CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
← Back to feed

DeepSeek V4.1 Flash Tops September LiveBench Agentic Coding at 77.3, Priced at $0.03 a Task

DeepSeek V4.1 Flash holds the top position on LiveBench’s Agentic Coding category in the September 2026 run, scoring 77.3 against the next-best score of 66.1 from Claude Fable 5.1 Max Effort. The gap to GPT-6 Astra is 20 points. It costs $0.029 per successful task — 42 times cheaper than Fable 5.1’s $1.212 and 25 times cheaper than GPT-6 Astra’s $0.736.

The result is notable partly because of what came before it. DeepSeek V4 Flash 0731, the model’s July predecessor, scored 46.8 on the same agentic coding metric — a 30-point gap to V4.1 Flash in roughly eight weeks.

The September LiveBench picture

On overall LiveBench score, V4.1 Flash ranks fifth at 81.1, behind Claude Fable 5.1 (83.4), Claude Fable 5 (83.0), GPT-6 Astra (82.2), and Muse Spark 1.3 (81.6). But the Agentic Coding subcategory, which tests models on realistic multi-step software tasks, reverses that order: V4.1 Flash at 77.3 is the only model above 66.1 in the field.

Its broader September scores: Reasoning 86.7, Coding 80.0, Mathematics 93.3, Data Analysis 79.3. The Instruction Following category, at 70.0, is the model’s relative weak point relative to the leaderboard leaders.

The model carries the “open” tag on LiveBench, meaning weights are publicly available.

What changed

DeepSeek has not published a detailed changelog between V4 Flash 0731 and V4.1 Flash, but the agentic coding jump is not matched by equivalent gains in core coding (the older model’s LiveBench coding score was 75.0, V4.1 Flash scores 80.0) or reasoning (86.6 to 86.7). The agentic coding delta is disproportionate, suggesting targeted post-training on multi-step agent tasks rather than a general capability upgrade.

DeepSeek V4 Pro, the larger open-weight model in the same family, scores 54.9 on agentic coding at $0.044/task — lower than V4.1 Flash despite the larger compute footprint and slightly higher price.

The cost angle

The 42x cost gap between V4.1 Flash and Fable 5.1 on agentic tasks is the kind of number that resets procurement conversations. At $0.029 per task, a team running 100,000 agentic tasks per month pays $2,900 with V4.1 Flash versus $121,200 with Fable 5.1. If the task distribution fits the model’s Agentic Coding profile, the open-weight option currently outperforms the most capable closed model on that axis by a wide margin.

The broader pattern holds: DeepSeek V4.1 Flash on overall LiveBench at 81.1 sits close to models costing 7 to 50 times more per task. On the one subcategory that most directly measures agent utility, it leads the field outright.