registry. And developers who instrument logprobs and tool traces will build more reliable systems. Than those who treat the chat box as a black box.
Frequently Asked Questions
What is the 'work' tool in ChatGPT?
According to Simon Willison's August 2026 post, 'work' is the internal function-call envelope ChatGPT uses to dispatch first-party tools such as web_search, file_search, and code_interpreter, architecturally identical to third-party tool calls.
How can developers access ChatGPT logprobs?
Developers request the top_logprobs parameter on Responses API completions to receive raw probability scores for each candidate token, useful for confidence scoring and hallucination detection.
Does ChatGPT use temperature to pick the next token?
Yes. Willison demonstrated that ChatGPT samples tokens from a probability distribution reshaped by temperature and top_p, with low temperature favoring high-likelihood completions and high temperature producing more varied output.
Related Resources
For more context, check our related article on OpenAI's ChatGPT Poised to Introduce PIN and Fingerprint Locked Chats for Enhanced Privacy.
