Once authenticity checks are live, this shows the distribution distance from the official API, capability-question pass rates side by side, and usage count drift. Routes that fail are greyed out on every leaderboard, with the reason shown in red on hover.

ⓘ Method: see the blog post “Did the model really get dumber?” Read

Speed

Measured by time to first token (seconds) and output speed (tokens / second)

Time to first token

Seconds until the first token · lower is better

1 relays

ⓘ Output speed: not yet available from the live API; shown once connected

Price

Cost per coding session

Gemini 3.8 Flash · 2M input and 40k output tokens (¥) · split by token type · lower is better

3 relays

OutputInput

ⓘ USD prices are converted at 1 USD = 6.7698 CNY

Prices: cache hit, input and output

¥ / 1M tokens

Cache hitInputOutput

Cache

Coding requests repeat much of their context, so the cache hit rate largely decides what you actually pay. Drag the hit rate and the costs and ranking below are reweighted with cache prices.

Cache discount

Not provided

Cache hit — · input ¥1.275 (per 1M tokens)

Measured hit rate

Pending

Measured by replaying multi-turn sessions: cached-read tokens in usage as a share of all input

Session cost at 80% hit rate

¥2.80

¥2.80 without cache, 0% saved

Cache-weighted session cost

Gemini 3.8 Flash · of 2M input tokens, the hit share is billed at the cache price and the rest at the input price, plus 40k output (¥) · lower is better

Rank 1 / 3

Cache hit rate80%Coding agents typically see 70%–90%

Cached inputUncached inputOutput

ⓘ Routes without a cache price are billed at the input price; cache write fees are not included

Stability

Availability

Request success rate, last 24 hours · higher is better

2 relays

Status of each model, last 24 hours

One cell per hour · green ok, yellow delayed, red failed

Model	Group	Availability	Last 24 hours	First token	Samples
DeepSeek V4.1 Flash	国模分组稳定版本官方2.5折	—	None yet	—	—
GPT 6 Astra	GPT-Pro++ 独享满血不降智	—		—	0
Kimi K3	国模分组稳定版本官方2.5折	—	None yet	—	—
Claude Sonnet 5	Claude-Kiro-D 企业正价号智商拉满不掺降智号	—	None yet	—	—
Qwen 3.8 Max	国模分组稳定版本官方2.5折	—	None yet	—	—
DeepSeek V4 Pro	国模分组稳定版本官方2.5折	—	None yet	—	—
Claude Opus 5	Claude-Kiro-D 企业正价号智商拉满不掺降智号	—	None yet	—	—
GPT 5.6 Sol	GPT-Pro++ 独享满血不降智	—	None yet	—	—
Claude Fable 5	Claude-Kiro-D 企业正价号智商拉满不掺降智号	—	None yet	—	—
GPT 5.6 Terra	GPT-Azure官key	—	None yet	—	—
Grok 4.6	Grok-heavy 高智商满血版	—	None yet	—	—
Gemini 3.8 Flash	Gemini-vertex 官转	—	None yet	—	—
Gemini 3.1 Pro Preview	Gemini-vertex 官转	—	None yet	—	—
GLM 5.3 Flash	国模分组稳定版本官方2.5折	—	None yet	—	—

Service details

Minimum top-upNot provided

Sign-up bonusNot provided

Support responseNot provided

Payment methodsOther

China invoiceNot verified

International invoiceNot verified

RefundsNot verified

PromotionsNot provided

FAQ

How much does Gemini 3.8 Flash cost on aruhub API?

Input ¥1.275 and output ¥6.375 per 1M tokens; at 2M input and 40k output tokens, a coding session costs about ¥2.80, an API rate of 0.25.

How does the cache hit rate affect price?

aruhub API has not published a cache price for this route, so cache hits are billed at the input price.Measured hit rates will replace this assumption once available.

Which models does aruhub API support?

Folkbench currently lists 14: GPT 6 Astra, DeepSeek V4.1 Flash, Claude Sonnet 5, GPT 5.6 Sol, Claude Fable 5, Claude Opus 5, GPT 5.6 Terra, Kimi K3, DeepSeek V4 Pro, Grok 4.6, Gemini 3.8 Flash, Qwen 3.8 Max, Gemini 3.1 Pro Preview, GLM 5.3 Flash.

What does the data security level mean?

Folkbench rates relays from 0 to 4 based on public policies and verifiable evidence; aruhub API is currently level 0. The check covers leaks in responses and error messages only, and does not prove data retention or training policies.

How often is this data updated?

Availability and time to first token use a 24-hour window, and this page syncs live data every 15 minutes.