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

10 relays

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

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

90%

Cache hit ¥0.35 · input ¥3.5 (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

¥2.66

¥7.70 without cache, 65% saved

Cache-weighted session cost

GPT 6 Astra · 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 7 / 10

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

10 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
GPT 6 Astra	GPT-Pro++ 独享满血不降智	—		—	0
DeepSeek V4.1 Flash	国模分组稳定版本官方2.5折	—	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 Sonnet 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	—	—
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 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	—	—
Kimi K3	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 GPT 6 Astra cost on aruhub API?

Input ¥3.5 and output ¥17.5 per 1M tokens; at 2M input and 40k output tokens, a coding session costs about ¥7.70, an API rate of 0.052.

How does the cache hit rate affect price?

On aruhub API, the cache-hit price is 10% of the input price. At an 80% hit rate a coding session costs about ¥2.66, versus about ¥7.70 without cache.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, DeepSeek V4 Pro, Grok 4.6, Kimi K3, 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.

aruhub API · GPT 6 AstraAuthenticity, speed and price analysis

Comparison summary

Authenticity

Model consistency