Pi 0.87.1: Three Frontier Models Land in One Coding Agent Release

Pi shipped version 0.87.1 on September 22, and the changelog reads like a model lineup: Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna arrive together, and Grok 4.7 becomes the default for new xAI sessions. A coding agent that connects three frontier model families in a single release has stopped being a footnote. It marks the point where the competition between the labs moved inside the agent harnesses. I run pi as the worker on my Pixel cyberdeck, orchestrated through herdr, so my question is a practical one. What does a release like this change in day-to-day agent operations, and what stays exactly as stubborn as before? The answer is more sober than the release headline, and that soberness is the point of this article.

The details are concrete. Anthropic users in pi now get Claude Opus 5.5 with adaptive thinking and a one million token context window. GPT-6 Sol and GPT-6 Luna arrive through OpenAI API keys and Codex subscriptions, and GitHub Copilot accounts gain access to all three families at once, with no additional contract required. The xAI default jumps to Grok 4.7. None of this is routine maintenance. It restructures the model layer of a harness in one release, exactly one day after 0.86.1 introduced Meta Muse as a provider. Anyone who operates agents knows this tempo: the harness stays in place while the models rotate underneath it. My whole setup is built on that assumption, and this post records where the release keeps the promise and where it does not.

The Added section of the changelog holds four lines, yet the reach is larger than the count suggests. The Claude Opus 5.5 support for Anthropic ships with adaptive thinking and the one million token context window. GPT-6 Sol and GPT-6 Luna run through OpenAI API keys and OpenAI Codex subscriptions. GitHub Copilot receives all three families in the same stroke. Read that as a distribution move: a paid Copilot seat suddenly reaches the frontier models of two additional vendors without a single new agreement. That is the core function of a harness like pi. It owns no intelligence, it routes it, and the intelligence remains interchangeable inventory while the value accumulates in the layer that governs operations.

The default switch at xAI exposes the other side of the same principle. New sessions start on Grok 4.7 instead of the previous model, and the user has to configure nothing for that to happen. That is an operational decision by the harness maintainer: pi chooses the model that most of its users will meet without any setup. I consider the move honest, because a tool that refuses to set defaults pushes integration work onto every individual user. The cost is mandatory reading. If you skip the changelog, you no longer know which model your runs actually use. This is why a weekly changelog pass belongs to my craft the same way counting words belongs to every blog article I publish.

The order of the two days shows how pi works today. I pulled the entries from the official release notes, with retrieval dates in the sources section.

None of the four fixes will appear in a marketing slide, and that is precisely why they are the most valuable section of the release. An agent harness connects dozens of providers, and every connection point has its own rules. The Fable fix shows this best: compaction summaries were refused by Claude Fable 5.1 because the split between conversation and continuation instructions was sloppy. The fix in the pull request is not a model change, it is a discipline measure, clean separation of sections and continuation-oriented instructions at the handoff. The defect lived neither in the model nor with the user, it lived in the format of the handoff between components.

The remaining three fixes follow the same pattern. Missing or invalid --mode values were silently ignored, so the agent appeared to run while working without a mode, exactly the kind of failure that wrecks overnight automation jobs because you only notice in the morning that nothing happened. Image-only user messages carried an empty text part that some OpenAI-compatible providers rejected outright. And the Anthropic OAuth flow reported an outdated client version in its headers. Four defects, four different providers, one shared lesson: in agent operations the quality of the handoff points decides whether a run survives, and that argues for reading fix lists more carefully than feature lists.

On my deck, pi is the worker harness. Orchestration runs through herdr, the worker holds a fixed model assignment, and the governing constraint is my request budget: a five hour window plus a weekly cap. Against that backdrop, 0.87.1 changes nothing about my architecture. What it changes is the menu underneath. The release puts the GPT-6 line within reach, but a model switch costs me requests, not merely attention. Every test round against a new model burns budget that producing runs need. A one million token window is not a savings machine either. A large window invites larger contexts, and larger contexts cost requests no matter how wide the window is.

The durable lesson sits in an earlier article of mine: the system stays, the model swaps. That is exactly why a release like 0.87.1 can leave my infrastructure untouched while it rearranges the model layer completely. Armin Ronacher described something similar in his pi workflow, the value lives in the harness rather than in any single model. The Hermes update line tells the same story from another angle, where patch releases carry state management and handoff fixes rather than headline features. If you bet on frontier models today, the sharper question is not which model wins the benchmark race but which harness survives a model swap without damage. By that measure the strength of 0.87.1 is that the whole model layer rebuild fits into one ordinary changelog evening.

About the author: I am Marcel and I run the HUNTER cyberdeck on a used Google Pixel 6a. I orchestrate my agent family with Hermes, pi and herdr, and I write these articles from live operation. The observations in this article come from September 2026 operation. Last reviewed on September 23, 2026.

Sources: Pi 0.87.1 release notes (September 22, 2026) and Pi 0.86.1 release notes (September 21, 2026), pi.dev, retrieved September 23, 2026. Repository earendil-works/pi on GitHub, retrieved September 23, 2026. The best-of-agent-harnesses list (ryanalberts, as of September 20, 2026), retrieved September 23, 2026. My own radar counter: GitHub API star counts for pi, as of September 10, 2026. Own measurements: Pixel cyberdeck operation with herdr orchestration, September 2026.