blockrun-llm-vip

Genuine native passthrough for Anthropic and OpenAI through the BlockRun gateway — pay per call in USDC (x402) on Base or Solana, with zero model substitution and zero response reshaping.

Unlike a normal aggregator, these clients subclass the official anthropic and openai Python SDKs and only swap the transport (to add x402 payment) and the base URL. The gateway returns the upstream provider's response verbatim, so the official SDK parses the real signals:

• Claude: real thinking-block signature, native content[] (text / thinking / tool_use), usage.cache_creation_input_tokens / cache_read_input_tokens, native signature_delta streaming — routed to Anthropic's native /v1/messages.
• GPT: native id (chatcmpl-*), system_fingerprint, usage.*_tokens_details, honest response_format (JSON mode) + stop + nested errors. gpt-4o / gpt-4o-mini are served OpenAI-direct.

A Claude / OpenAI relay detector (e.g. cctest.ai) sees a direct upstream call.

Install

pip install blockrun-llm-vip

Use — it's a drop-in

from blockrun_llm_vip import Anthropic, OpenAI

# Claude — exactly the official anthropic SDK API
claude = Anthropic()                      # wallet auto-loaded from ~/.blockrun/.session
r = claude.messages.create(
    model="claude-opus-4.8",              # current flagship (adaptive thinking)
    max_tokens=2048,
    thinking={"type": "enabled", "budget_tokens": 1024},
    messages=[{"role": "user", "content": "What is 23*47?"}],
)
for block in r.content:
    if block.type == "thinking":
        print("signature:", block.signature)   # real Anthropic signature

# GPT — exactly the official openai SDK API
gpt = OpenAI()
r = gpt.chat.completions.create(
    model="gpt-4o",                       # served OpenAI-direct (relay-detector-proof)
    messages=[{"role": "user", "content": "hi"}],
)
print(r.system_fingerprint, r.model)        # genuine OpenAI direct

Async: from blockrun_llm_vip import AsyncAnthropic, AsyncOpenAI.

Models

You name the model; the gateway never substitutes it. Pass any current id verbatim:

• Claude (native /v1/messages): claude-opus-4.8 · claude-opus-4.7 · claude-opus-4.6 · claude-opus-4.5 · claude-sonnet-4.6 · claude-sonnet-4.5 · claude-haiku-4.5. Opus 4.7/4.8 use adaptive thinking — the standard thinking={"type": "enabled", "budget_tokens": N} is honored.
• GPT (/v1/chat/completions): gpt-5.5 · gpt-5.4 · gpt-5.4-pro · gpt-5.3 · gpt-5.2 · gpt-4.1 · gpt-4o · gpt-4o-mini, reasoning o3 / o4-mini, and more. GPT‑5.x / o-series are reasoning models — omit max_tokens / temperature / top_p (the gateway normalizes them); gpt-4o / gpt-4o-mini are served OpenAI-direct.

The full live catalog (66+ models incl. xAI Grok, DeepSeek, Llama, Mistral, Google Gemini) is at https://blockrun.ai/api/v1/models.

Solana — pay in USDC on Solana

Pass chain="solana" to any client to pay USDC on Solana (routed through sol.blockrun.ai) instead of Base. The response is identical native passthrough — only the payment leg changes (x402 SVM signing instead of EIP-712). Needs the [solana] extra:

pip install blockrun-llm-vip[solana]

from blockrun_llm_vip import Anthropic, OpenAI, Video

claude = Anthropic(chain="solana")   # bs58 key auto-loaded from ~/.blockrun/.solana-session
gpt    = OpenAI(chain="solana")
video  = Video(chain="solana")
# default stays Base:
claude_base = Anthropic()

Works on every client (Anthropic, OpenAI, Video, RealFace, VirtualPortrait, Image, Search, Exa, Audio, Voice, Phone + async). Signing needs a Solana RPC for the blockhash — it defaults to BlockRun's free proxy and is overridable via rpc_url= or the SOLANA_RPC_URL env var.

Video — Seedance, incl. real-person (RealFace)

Generate short videos through the gateway with ByteDance Seedance. Video.generate() runs the async submit→poll loop for you (x402-paid both legs by the same wallet) and returns the gateway's verbatim completed-job JSON — data[0].url is a permanent BlockRun-hosted MP4. Defaults to bytedance/seedance-2.0-fast.

from blockrun_llm_vip import Video

video = Video()
job = video.generate(
    "a neon-lit cyberpunk street, slow dolly forward",
    duration_seconds=5,
    aspect_ratio="16:9",
)
print(job["data"][0]["url"])          # permanent MP4 URL

Real people are supported. A specific, real human can appear consistently across clips — you don't upload a face per call (raw face uploads to Sora / generic image-to-video are blocked by design). Instead you enroll the person once via RealFace (one-time $0.01, ~1-min on-phone liveness for consent, no KYC), get a ta_xxxx, and pass it as real_face_asset_id on Seedance 2.0 / 2.0-fast:

Subject	Use	Liveness	KYC	Enroll
A real, specific person	RealFace	~1 min on phone (consent)	No	init → liveness → enroll
AI character / mascot	VirtualPortrait	None	No	single enroll call

Full flow, state machine, and error states: docs/real-person-flow.md.

from blockrun_llm_vip import Video, RealFace

rf = RealFace()
started = rf.init("Spokesperson — Q3 campaign")
print("Open on the rights-holder's phone:", started["h5_link"])   # QR / mobile link
rf.wait_until_active(started["group_id"])                         # after they nod + blink
asset = rf.enroll(
    name="Spokesperson — Q3 campaign",
    image_url="https://example.com/person.jpg",
    group_id=started["group_id"],
)

video = Video()
job = video.generate(
    "she smiles warmly and waves at the camera in soft studio light",
    model="bytedance/seedance-2.0",
    real_face_asset_id=asset["asset_id"],   # ta_xxxx
)
print(job["data"][0]["url"], job["payment"]["tx_hash"])

real_face_asset_id is mutually exclusive with image_url and only works on Seedance 2.0 / 2.0-fast. Other generate() options: resolution, generate_audio, seed, watermark, return_last_frame, plus timeout / poll_interval for the poll loop.

AI character (mascot / avatar, no liveness) — enroll a Virtual Portrait instead of a RealFace; same ta_xxxx → real_face_asset_id flow:

from blockrun_llm_vip import VirtualPortrait

vp = VirtualPortrait()
asset = vp.enroll(name="Mascot", image_url="https://example.com/character.jpg")  # $0.01
# pass asset["asset_id"] as real_face_asset_id on Seedance 2.0 / 2.0-fast

List what a wallet has enrolled: RealFace().list() / VirtualPortrait().list() (free). Async: AsyncVideo, AsyncRealFace, AsyncVirtualPortrait.

Image — generate & edit

Image.generate() / Image.edit() return the gateway's verbatim job (data[].url is a permanent BlockRun-hosted image). They block by default and transparently handle the gateway's hybrid flow (fast models return inline; slow ones poll). Pass wait=False to get the raw job back and Image().poll(job) yourself.

from blockrun_llm_vip import Image

img = Image()
out = img.generate("a red fox in fresh snow, soft studio light", model="openai/gpt-image-1")
print(out["data"][0]["url"])

# edit / fuse — pass a base64 data URI (or a list of up to 4); helper encodes bytes:
from blockrun_llm_vip import encode_data_uri
edited = img.edit("make it night", image=encode_data_uri(open("fox.png", "rb").read()))

Models: openai/gpt-image-1 · gpt-image-2 · google/nano-banana · nano-banana-pro · xai/grok-imagine-image · zai/cogview-4. Image().models() lists them (free).

Search — Grok Live Search & Exa

from blockrun_llm_vip import Search, Exa

# xAI Grok live search over X / web / news — grounded summary + citations
r = Search().search("latest on x402 micropayments", sources=["x", "news"], max_results=15)
print(r["summary"], r["citations"])

# Exa neural web search / similar pages / full-text extraction / grounded answer
exa = Exa()
hits = exa.search("x402 protocol", num_results=5, category="github")
text = exa.contents([h["url"] for h in hits["results"]])
ans  = exa.answer("what is the x402 payment header?")

Audio — speech, music, sound effects

from blockrun_llm_vip import Audio

audio = Audio()
speech = audio.speech("Hello there.", voice="sarah")           # ElevenLabs TTS
track  = audio.music("dreamy lo-fi beat", instrumental=True)   # MiniMax music (~2 min)
sfx    = audio.sound_effects("distant thunder over rain")      # ElevenLabs SFX
print(speech["data"][0]["url"], track["data"][0]["url"])

audio.voices()   # FREE: list TTS voices      audio.models()  # FREE: music models

Voice & Phone — AI phone calls

Lease a number, then place an AI-driven outbound call. Voice.call() blocks until the call ends and returns the transcript + recording; wait=False + Voice().poll(call_id) for control.

from blockrun_llm_vip import Phone, Voice

num = Phone().buy_number(country="US", area_code="415")   # $5, 30-day lease, wallet-bound
result = Voice().call(
    to="+14155551234",
    task="Ask if they're open Sunday, confirm hours, then thank them and end the call.",
    max_duration=3,
)
print(result["ended_by"], result["transcript"], result.get("recording_url"))

Phone also does lookup() / lookup_fraud(), list_numbers(), renew_number(), release_number(). Async twins: AsyncImage, AsyncSearch, AsyncExa, AsyncAudio, AsyncVoice, AsyncPhone — and chain="solana" works on all of them.

Wallet

The private key is used only for local signing (EIP-712 on Base, SVM on Solana) and never leaves your machine.

• Base: private_key= arg → BLOCKRUN_WALLET_KEY env → BASE_CHAIN_WALLET_KEY env → ~/.blockrun/.session.
• Solana (chain="solana"): private_key= arg (bs58) → SOLANA_WALLET_KEY env → ~/.*/solana-wallet.json → ~/.blockrun/.solana-session.

Access

Give BlockRun your wallet address to enable VIP, then pay per call from that wallet.

Contact: vicky@blockrun.ai