BlogLLM Benchmark Scores as Published, Side by Side

Claude Fable 5.2 vs DeepSeek-V4.1-Flash-1017: Benchmark Scores Side by Side

Every test both Claude Fable 5.2 by Anthropic and DeepSeek-V4.1-Flash-1017 by DeepSeek have a published number for: 15 tests in 5 areas. Each number is as printed, with the page it comes from and the setup it was run under; nothing is ranked or rescored, and two numbers under different setups are two results, not one comparison.

Models Claude Fable 5.2, DeepSeek-V4.1-Flash-1017Data version 2026111201, as of 2026-11-12

Model

What each developer prints on its own model page; a dash means it prints none. Runs on is the smallest device that holds the model at 4-bit: the developer's words where it prints them, otherwise our estimate from the size.

FactClaude Fable 5.2DeepSeek-V4.1-Flash-1017
Released2026-11-092026-10-17
Context window1M1M
Max output128K384K
InputText, imageText, image
OutputTextText
WeightsClosedOpen (MIT)
Size-748B
Runs on-8 × H100 (about 496 GB at 4-bit, estimate)

Knowledge Work

TestClaude Fable 5.2DeepSeek-V4.1-Flash-1017
AutomationBench
51.4%Anthropic, Claude Fable 5.2 and Mythos 5.2, checked 2026-11-09; printed on 1 more page
60.5%DeepSeek, DeepSeek-V4.1-Flash-1017 model card, checked 2026-10-18; printed on 2 more pagesDataset revision: v1.0.6 (printed in the label)
OfficeQA Pro
62.4%Moonshot AI, Introducing Kimi K3.1 Pro, checked 2026-10-24Effort: max; Harness: Claude Code
GDPval-AA v2.1
1746Moonshot AI, Introducing Kimi K3.1 Pro, checked 2026-10-24Effort: max
AA-Briefcase v1.1
1785Anthropic, Claude Fable 5.2 and Mythos 5.2, checked 2026-11-09; printed on 1 more page
1570Moonshot AI, Introducing Kimi K3.1 Pro, checked 2026-10-24Effort: max

Agentic Coding

TestClaude Fable 5.2DeepSeek-V4.1-Flash-1017
DeepSWE v1.1
77%DeepSeek, DeepSeek-V4.1-Flash-1017 model card, checked 2026-10-18; printed on 2 more pagesHarness: Mini-SWE-V2
Terminal-Bench 4.0
67.6%Anthropic, System Card: Claude Fable 5.2 & Claude Mythos 5.2, checked 2026-11-09Effort: max; Harness: Claude Code --bare; Fallback: safeguards on; the fallback answered 2.5% of requests, affecting 10% of trials
42.3%DeepSeek, DeepSeek-V4.1-Flash-1017 model card, checked 2026-10-18; printed on 3 more pagesHarness: Claude Code
Vibe Code Bench
93.77%Vals AI, Vibe Code Bench v1.1 leaderboard, checked 2026-11-09Harness: OpenHands
87.73%Vals AI, Vibe Code Bench v1.1 leaderboard, checked 2026-11-02Harness: OpenHands

Science and Math

TestClaude Fable 5.2DeepSeek-V4.1-Flash-1017
GPQA Diamond
93.9%Epoch AI, Epoch AI Benchmarking Hub, checked 2026-11-10Effort: max
92.2%Moonshot AI, Introducing Kimi K3.1 Pro, checked 2026-10-24Effort: max
Terminal-Bench Science 0.1
74.6%Anthropic, Claude Fable 5.2 and Mythos 5.2, checked 2026-11-09; printed on 1 more page
9.3%Moonshot AI, Introducing Kimi K3.1 Pro, checked 2026-10-24Effort: max; Model as printed: DeepSeek-V4.1-Flash-1017
15.7%Moonshot AI, Introducing Kimi K3.1 Pro, checked 2026-10-24Effort: max; Model as printed: DeepSeek-V4.1-Flash-1017 (max)
FrontierMath Tier 4 (v2)
100%Epoch AI, Epoch AI Benchmarking Hub, checked 2026-11-10Effort: max
46.3%Epoch AI, Epoch AI Benchmarking Hub, checked 2026-10-30Effort: max
FrontierMath Tier 1-3 (v2)
95.1%Epoch AI, Epoch AI Benchmarking Hub, checked 2026-11-10Effort: max
71.2%Epoch AI, Epoch AI Benchmarking Hub, checked 2026-10-30Effort: max

Multimodal Understanding

TestClaude Fable 5.2DeepSeek-V4.1-Flash-1017
Chartography (with tools)
83%DeepSeek, DeepSeek-V4.1-Flash-1017 model card, checked 2026-10-18; printed on 1 more pageTools: enabled; exact tools in source

General Knowledge and Reasoning

TestClaude Fable 5.2DeepSeek-V4.1-Flash-1017
Humanity's Last Exam (no tools)
42.6%DeepSeek, DeepSeek-V4.1-Flash-1017 model card, checked 2026-10-18; printed on 2 more pages
44%DeepSeek, DeepSeek-V4.1-Flash-1017 model card, checked 2026-10-18Dataset revision: text-only subset of HLE (dagger on the card)
Humanity's Last Exam (with tools)
74.2%Anthropic, Claude Fable 5.2 and Mythos 5.2, checked 2026-11-09; printed on 1 more page
58.8%Moonshot AI, Introducing Kimi K3.1 Pro, checked 2026-10-24Effort: max; Tools: with tools
SimpleQA Verified
94%Epoch AI, Epoch AI Benchmarking Hub, checked 2026-11-10Effort: max
42.5%Epoch AI, Epoch AI Benchmarking Hub, checked 2026-10-30Effort: max

Price

API list prices in US dollars per one million tokens, as printed on the developer's pricing page on the day checked; a dash means the developer lists that price in another unit or not at all. Cost per task is measured by Vals AI.

PriceClaude Fable 5.2DeepSeek-V4.1-Flash-1017
Input
$8Anthropic, Anthropic API pricing, checked 2026-11-09
$0.3DeepSeek, DeepSeek API pricing, checked 2026-10-18
Cached input
$0.2Anthropic, Anthropic API pricing, checked 2026-11-09
$0.006DeepSeek, DeepSeek API pricing, checked 2026-10-18
Cache write
$10Anthropic, Anthropic API pricing, checked 2026-11-09Cache: 5 minute cache
-
Cache write, 1 hour
$16Anthropic, Anthropic API pricing, checked 2026-11-09
-
Output
$40Anthropic, Anthropic API pricing, checked 2026-11-09
$1.2DeepSeek, DeepSeek API pricing, checked 2026-10-18
Input, batch
$4Anthropic, Anthropic API pricing, checked 2026-11-09
-
Output, batch
$20Anthropic, Anthropic API pricing, checked 2026-11-09
-
Input, fast
$16Anthropic, Anthropic API pricing, checked 2026-11-09
-
Output, fast
$80Anthropic, Anthropic API pricing, checked 2026-11-09
-
Input, off-peak hours-
$0.15DeepSeek, DeepSeek API pricing, checked 2026-10-18
Cached input, off-peak hours-
$0.003DeepSeek, DeepSeek API pricing, checked 2026-10-18
Output, off-peak hours-
$0.6DeepSeek, DeepSeek API pricing, checked 2026-10-18
Cost per taskVals AI, USD, lower is better
$22.13Vals AI, Vibe Code Bench v1.1 leaderboard, checked 2026-11-09Harness: OpenHands
$0.58Vals AI, Vibe Code Bench v1.1 leaderboard, checked 2026-10-27Harness: OpenHands

Claude Fable 5.2 has 31 more and DeepSeek-V4.1-Flash-1017 has 38 more: tests only one of them has a number for, listed on its model page.

More side by side

Every number here is read from the same catalog file as the benchmark table, data version 2026111201. A pair has a page when both models are their developers' flagships and both have a number for at least 10 of the same tests. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.