Pi 1.0 and Pi Durable address different uses of the same AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →-building approach. The 1.0 announcement presents Pi as a minimal, extensible agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition →, adding support for changes to prompts and tools during a conversation. It describes Pi Durable as a separate experimental package for building long-running agent applications that are not limited to the coding agent or the terminal.
The distinction explains the paired release: Pi 1.0 continues the single harness, while Durable is a separate experimental foundation for applications that need longer-running conversations. The 1.0 version number applies to Pi, while Durable retains its experimental status.
Key takeaways · AI-distilled
Pi Durable stores application state, such as a to-do list, in documents kept next to the conversation transcript, so several users or UIs can connect to, watch and steer the same agent at the same time.
Pi Durable compacts older messages in the background to stay within tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → limits without pausing the agent's active work, and lets tool and extension code be updated while the agent runs, with the next tool call picking up the new code.
Artificial Analysis, as summarized by AINews, puts GPT-6.1 Sol at $0.72 per Intelligence Index task at maximum effort, versus $1.04 for GPT-6 Sol and $3.26 for Astra, and attributes the gain to fewer turns and cheaper cache reads rather than simply fewer generated tokens.
Upstage's Solar Mini 4 (vendor-reported 35B total, 3B active parameters, 1M context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →) scores 83% on long-context reasoning but 1% on Terminal-Bench 4.0 per Artificial Analysis; about 88K output tokens per task mean a 7.1-minute average completion despite 208 tokens/s.
Tavus claims 48% of live participants mistook its Griffin video-interaction model for a human, versus under 3% for earlier systems; AINews cautions against reading this as an unrestricted Turing-test pass without the test protocol.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
