I wanted to compare how several current models approach the same tasks—not just their final answers, but the intermediate work they are willing to expose.
I ran three prompts through Kimi K3, DeepSeek V4 Pro, GPT-5.6 Sol, Gemini Flash Latest, and Claude Opus 5 in pi. Each model received the same tools, an isolated workspace, and medium reasoning effort. I also ran a second condition inspired by an external-scratchpad experiment: before answering, the model had to call a think tool with its constraints, derivation, assumptions, and verification. This makes the comparison less dependent on whether a provider happens to return its native reasoning text.
The viewer below contains both conditions. Choose a prompt at the top, filter to one model, and switch each card between its thinking trace and final answer. Think-assisted cards contain the external scratchpad; cards without that label show the provider's native exposed trace.
What the traces mean
These are recorded model outputs, not a way to recover inaccessible hidden chain-of-thought. The native cards show only what each provider exposed. The think-assisted cards show reasoning deliberately elicited through a tool call, much like asking a model to keep a structured scratchpad.
That distinction matters. A short native trace does not necessarily mean a model did less internal computation, and a long scratchpad does not necessarily mean it reasoned better. The useful comparison is how each model decomposes a task when asked to leave an inspectable record—and whether that record agrees with the work it ultimately produces.
The three prompts cover different kinds of work:
- A one-shot, multi-device todo application that requires planning and writing working files.
- A river-speed algebra problem with an exact answer and easy verification.
- An interactive visualization explaining why the sky is blue, mixing science, interface design, and implementation.
The complete raw event streams and canonical pi sessions were retained alongside the extracted traces so that tool use, timing, generated artifacts, and final responses can be inspected separately.