TokenRiver Review: Pricing & Comparison

Doing the math first: how much does ¥1=$1 actually save

TokenRiver’s single biggest selling point is its exchange rate: RMB and USD convert 1:1, while the current market forex rate is around ¥7.2/USD. Turned into real numbers:

Model	Official USD price	Paid directly at 7.2 rate	TokenRiver price	Actual savings
Claude Opus 4.8 (input)	$5/M tokens	¥36/M	¥5/M	86%
Claude Opus 4.8 (output)	$25/M tokens	¥180/M	¥25/M	86%
GPT-5.5 (input)	$5/M tokens	¥36/M	¥5/M	86%
DeepSeek V4 Flash	$0.07/M tokens	¥0.5/M	¥0.07/M	86%

For a development team calling more than 10M tokens a month, this exchange-rate gap adds up to a substantial number. It’s also the fundamental reason TokenRiver has built a reputation among price-sensitive users.

From ShiyunApi to TokenRiver

TokenRiver didn’t appear out of nowhere. In June 2026, ShiyunApi (shiyunapi.com) officially shut down, and its site now redirects straight to tokenriver.cn.

That background matters:

• TokenRiver inherited ShiyunApi’s user base and operating history — it isn’t a brand-new platform starting from zero
• Former ShiyunApi users have a direct migration path to TokenRiver, minimizing the cognitive cost of switching
• The underlying team brings accumulated experience running a domestic AI API relay

Migrating from the former ShiyunApi only requires changing the base_url, but you’ll need to apply for a new API key on TokenRiver — how the old account balance is handled should be confirmed via the official announcement.

Multi-channel failover architecture

TokenRiver claims an architecture built on automatic multi-channel failover: when one upstream channel has an outage, the system automatically switches to a backup channel, reducing service-interruption time.

For production environments running 24/7, this kind of architecture has an availability edge over a relay running on a single channel. Specific failover latency, SLA figures, and the number of channels aren’t publicly disclosed in detail and would need to be validated against real operating data.

Support for all three protocols and model coverage

650+ models are accessible through three native protocol formats:

# OpenAI format (GPT, DeepSeek, domestic models)
export OPENAI_BASE_URL="https://api.tokenriver.cn/v1"
export OPENAI_API_KEY="your key"

# Anthropic format (configure directly in Claude Code)
export ANTHROPIC_BASE_URL="https://api.tokenriver.cn"
export ANTHROPIC_API_KEY="your key"

# Gemini format (Gemini CLI)
export GEMINI_BASE_URL="https://api.tokenriver.cn"

Native support for all three protocols (not a conversion layer) means Claude-native features like Prompt Caching work normally, and Claude Code configuration needs no format adaptation.

Model coverage includes: Claude Fable 5/Opus 4.8/Sonnet 4.6, GPT-5.5/the o-series, Gemini 3.5, DeepSeek V4, Grok 4.3, Qwen3, Kimi K2.7, MiniMax, plus 650+ versioned snapshots.

Who it fits

Strong fit:

• High-volume development teams — the ¥1=$1 rate adds up to significant savings at scale
• Former ShiyunApi users — the shortest migration path, just switch the base_url
• Production environments that need multi-channel failover protection
• Day-to-day Claude Code usage (native Anthropic protocol, zero adaptation)

Worth noting:

• The ¥1=$1 rate is the core selling point, but confirm this rule is still in effect on the official site — exchange-rate policy can change in a business environment
• The 650+ model count includes versioned snapshots; defer to the official model list for the actual count of distinct model capabilities
• SLA and compensation terms aren’t published — high-compliance use cases should confirm directly with TokenRiver

If you need multimodal coverage (Midjourney/Suno/Kling) while keeping native support for all three protocols, UniAPI likewise supports the native Anthropic/OpenAI/Gemini protocols and integrates multimodal capability, giving it an edge for image/music/video generation. For transparent discounting, UiUiAPI publishes a quantified discount percentage for each model, offering another way to work out costs before topping up. If you need the 650+ model scale plus Sora 2 video generation, Shenma Relay API has a broader multimodal model matrix.

Information verified 2026-08-11. On re-check in 2026-08, the homepage now leads with “new users get 1 million free Tokens upon login” and no longer shows the “¥1=$1 exchange rate”, 650+ models, Claude/GPT/Gemini/Grok coverage, or ShiyunApi migration claims; these should be confirmed by logging into the console or contacting the vendor before drawing any conclusions.

Related reviews

• AnPin AI: 1Gbps dedicated line + multi-node routing, 99.99% uptime commitment, the operator responds in real time on X
• FlintAPI: unified API access to 43 domestic large models, one key covers every major Chinese-language model, $2 free trial credit
• UU API: MAX account pool, image generation at ¥0.04/image, ¥1 new-user bonus
• 4SAPI / Starlink 4SAPI: Enterprise-grade