Amp released major updates in September, introducing Fable 5.1 for ultra mode, Claude Opus 5.5 as the new default medium model, and significant improvements to runner infrastructure and customization options.
Models and Inference
- Fable 5.1 in ultra mode: Delivers better long-horizon work, mid-task steering for multiple instructions, and 35% cheaper ultra threads. Excels at writing prose, proving work with videos and screenshots, and iterating autonomously for hours.
- Claude Opus 5.5 as default medium mode: Replaces GPT-5.6 Sol, solving 65% of tasks at 10% lower cost. Responds to mid-work corrections without restarting, with high reasoning effort built in.
- Plaid speed for GPT-6 Astra: Enables inference up to 6× faster at 6× cost per token via OpenAI's ultrafast tier. Toggle in the new thread dialog or existing threads.
Agent Control and Workflow
- Immediate user messages during execution: Feedback now takes effect sooner instead of queuing, avoiding redundant verification steps. Built-in actions like Ship and Review still queue as before.
- Collapsed agent work steps by default: Intermediate details are hidden while keeping results visible, letting agents handle longer tasks independently. Expand steps when needed.
- Restack button: Reorganizes all changes in a thread into cleaner commits, letting you split large diffs, tidy intermediate commits, or combine small related ones while preserving final contents.
- Desktop tab: See and control a high-resolution interactive Linux desktop directly in any orb thread, useful for verifying agent work on Linux, Android, Chrome extension APIs, and file formats requiring desktop applications.
Customization and Modes
- Customize Your Dial: Select models behind built-in modes and add your own agents. Tune Your Modes lets you set model and reasoning effort using your own API key or ChatGPT subscription. Build Your Own Dial combines plugin agents with modes, drag 2-4 modes into order, and save for personal or shared use.
- Intelligently ordered diffs: Highlight essential files first and dim tests and generated code, helping you review changes faster by reading what matters most.
Runners and Infrastructure
- Shared Runners: Start a runner with --share so your entire workspace can spawn agents on your machine from ampcode.com, useful for GPUs, specialized hardware, or network-specific dev boxes. The macOS app now runs as a runner automatically. A "Keep This Mac Awake" option prevents sleep while the runner is active and plugged in.
- Multiple Directory Serving: A single runner can serve specific paths or auto-discover Git repositories within two levels. Directories can be added or removed while the runner is running, and changes persist for the next start. Self-updating runners check hourly and restart with the new version, preserving ID, flags, and served directories.
- Git worktrees and secrets: Runners can create fresh checkouts by picking a repository and naming a branch without touching main code. Access Secrets & Env Vars configured on ampcode.com, applied per thread with the --amp-env flag. Threads in runner-created worktrees get Archive and Remove Worktree actions.
Pricing and Access
- Free Hobby and Teams tiers: Hobby tier lets users run agents on their own hardware or Amp's remote orbs without subscription or BYOK token fees. Teams tier allows unlimited members, shared threads, portals, multiplayer sessions, and free SAML/OIDC SSO when any member is on a paid tier. Zero-data-retention policy now applies to every user and workspace.
- Expanded BYOK options: All users can route to many providers including OpenRouter, Amazon Bedrock, Google Cloud Agent Platform, Azure Foundry, Vercel AI Gateway, Cloudflare AI Gateway, Ollama Cloud, OpenCode Go, or custom URLs.