Pi
Quick Facts
- Vendor
- Inflection AI (Palo Alto)
- Founded
- 2022 — Mustafa Suleyman, Karén Simonyan, Reid Hoffman
- Current state
- Post-March 2024 Microsoft acquihire; pivoted to enterprise. Model line continues as Inflection for Enterprise.
- Product
- Pi (consumer personal AI); Inflection for Enterprise (B2B API)
- Model line
- Inflection-1 (2023) · Inflection-2 · Inflection-2.5 (2024) · Inflection-3
- License
- Proprietary; hosted API only
- Hosting
- Inflection API; licensed deployments for enterprise customers
- Context window
- ~32K tokens (varies by tier); persistent cross-session memory at product layer
- Modalities
- Text and voice (real-time conversational voice is a product differentiator)
Summary
Inflection was one of the most well-funded 2022-era frontier labs, raising $1.3B by mid-2023 on the thesis that personal, emotionally-intelligent conversation was an underserved product wedge distinct from OpenAI's assistant framing or Anthropic's safety-first positioning. The flagship product — Pi — launched in 2023 with a tuned-for-warmth voice, a persistent memory of the user's ongoing context, and a real-time voice mode that was materially ahead of competitors at the time.
In March 2024 Microsoft struck a deal that moved Mustafa Suleyman and most of Inflection's research staff to lead Microsoft AI, licensed Inflection's models for Microsoft's use, and left Inflection itself as a going concern refocused on enterprise B2B. The consumer Pi product still runs, but Inflection's forward roadmap is now "Inflection for Enterprise" — licensing the model and the conversational agent stack to enterprises that want a supportive, non-technical-feeling AI layer on top of existing workflows.
For infrastructure teams, Pi's interesting properties are (1) the baseline conversational style — warmer and more naturally supportive than stock Claude or GPT without prompt-engineering effort — and (2) the product-layer memory model, where the agent remembers facts about the user across sessions without a separate RAG pipeline. These matter for workloads where the LLM faces humans, not other services.
Model Lineup
- Inflection-3 — current enterprise tier. Competitive on general benchmarks; optimized for sustained conversation rather than one-shot completion.
- Inflection-2.5 — released March 2024, the line's benchmark breakthrough. Claimed near-GPT-4-class performance on several evals at Inflection's inference efficiency targets.
- Inflection-2 — mid-2023. The jump that made Inflection a credible frontier peer.
- Inflection-1 — initial model behind consumer Pi at launch.
- Pi (product) — the consumer personal-AI product. Wraps the Inflection model line with Inflection's memory system, voice stack, and persona tuning.
Where Pi Fits
Pi is the default pick when the agent faces end users and the quality bar is feel before it's capability — customer support, wellness / coaching products, concierge assistants, consumer onboarding flows. Out of the box, Pi is less likely to produce terse, detached, or robotic output than any of the technical-first frontier models. For workloads where the human on the other side is paying per minute of attention, that baseline matters.
It is not the default pick for coding, technical analysis, long-document reasoning, or any agent loop where tool-use reliability is the bottleneck. For those, Claude, Qwen3-Coder, or Hermes remain better picks.
Tradeoffs
- Hosted-only. No self-hosting. Licensed deployments exist for large enterprises but are not a standard product.
- Post-acquihire uncertainty. Inflection's research bench moved to Microsoft. The enterprise Inflection continues, but forward model velocity is an open question vs. Anthropic / OpenAI / Google.
- Tool use. Pi is tuned for conversation, not for dispatching tool calls. Agent frameworks that depend on structured tool output work better on Claude, GPT, or Hermes Pro.
- Context window. Shorter than frontier peers. Pi compensates at the product layer with memory, but raw context for long-document work is smaller.
- Ecosystem. Third-party tooling and community integrations are a fraction of Claude / GPT / Llama. Most things require custom glue.
Deployment Notes
Within the Claw ecosystem, Pi is routed through the provider arbitrage layer as a specialist empathic tier — requests that match human-facing support, coaching, or onboarding patterns get routed to Pi; everything else goes to Claude, Qwen, or the usual providers. On a Claw Mac Mini, Pi is reachable via the Hermes agent gateway as a delegated tool: the Hermes gateway fronts all agent traffic (including traffic from Raspberry Pi terminals — see Pi × LLM), and when a turn is classified as emotion- or support-heavy, Hermes routes it to the Pi backend rather than answering directly.
This "Hermes front door, Pi specialist behind it" pattern gives deployments a single consistent gateway while still using the right model per turn. FrawdBot sits inline on the gateway regardless of which backend handles the request.
Related
- Hermes — the gateway model that fronts Pi in Claw deployments.
- Pi × LLM — the Raspberry Pi hardware recipe (same gateway, different "Pi").
- Claude — technical counterpart; typical backend for non-conversational turns.
- OpenClaw — the runtime hosting the gateway on the Mac Mini.