pi-gpt-fast-mode

A pi extension that controls OpenAI's service tier for GPT-5.4 / GPT-5.5 from inside pi.

Unlike a simple on/off "fast mode", this lets you pick the tier:

• priority — faster, premium (the classic "fast mode")
• flex — cheaper, slower (economy)
• default — standard
• auto — let OpenAI choose

It shows the active tier in the TUI and hands the preference off to subagents automatically.

npm · Pi Package Gallery

Install

pi install npm:pi-gpt-fast-mode

Or install from git:

pi install git:github.com/devwithpug/pi-gpt-fast-mode

Try it for a single run without installing:

pi -e npm:pi-gpt-fast-mode

Usage

/fast            # toggle Fast Mode (uses the selected tier, default: priority)
/fast on         # enable
/fast off        # disable
/fast priority   # select the fast/premium tier and enable
/fast flex       # select the economy tier and enable
/fast default    # standard tier and enable
/fast auto       # let OpenAI choose, and enable
/fast status     # report current state and tier

Start pi with Fast Mode already requested (priority tier):

pi --fast

How it behaves

The extension separates what you want from what is applied:

• desired — you asked for Fast Mode (/fast, --fast, or persisted config)
• active — desired and the current model is supported → requests get service_tier: <tier>

Switching to an unsupported model temporarily stops the request without losing your preference or your selected tier. Switch back to a GPT-5.4/5.5 model and it resumes.

Subagents

Turn Fast Mode on once in your parent pi session, then launch subagents as usual. The preference is exported as the PI_GPT_FAST_MODE environment variable (1 / 0), which child pi processes inherit on startup. A subagent only sends the tier request when it is also on a supported model.

Parent on an unsupported model is fine. If your parent session runs a non-GPT model (e.g. Claude) and you only use GPT-5.4/5.5 in subagents, just run /fast in the parent anyway. The preference is still recorded and handed off — the parent simply won't request a tier for itself, but subagents on GPT-5.4/5.5 will. The TUI indicator shows fast⇢ in that "armed" state to signal the preference is passed downstream.

To verify, ask a subagent to print PI_GPT_FAST_MODE; 1 means the preference was handed off.

Supported models

By default:

[
  "openai/gpt-5.4",
  "openai/gpt-5.5",
  "openai-codex/gpt-5.4",
  "openai-codex/gpt-5.5"
]

Model keys must match pi's exact provider/model key.

Configuration

Config files are resolved in this order (first that exists wins):

1. Project: .pi/extensions/pi-gpt-fast-mode.json
2. User: ~/.pi/agent/extensions/pi-gpt-fast-mode/config.json

{
  "persist": false,
  "desired": false,
  "tier": "priority",
  "models": [
    "openai/gpt-5.4",
    "openai/gpt-5.5",
    "openai-codex/gpt-5.4",
    "openai-codex/gpt-5.5"
  ],
  "indicator": "status"
}

Field	Default	Meaning
persist	false	Remember the on/off choice and tier between pi runs.
desired	false	Saved preference (only honored when persist is true).
tier	"priority"	Service tier: priority, flex, default, or auto.
models	see above	Exact provider/model keys allowed to use Fast Mode.
indicator	"status"	TUI feedback: status, widget, or off.

Fast Mode starts disabled and session-only by default. Set "persist": true to remember your choice between runs.

How it differs from similar extensions

The core mechanism — injecting service_tier via before_provider_request — is necessarily the same as other OpenAI fast-mode extensions. This one adds:

• Multi-tier control (priority / flex / default / auto) instead of a single hardcoded priority toggle, so it doubles as an economy (flex) switch.
• A /fast status subcommand and a tier-aware TUI indicator.

Troubleshooting

I turned Fast Mode on but I don't see the indicator. Check that your current model is in models, and that indicator is not off.

My choice isn't remembered after restart. Set "persist": true in your config.

Does this guarantee faster responses? No. It requests a service tier when possible. Actual latency, availability, and billing depend on OpenAI and your account. Not every model/account supports every tier; OpenAI rejects unsupported tiers.

Development

npm install
npm run check   # typecheck + tests

License

MIT © devwithpug

Inspired by pi-openai-fast-mode and pi-openai-fast.