Time per task: structured extraction

2,000 tokens of source text in, 200 tokens of JSON out. Time = time to first answer + output tokens ÷ output speed.

As of 25 Oct 2026, BenchLeader ranks Celeris-1 first on the Time per task: structured extraction board at 0.7 s, ahead of Gemini 2.5 Flash-Lite (1.0 s) and Gemini 2.5 Flash (1.3 s).

520 models · data as of 25 Oct 2026

1. 1Celeris-10.7 s
2. 2Gemini 2.5 Flash-Lite (none)1.0 s
3. 3Gemini 2.5 Flash (none)1.3 s
4. 4DeepSeek V4.1 Flash (none)1.6 s
5. 5Nemotron 3 Nano 30B A3B (none)1.6 s
6. 6Nova Micro1.6 s
7. 7GPT-4.1 nano1.6 s
8. 8Ministral 3 3B1.6 s
9. 9Llama 3.2 3B Instruct1.7 s
10. 10Llama 3.2 1B Instruct1.7 s
11. 11GPT-5.4 Mini (none)1.8 s
12. 12GPT-5.4 Nano (none)1.8 s
13. 13Gemini 3.5 Flash (minimal)1.9 s
14. 14Gemini 3 Flash (none)1.9 s
15. 15Beam 1.0 501B A23B1.9 s

SearchAll providersAlibabaAmazonAnt GroupAnthropicArcee AIByteDanceCelerisCohereDeepSeekGoogleIBMInceptionMBZUAIMeituanMetaMicrosoftMiniMaxMistral AIMoonshot AIMultiverse ComputingNous ResearchNVIDIAOpenAIReflectionRekaSapiens AISarvamSpaceXAIStepFunSwiss AIThinking MachinesUpstageXiaomiZhipu AIOpen weightsNew modelsBest per model (244)244 of 244

#	Model	seconds ↓	Index ↕
1	DeepSeek V4 Flash 0420maxDeepSeekopen	2.1 min	57.0
2	Qwen3.6 PlusAlibaba	1.8 min	58.8
3	DeepSeek R1 0528DeepSeekopen ↗	1.1 min	49.5
4	GPT-5.3-CodexxhighOpenAI	1.0 min	63.8
5	Seed-OSS-36B-InstructByteDanceopen ↗	1.0 min	46.8
6	Qwen3.8 2.4T A95BAlibabaopen ↗	54.1 s	61.0
7	MiniMax M2.7MiniMaxopen ↗	50.5 s	55.6
8	Ling Flash 2.0Ant Groupopen	50.4 s	44.4
9	Qwen3.8-Flash-NextAlibabaopen ↗	44.9 s	60.3
10	Kimi K2.7 CodeMoonshot AIopen ↗	43.3 s	55.8
11	MiMo-V2.6-ProXiaomiopen ↗	42.9 s	64.0
12	MiMo-V2.6-FlashXiaomiopen ↗	41.1 s	59.7
13	Ling 3.0 TinyAnt Groupopen	40.9 s	48.7
14	Qwen3.7 PlusAlibaba	39.4 s	58.5
15	North Mini CodeCohereopen	37.0 s	45.9
16	Grok Build 0.1SpaceXAI	35.3 s	54.8
17	MiniMax M2.5MiniMaxopen ↗	35.3 s	53.6
18	GLM 5.3 FlashZhipu AIopen ↗	35.2 s	62.7
19	GPT-5.4 ProOpenAI	34.6 s	63.3
20	MiniMax-M2.1MiniMaxopen ↗	34.1 s	53.8
21	Grok 3 minithinkingSpaceXAI	33.9 s	49.4
22	NewKimi K3.1 ProMoonshot AI	32.9 s	66.9
23	LongCat 2.0Meituanopen ↗	31.4 s	49.4
24	GLM-4.5-AirZhipu AIopen ↗	31.2 s	46.6
25	MiMo-V2.5Xiaomiopen ↗	30.6 s	54.3
26	Granite 4.2 30BIBMopen ↗	30.2 s	48.7
27	Magistral Small 1.2Mistral AIopen	29.6 s	40.3
28	GLM 5.3lowZhipu AIopen ↗	27.8 s	55.6
29	o4-minihighOpenAI	26.5 s	50.1
30	Mistral Large 4Mistral AIopen	26.4 s	56.7
31	Reka Flash 3Rekaopen ↗	26.4 s	38.9
32	MiniMax M3MiniMaxopen ↗	26.1 s	56.4
33	Granite 4.2 8BIBMopen ↗	26.1 s	46.1
34	MiniMax-M1MiniMaxopen	25.0 s	47.5
35	NewMiniMax M3.1 FlashMiniMaxopen ↗	24.3 s	61.9
36	Gemini 2.5 ProGoogle	24.1 s	53.9
37	MiniMax-M2MiniMaxopen ↗	23.6 s	50.8
38	Granite 4.2 3BIBMopen ↗	22.9 s	44.8
39	Gemini 3.1 ProGoogle	22.6 s	60.5
40	Ring-2.6-1TAnt Groupopen	22.0 s	49.3
41	DeepSeek V3.1thinkingDeepSeekopen ↗	21.0 s	48.1
42	Agnes 2.5 Pro BetaSapiens AI	20.9 s	59.1
43	Kimi K2 ThinkingthinkingMoonshot AIopen ↗	20.1 s	52.4
44	Claude Opus 4.1thinkingAnthropic	19.4 s	56.4
45	Qwen3.5-4Bno reasoningAlibabaopen ↗	19.0 s	42.2
46	GPT-5.5 ProxhighOpenAI	18.8 s	65.3
47	DeepSeek R1DeepSeekopen ↗	18.8 s	46.8
48	Step 3.5 FlashStepFunopen ↗	18.6 s	53.7
49	Muse Spark 1.2xhighMeta	18.3 s	63.2
50	Step 3.5 Flash 0202StepFunopen	18.2 s	50.7
51	Granite 3.3 8Bno reasoningIBMopen ↗	17.6 s	39.2
52	Solar Pro 3Upstage	17.2 s	43.2
53	Mistral Medium 3.5Mistral AIopen	16.1 s	49.4
54	GLM-4.5Zhipu AIopen ↗	16.0 s	51.7
55	GLM 5 TurboZhipu AI	15.9 s	55.5
56	Nemotron 3 Super 120B A12BthinkingNVIDIAopen ↗	15.5 s	50.6
57	Muse Spark OpenxhighMetaopen ↗	15.4 s	62.0
58	Step 3.7 FlashStepFunopen ↗	15.3 s	51.5
59	Command A+Cohereopen	15.3 s	50.5
60	Qwen3.7maxAlibaba	15.1 s	60.6

Show 100 more

Cite as: BenchLeader, “Time per task: structured extraction”, data as of 25 Oct 2026, https://www.benchleader.com/leaderboards/time-extraction.