Pay-2-Go: `ipx p2g`#

ipx p2g uses the Pay-2-Go API (/v1): each request is metered against your credit wallet. It is the production-oriented API and the one that lets an organization pay (set a context with ipx auth context, see Authentication and Organization Context).

chat, text, image, video, audio and models work exactly as described in Lightning: ipx gen (replace gen with p2g); after each generation the credits charged are shown. This page covers what is specific to Pay-2-Go: 3D, embeddings, realtime voice, wallet, keys and deployments.

$ ipx p2g chat                    # per-token billing, footer shows credits per answer
$ ipx p2g image "a red fox" -o fox.jpg
$ ipx --json p2g text "hi" | jq .usage.credits

Wallet and pricing#

$ ipx p2g balance          # balance, purchased/spent and usage by modality
$ ipx p2g ledger --limit 50
$ ipx p2g pricing          # live credit prices and top-up packs (no sign-in needed)

When an organization context is set, requests are billed to the organization’s wallet. Use --personal to bill your own wallet for one command. Top up credits in the console at https://console.inferenceport.ai.

3D generation#

$ ipx p2g 3d --image chair.jpg -m tripoSR -o chair.glb
$ ipx p2g 3d --image logo.png -m trellis2 --resolution high
$ ipx p2g 3d --image obj.png -m asset-harvester --prompt "a wooden chair"
$ ipx p2g 3d --image chair.jpg --no-wait          # print the job id
$ ipx p2g 3d --image chair.jpg --job a3f1c2e4b5   # fetch the result of an earlier job

Models: tripoSR (default), sf3d, trellis2 (with --resolution low|medium|high) and asset-harvester (needs --prompt; produces a Gaussian-splat .ply and an orbit .mp4). The job runs asynchronously and credits are refunded automatically if it fails. This feature is in early access; see Pay-2-Go (P2G) API.

Embeddings#

$ ipx p2g embed "The quick brown fox"                           # one input, shows the vector summary
$ ipx p2g embed "first" "second" -m text-embedding-3-large --dimensions 256
$ ipx p2g embed -f notes.txt --lines -o vectors.json            # one input per non-empty line; full response saved
$ cat docs.txt | ipx p2g embed --lines --json | jq '.data[0].embedding | length'
$ ipx p2g models --type embedding                               # models and prices

Inputs can be arguments, files (-f, repeatable; a whole file is one input unless --lines), or standard input. The default model is text-embedding-3-small. The API takes at most 32 inputs per request, so larger sets (up to 1,024) are sent in batches of 32 and merged; the token and credit totals are summed. --dimensions accepts 128-4096 (model limits apply), --input-type query|document is for Cohere and --task-type for Gemini. You are billed per input token and the credits charged are printed after the result; --json prints one JSON document (the merged response, including usage.payg_credits_charged). See Embeddings API.

Realtime#

$ ipx p2g realtime chat "Tell me a one-line joke"                   # streams the reply as text
$ ipx p2g realtime chat "Say hello" --voice marin --audio-out hello.wav
$ ipx p2g realtime chat --audio-in question.wav                     # ask by speaking a WAV file
$ ipx p2g realtime chat                                             # interactive: one message per line, Ctrl+D to end
$ ipx p2g realtime transcribe meeting.wav
$ ipx p2g realtime transcribe talk.wav -m scribe-v2-realtime -o talk.txt
$ arecord -f S16_LE -r 24000 -c 1 -t raw | ipx p2g realtime transcribe - --rate 24000
$ ipx p2g models --type realtime

realtime chat talks to a conversational model (default gpt-realtime-2.1). Replies are text only unless you pass --audio-out FILE (then the model speaks, the transcript is still printed, and the audio is saved as a WAV, even if the session fails part-way), so a plain text chat is never billed for audio. realtime transcribe streams a recording (a WAV file, or raw PCM16 on standard input with - and --rate) to a transcription model (default gpt-live-transcribe) and prints the transcript; a conversational model is refused with a pointer to chat, and the other way round. Audio is converted to 24 kHz mono PCM16 while it streams (nothing is held in memory, and a live arecord pipe is sent as it is recorded). ipx has no microphone capture of its own; pipe from arecord, sox or ffmpeg.

Sessions use the model’s default from GET /v1/realtime/models (the GA Realtime shape). Models with manual commits (mai-transcribe-2-streaming, scribe-v2-realtime) are committed when the audio ends; models with server-side turn detection (gpt-live-transcribe) get 1.5 seconds of silence at the end so the last sentence is closed. The command then waits for the transcript itself (--wait, default 60 s) and until the server has been quiet for --idle seconds. An error event from the server, or a connection that drops before the end, fails the command with a clear message and a non-zero exit status.

Useful options: --pace sends audio at real-time speed, --session '{...}' replaces the default session.update payload, and --events prints every raw event to standard error. With --json, standard output is exactly one JSON document (text, usage and audio_file for chat; text, segments and audio_seconds for transcribe). The WebSocket uses the same base URL as every other command (path prefixes are kept) and is always encrypted except for localhost. Sessions are billed while they run (per token for gpt-realtime-*, per streamed second for the transcription models); ipx p2g ledger shows the exact charges, and a session ends with a clear message and hint when credits run out or an organization policy blocks the model. See Realtime API.

API keys#

$ ipx p2g keys list
$ ipx p2g keys create ci-bot --expires 2027-01-01      # the secret is shown once
$ ipx p2g keys revoke <key-id>

Personal keys (ipa_live_...) are used with Authorization: Bearer or ipx auth login --api-key. Organization keys, with spend caps and restrictions, are managed with ipx org keys (Organizations: ipx org).

Deployments#

Personal Featherless deployments billed through your wallet. For organization deployments and the full command set (list, restore, terminate, cost and state checks) use Featherless deployments: ipx featherless:

$ ipx p2g deploy models                 # what can be deployed
$ ipx p2g deploy status <model-id>      # availability
$ ipx p2g deploy create <model-id>
$ ipx p2g deploy list
$ ipx p2g deploy get <deployment-id>
$ ipx p2g deploy usage <deployment-id>
$ ipx p2g deploy cancel <deployment-id>   # and: uncancel, delete

Errors you may see#

402 Insufficient credits: The wallet is below the minimum for the operation. Check ipx p2g balance and top up.
429: Rate limit reached (120 requests per minute per identity by default).

Decision models#

Evaluate typed choice, noul and score questions through the System One endpoint:

$ ipx p2g models --type decision
$ ipx p2g prices --type decision
$ ipx p2g decision --file request.json --output result.json
$ ipx p2g decision 'Please refund my payment' --questions questions.json --json
$ cat request.json | ipx p2g decision --file - --json

request.json contains model (default microsoft/decision-1), state (text or JSON) and named questions. questions.json contains just the named questions object. --model overrides a request file’s model when explicitly supplied. Responses include provider answers, input tokens and credits charged. Normal authentication and organization billing context apply. See Decision models / System One API for complete examples and provider configuration.

Pay-2-Go: ipx p2g#