[AINews] Pi 1.0, Pi Durable, and AIE NYC

What it claims

Earendil, which Pi joined, shipped two things. Pi 1.0 is a feature release of the Pi agent harness. Pi Durable is a TypeScript port of Pi that moves all of its stateful components out of the process. The same issue includes Artificial Analysis cost and capability figures for GPT-6.1 Sol and Upstage Solar Mini 4, and vendor capability claims for FLUX 3 and Tavus Griffin.

How it works

Pi 1.0 adds:

• Codemode, with native support for MCP, Jev and image models
• Extension support for virtual models
• Deferred tool loading
• Cache warming for Anthropic models
• Mid-conversation system messages, so prompts and tools can change with awareness of the transcript
• A new TUI theme, with full-screen mode on by default

Pi Durable mechanics the summary does not spell out:

• Checkpointing. Every step is recorded as a checkpointed task. After a failure or restart, agents and subagents resume from their last exact state.
• Runtimes. It runs on any JavaScript runtime; Node, Bun and Cloudflare are named. Execution can be local or remote.
• Extensions. These bundle system prompts, tools, hooks and durable tasks. The page's example is a multi-step checkout with rollback.
• Compaction. Older messages are summarised in the background, without pausing the agent's active work.
• Hot-swapping. When tool or extension code is updated, the new code takes effect on the next tool call.
• Shared state. App state, for example a to-do list, lives in documents stored next to the transcript. Multiple users or UIs can watch and steer the same agent.

Evidence

The page gives no benchmarks or reliability numbers for Pi or Pi Durable. Its only evidence is that both reached the HN front page.

The other launches have figures, mostly from Artificial Analysis:

• GPT-6.1 Sol, cost per Intelligence Index task at maximum effort:

Artificial Analysis attributes the saving to fewer turns and cheaper cache reads, not mainly to fewer generated tokens. A separate image-encoding fix gave Luna +1 Intelligence Index point; Sol barely changed.

• Solar Mini 4:
• Size and limits: 35B total / 3B active parameters, 1M-token context, 262K max output, text-only, proprietary.
• Price: $0.10 input / $0.40 output / $0.01 cache-hit per million tokens.
• Scores: 24 overall and 83% on long-context reasoning, but 1% on Terminal-Bench 4.0.
• Speed and cost: output runs at 208 tok/s, but about 88K output tokens per task gives a 7.1-minute average completion time and roughly 5× Luna's cost per task.
• FLUX 3: native generation up to 4K, up to 10 reference images, bounding-box layout control, and multi-turn targeted editing.
• Tavus Griffin: 48% of live participants mistook it for a human, versus under 3% for earlier systems.
• Synthesia Sessions: conversational avatars for roleplay and survey interviews.

Limits

• Pi / Pi Durable: the page has no install commands, repo link, licence, API shape or storage-backend details beyond the names Memory, SQLite and JSONL. You need the project's own docs before evaluating it.
• Solar Mini 4: weights are unreleased, so the parameter counts are vendor-reported. A low per-token price does not mean a low per-task cost, because of how many tokens it generates per task.
• Gemini 4 Argon: coding quality is disputed. A Bloomberg report citing anonymous insiders is contradicted by a reported DeepMind rebuttal, and the newsletter calls it unresolved. Google's claims about how it was built come from its own developers.
• FLUX 3: "preserves every untouched pixel" is flagged as an unverified vendor claim.
• Tavus Griffin: the 48% figure comes without a published test protocol and should not be read as a general Turing-test result.
• Dates: the issue's date range ("10/01/2026–9/30/2026") appears garbled.