Raspberry Pi 5 in 2026: the news, and where a coding agent actually belongs

Four things, all verifiable on the official Raspberry Pi news page. First, memory pricing: a global LPDDR4 crunch driven by AI infrastructure demand pushed the 16GB Pi 5 to about $305, up from its $120 launch price in March 2025, with the 8GB around $175 and the 4GB around $110 as of April 2026. Second, a new budget 1GB Pi 5 launched at $45. Third, the AI HAT+ 2 shipped on January 15, 2026, at $130, pairing a Hailo-10H accelerator (40 TOPS INT4, 8GB onboard RAM) so the Pi 5 can run small generative-AI models locally. Fourth, the Raspberry Pi 6 is not expected before 2028.

The Raspberry Pi Foundation attributes it to an unprecedented rise in the cost of LPDDR4 memory, driven by competition for memory fab capacity from the AI infrastructure roll-out. Memory is the single largest component cost on the higher-capacity boards, so the 8GB and 16GB models took the steepest increases while the 1GB variant stayed at $45. The Foundation has called the situation painful but temporary and said it expects to unwind the increases once memory supply loosens.

Yes, with caveats. The AI HAT+ 2 (Hailo-10H, 40 TOPS INT4, 8GB onboard RAM) runs a defined set of small models at launch: DeepSeek-R1-Distill 1.5B, Llama 3.2 1B, and Qwen2.5-Coder, Qwen2.5-Instruct, and Qwen2 at 1.5B. That is genuine on-device generative AI, and it is great for classification, lightweight chat, and edge inference. It is not a model class that can drive a serious agentic coding loop, where you want frontier-grade reasoning over a large live context. For that, the practical answer in 2026 is still a hosted Sonnet- or Opus-class model.

Run the heavy agent on the Mac you already own, and use the Pi as the always-on Linux node. Claude Code's CLI is Node-based and will run on a Pi over SSH, but the Pi's value is being a cheap, low-power, always-on box for small local models, MCP servers, cron jobs, and self-hosted services. The reasoning-heavy coding work calls a hosted model regardless of which machine runs the CLI, so the deciding factor is the surface you work in. fazm gives the Mac side a native UI with persistent sessions, one-click forking, and reach beyond the terminal into the browser and native apps.

No. fazm is a native macOS app and requires macOS 14 or newer. It does not run on Raspberry Pi OS or any Linux. That is deliberate: fazm uses macOS accessibility APIs and screen context to reach beyond the terminal into native Mac apps and the browser, which has no equivalent on a headless Pi. The honest split is a Pi 5 as your always-on Linux node and a Mac running fazm as the agent surface. fazm wraps the real Claude Code loop via the claude-agent-acp adapter (0.29.2) and Codex via codex-acp (0.12.0), both pinned in acp-bridge/package.json.

Not before 2028, with early 2028 the absolute earliest. CEO Eben Upton has described a 4 to 4.5-year cadence from the Pi 5 launch and said the Pi 6 will bring quantitative changes, not qualitative ones, meaning the same overall form factor and ports rather than a redesign. He also noted the Pi 5 is still a capable flagship that can comfortably hold its position in the meantime.

It depends on what you want from it. At $130 it is a clean way to run 1B to 1.5B models on-device for edge use cases: a local classifier, a privacy-bound assistant, a vision-plus-language pipeline that never leaves the Pi. If your goal is to offload agentic coding from a laptop, it will disappoint, because that workload needs a much larger model and a long live context. Think of the AI HAT+ 2 as an edge-inference accelerator, not a replacement for a hosted coding model.