A cheap Flash model,
ninety-two of a hundred,
a tenth of the bill.

On October 31 I posted five Qwen3.8-27B agents against Fable 5 on ten LiveCodeBench problems. The main feedback: ten is not enough. So this time it is all 100 of GVS5H’s hardest LiveCodeBench problems. I ran team-solve-fanout, Memdoor’s team mode with three attempts in parallel, once on DeepSeek V4.1 Flash. Every test was graded, public and hidden.

	solved	$ per solved	Memdoor team-solve-fanout against it
Fable 5, single call	90.4%	0.68	about 1/11 the price, +1.6 points
GVS5H manager, Qwen3.8-27B	92.4%	0.52	about 1/9 the price, level within noise
GVS5H manager, DeepSeek V4.1 Flash	92.0%	0.054	same price, same score
GVS5H manager, Qwen3.8-Flash-Next	93.0%	0.062	same price, 1 point lower
Memdoor team-solve-fanout, DeepSeek V4.1 Flash	92%	0.060

Fable 5’s score on a Flash model, for 1/11 of its bill.

• Mine is one pass, priced at what OpenRouter billed ($5.49 for the 100). The other rows are the paper’s means of five passes at list rates, except the DeepSeek row: GVS5H’s own token counts (runs/per_problem_tokens.json) at the rates OpenRouter billed my run. At DeepSeek’s list rate the paper prices that row at $0.17.
• On DeepSeek the two harnesses write the same output (12.4M and 12.3M tokens a pass). Mine reads more input (88M against 1.5M), most of it from cache. The gap with Fable 5 is the model’s price.
• Grading follows the paper’s rule for the problems that accept more than one answer. The stock grader gives 91.

The ranking with every row · Per problem · The 8 misses, and what is next · Memdoor