Simon Willison
CreatorLast updated: Sep 15, 2026
| Role | Independent open-source developer |
|---|---|
| Earns | Creator |
| Sources | 25 cited fields · high confidence |
GPT-6 Astra at extra-high via ChatGPT Work for long-running tasks, Claude Fable 5.1 and Codex for code, scripts and skills over MCP.
“Claude Code is, with hindsight, poorly named. It’s not purely a coding tool: it’s a tool for general computer automation.” simonwillison.net
Loadout
aisettingsShare on X
Simon Willison’s loadout
Creator
HarnessChatGPT Work
Uses ChatGPT Work for long-running tasks with GPT-6 Astra, files, web access, and generated artifacts.
ModelGPT-6 Astra
extra high
Also usesGPT-5.6 Sol
Compared directly with Astra.
ToolsLLM CLI 0.32+7 more
3 views · 0 equipped
Models
GPT-6 Astraextra high
Used at Extra High for a documentation-driven web UI and at Max for a long-running route-generation task.
Claude Fable 5.1
Used through llm-anthropic with multiple reasoning levels.
Gemini 3.7 Flash
Supported by llm-gemini and tested at multiple reasoning levels.
Gemini 3.8 Live
Google’s Live audio model used in the documented tool experiment.
Gemini 3.8 Live Extended Thinking
Named alongside Gemini 3.8 Live.
Tools
LLM CLI 0.32
Simon released LLM 0.32 with GPT-5.6 family support.
llm-anthropic
Supports Claude 5 family and Anthropic WebSearch, WebFetch, CodeExecution, and MCP tools.
llm-gemini
Plugin used for Gemini model support.
mcp-explorer
One of three MCP tools he built that week.
datasette-mcp
An MCP tool he built that week.
llm-mcp-client
An MCP tool he built that week.
visualize skill
Used by ChatGPT Work to create a route visualization.
Nominatim and Overpass
Used as the geospatial data sources for the running-route task.
Workflow
- 1Uses ChatGPT Work with Astra for long-running tasks that browse, write files, call services, and produce visual artifacts.simonwillison.net, 2026-09
- 2Points Astra at documentation at Extra High and has it build a working web UI.simonwillison.net, 2026-09
- 3Uses Codex to help build MCP tooling.simonwillison.net, 2026-07
- 4Delegate implementation to lower-power subagents; keep judgment, review, and synthesis in the main loop.simonwillison.net, 2026-07
- 5Use Sonnet for substantive implementation and Haiku for trivial edits, then review before committing.simonwillison.net, 2026-07
- 6Run long agent sessions from a phone, then switch to a laptop and review through the GitHub PR interface.simonwillison.net, 2026-07
Timeline
- 2026-09-15ModelGPT-6 Astrasimonwillison.net, 2026-09
- 2026-09-15Effortextra highsimonwillison.net, 2026-09
- 2026-09-15Also usesGemini 3.8 Livesimonwillison.net, 2026-09
- 2026-09-15Also usesGemini 3.8 Live Extended Thinkingsimonwillison.net, 2026-09
- 2026-09-12HarnessChatGPT Worksimonwillison.net, 2026-09
- 2026-09-12Toolsvisualize skillsimonwillison.net, 2026-09
- 2026-09-12ToolsNominatim and Overpasssimonwillison.net, 2026-09
- 2026-09-04Also usesGPT-5.6 Solsimonwillison.net, 2026-09
- 2026-09-04Also usesGPT-5.6 Terrasimonwillison.net, 2026-09
- 2026-09-04Also usesGPT-5.6 Lunasimonwillison.net, 2026-09
- 2026-09-01Also usesClaude Fable 5.1simonwillison.net, 2026-09
- 2026-08-30Also usesGPT-5.5simonwillison.net, 2026-08
Comments
Log in to comment