Price terms as published

Each line is the vendor’s own wording, split from the price capture checked on 2026-09-07. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.

2 price terms for AMD Radeon Cloud · Token Factory
#Plan / termChecked
1Free tier with daily account spend cap2026-09-07
2"$10/day" figure is tester-reported, not in official docs2026-09-07

One-liner: AMD’s Radeon Cloud “Token Factory” (developer.amd.com.cn) hands every registered account an auto-generated API key for a small catalog of open models, served through an OpenAI-compatible endpoint. It is a free channel, funded by AMD’s push to build developer loyalty to its GPUs, and it works from mainland China without a VPN.

Verified facts (checked 2026-09-07)

Attribute Value Source
Official docs amd-aim.github.io/radeon-cloud-docs checked same day
Entry point developer.amd.com.cn/radeon/modelapis official docs
API key Auto-issued on first visit (rc- + 48 hex), visible in Profile official docs (authentication), same day
Login methods Phone/email code, or ModelScope / CSDN / GitHub third-party official docs (login guide)
Rate limits 30 RPM per key · 120 RPM per IP · 8 concurrent per key official docs (rate-limits), checked same day
Daily cap Per-account daily spend cap, resets midnight Asia/Shanghai official docs, same wording
Endpoints OpenAI-compatible /v1/chat/completions and Anthropic-compatible /v1/messages (Claude Code named in docs) official docs
Error shape 429 with OpenAI-style body, token_rate_limit_exceeded code, Retry-After header official docs

Independent test record (2026-09-06, registered account)

A collaborator registered and drove the API directly. Findings, kept separate from official claims:

  • /v1/models returned four live models: DeepSeek-V4-Flash, DeepSeek-V4-Flash-Vision-Exp, Qwen3.8-Flash-Next, MiniCPM5-1B
  • Chat, vision, and the Anthropic /v1/messages endpoint all returned 200 (the vision model correctly identified a color test)
  • Cold-start first token took 23 seconds; steady-state responses 0.3–1.1 s
  • Rapid-fire requests hit 429 token_rate_limit_exceeded quickly — real usable throughput is well below the nominal 30 RPM, consistent with the account-level daily cap being reachable by small requests

The widely circulated claim of a “$10 free spend per day” number does not appear in official docs — the docs only say a per-account daily cap exists. Treat $10 as one tester’s measurement, not AMD’s published terms.

Gotchas worth knowing before you build on it

  1. Silent parameter dropping. The free gateway whitelists parameters; unsupported ones are dropped without error. Your temperature or tool-call config may simply not apply. (official docs — non-whitelisted params silently ignored)
  2. Chat and Anthropic endpoints only. embeddings, completions, image, audio, responses all 404 on the free endpoints.
  3. Small model catalog. Four models, subject to change; this is a showcase, not a marketplace.
  4. Shared pool. Platform-wide concurrency is a shared resource — global_concurrency_rate_limit_exceeded is a documented failure mode. AMD’s own answer for needing more is a paid dedicated endpoint.

Why this exists (the durable part)

GPU vendors giving away inference is a CUDA-moat counterplay: AMD binds developers to Radeon Cloud the way NVIDIA binds them to CUDA. For users in mainland China specifically, the China-domain entry with ModelScope/CSDN login makes it one of the few big-company, directly-reachable free API channels — no reseller middlemen, no card.

Compare with Keenable (search layer, free 100K req/mo), Cline API (tool-vendor gateway), Agnes AI (model-lab funnel) and ModelScope API-Inference (community-platform pool, 2,000 calls/day) — four different 2026 free-tier shapes: hardware vendor, search infra vendor, model lab, platform community, each giving away a different layer.

Sources