Alibaba / manual · models.dev

Qwen3.8 27B

Opened from a screened package f6c34298809c.

Harnesses

Same model, different drivers. ChatGPT is not OpenCode. $ / MU is only set when every official task passed. $ / pass still shows on a partial run; it does not rank. Time is output tokens at 80 tok/s, not wall clock.

Stack	Passed	Attempts	Time	$ / MU	$ / pass	$ / hr	Tokens / pass	Burn	$ / OpenRouter
Grokmedium	3/3	4	31m 55s	$0.000209	$0.2991	$0.8882	99,456	1.53×	$0.2991
Grokxhigh	3/3	4	14m 47s	$0.000091	$0.1297	$0.8291	64,974	1.00×	$0.1297

Official jobs

Pass or fail on each sealed job. Attempts include retries. Failed tries stay in the bill.

Harness	Durable work queue 2/2	JSON Patch (RFC 6902 subset) 2/2	Token bucket and sliding window 2/2
Grok · medium	pass · 2	pass · 1	pass · 1
Grok · xhigh	pass · 2	pass · 1	pass · 1

Tokens by job

Billed tokens on each official job, including retries. Cache reads drop out of this count.

Grok · medium
182,038Durable work queue
40,245JSON Patch (RFC 6902 subset)
76,084Token bucket and sliding window
Grok · xhigh
121,662Durable work queue
34,889JSON Patch (RFC 6902 subset)
38,370Token bucket and sliding window

Where the bill goes

Failed attempts stay in the bill. Cache hits are a discount, shown separately.

Grok · medium298,367
Cache hit 5,234,987 (discount, not extra work)
Grok · xhigh194,921
Cache hit 1,826,798 (discount, not extra work)

Uncached in
Output
Thinking
Cache write

Retry tax

Attempts are harness calls, including failed hidden-verifier retries. Extra tries stay in the bill.

Grokmedium4 attempts · 1 extra / 3 jobs
Grokxhigh4 attempts · 1 extra / 3 jobs

Cache hits

Hits divided by uncached input plus hits. No cache means the harness did not report a discount.

Grokmedium97%
Grokxhigh94%

Time vs $ / MU

Decode time at 80 tok/s against what that finish cost. A long token bill is not the same as an expensive one.

Endpoints

Provider	SKU	List in	List out	Cache hit	Cache write
OpenRouter	`qwen/qwen3.8-27b`	$0.43	$2.55	$0.09	$0.53

All stacks