pi v1.6.0 went out on November 20 at 11:35 UTC with a model swap: GPT-6.2 Sol is now the default across the openai, azure, and openai-codex providers, at the same price as the GPT-6.1 Sol it replaces. For anyone paying per token, that is a good deal, and it is the kind of quiet, user-favorable move this project keeps making.
The rest of the changelog is where the story gets less comfortable: one new feature, four additions, seven changes, one removal, and forty-eight fixes. Forty-eight. In a version published eleven days after v1.5.2, which itself followed a week that produced seventeen fixes. If you extend this agent rather than just run it, that number is not a bragging right. It is a bill, and it comes due in your code.
What Actually Shipped
Strip the framing and v1.6.0 is mostly provider plumbing:
- GPT-6.2 Astra and GPT-6.2 Sol join the openai, azure and openai-codex providers, with Sol as the default on all three.
- Both new models reject the
noneandminimalreasoning efforts, so the thinking-level-off switch is simply unavailable on them. - Context windows differ by route: 1.05M tokens through Azure, 500K for Sol and 272K for Astra through the Codex provider.
qwen4-max-previewarrives on the Qwen Token Plan Individual providers withdeepseek-v4.1-flashanddeepseek-v4.1-flash-1017, and the shared Qwen plan providers are renamed to Team and Team CN under unchanged IDs.- undici moves to 8.12.4 to close six security advisories, and the llama.cpp extension default port shifts to 9931.
Forty-eight fixes is a number, not a proof of quality. Groq Qwen models were failing with 400 Unexpected message role because non-OpenAI providers sent system messages with the wrong role. Mistral Large 4 tool calls stalled on strict tools. GitHub Copilot Responses models filled omitted optional arguments with invented values, the exact failure mode that makes an agent invent a filename and then confidently edit the wrong file. Codex sessions died against a sixty-minute WebSocket limit, and more than one provider was not retrying on socket-level errors.
None of that is scandalous. Provider APIs are a swamp, and this team ships fixes faster than it ships apologies. But a forty-eight item fixed list measures how much was broken, and a project that publishes one every other week is telling you something about the maturity of the surface it hands you.
The Breaking Change Nobody Reads
The notes flag exactly one breaking change. Extension components render() now returns readonly string[], and Container returns the same cached array while its children are unchanged. Any extension that mutates render output in place, including a super.render() call inside a Container subclass, has to copy the array first.
That is a one-line edit for whoever wrote it and a hunting expedition for everyone downstream. The community map lists 133 hand-verified extensions, themes, skills and tools. How many subclass Container, how many authors will see the note, and how many will instead find out through a corrupted TUI in the middle of a long session? The same issue fixed fullscreen scrolling rebuilding the entire transcript on every wheel event, so the change is defensible on performance grounds. Defensible is not safe.
The item that never got the label is worse. The find tool path patterns are now relative to the search directory: src/*.ts finds src/a.ts but no longer matches pkg/src/a.ts, and **/src/*.ts matches at any depth. That shipped under Changed.
Which means a rendering tweak and a change to which files an agent can see now sit in the same section. Every prompt, extension and saved skill that assumed the old behavior quietly reaches a different set of files. Nothing errors, nothing warns. The agent works on a slightly wrong slice of the repository and you find out three commits later.
Velocity Is the Product, and That Is the Risk
The velocity itself is remarkable. v1.0.0 landed October 1; v1.6.0 landed November 20. Six minor versions and four patches in sixty-eight days, with v1.0.1 through v1.0.4 all appearing in the first five days of October. One contributor landed eleven fixes in under three hours on November 16.
It also sits at roughly 122.9K stars, 15.9K forks and 266 open issues, with about 25.6M npm downloads a month. Claude Code is near 65.9M and Codex near 111.1M on the same yardstick. Downloads are a crude proxy and stars measure enthusiasm more than usage, but the gap says something: many people watch pi, comparatively few depend on it in production. Fast releases are cheap while the audience is spectators. They get expensive when it is not.
The skeptical read is not that pi is doing anything wrong. It is that release discipline has not scaled with release frequency. Three fixes would close most of the gap:
- Mark behavioral changes to tool semantics as breaking, not as Changed, even when no function signature moved.
- Keep a dated migration note per release for extension authors, separate from the aggregate fixed list.
- Hold the forty-eight-fix cadence for regressions the previous release introduced, and slow down when the list is mostly provider drift.
Pi's pitch is that it is small, legible and yours to extend. That survives on one condition: upgrading it must not quietly rewrite what your agent does. v1.6.0 gives users a better default model at the same price, a real win, then hands extension authors a breaking change, a semantic change filed under the wrong heading, and eleven days to notice both.
Comments