Time per task: agentic coding session

One long tool-using loop: 60,000 tokens in across turns, 4,000 tokens 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: agentic coding session board at 3.1 s, ahead of K2 Horizon 375B A23B (8.7 s) and Mercury 2.5 (9.5 s).

520 models · data as of 25 Oct 2026

1. 1Celeris-13.1 s
2. 2K2 Horizon 375B A23B8.7 s
3. 3Mercury 2.59.5 s
4. 4Gemini 2.5 Flash-Lite (none)13.9 s
5. 5DeepSeek V4.1 Flash (none)14.4 s
6. 6Gemini 3.1 Flash Lite14.8 s
7. 7Nova Micro15.2 s
8. 8Nemotron 3 Nano 30B A3B (none)15.7 s
9. 9Gemini 3.5 Flash Lite16.2 s
10. 10Mercury 216.7 s
11. 11Qwen3 Coder 30B A3B Instruct17.5 s
12. 12Trinity Large (thinking)18.1 s
13. 13Grok 4.20 Multi-Agent18.5 s
14. 14Beam 1.018.6 s
15. 15GPT-4.1 nano18.6 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	Ling Flash 2.0Ant Groupopen	16.0 min	44.4
2	GPT-5.4 ProOpenAI	9.6 min	63.3
3	Qwen3.5-4BthinkingAlibabaopen ↗	6.0 min	46.6
4	Claude Opus 4.1thinkingAnthropic	5.2 min	56.4
5	DeepSeek V3.2DeepSeekopen ↗	4.0 min	56.6
6	DeepSeek V4 Flash 0420maxDeepSeekopen	3.4 min	57.0
7	GLM 5 TurboZhipu AI	3.1 min	55.5
8	DeepSeek R1 0528DeepSeekopen ↗	3.1 min	49.5
9	Llama 3.2 Instruct 11B (Vision)Metaopen ↗	2.9 min	36.5
10	Qwen3.6 PlusAlibaba	2.9 min	58.8
11	Mixtral 8x7BMistral AIopen	2.7 min	31.7
12	Seed-OSS-36B-InstructByteDanceopen ↗	2.7 min	46.8
13	GPT-5.5 ProxhighOpenAI	2.7 min	65.3
14	GLM-4.5Zhipu AIopen ↗	2.6 min	51.7
15	Gemma 3 27BGoogleopen ↗	2.6 min	38.9
16	Qwen2.5 72BAlibabaopen ↗	2.5 min	42.2
17	Qwen3 VL 30B A3B InstructthinkingAlibabaopen ↗	2.5 min	44.0
18	Qwen3.8 2.4T A95BAlibabaopen ↗	2.4 min	61.0
19	Qwen2.5 Coder 32B InstructAlibabaopen ↗	2.2 min	42.1
20	GPT-4 TurboOpenAI	2.1 min	40.9
21	GLM-4.6no reasoningZhipu AIopen ↗	2.0 min	46.7
22	Qwen3.8 MaxxhighAlibaba	2.0 min	57.1
23	MiniMax M2.7MiniMaxopen ↗	2.0 min	55.6
24	Qwen3.8-Flash-NextAlibabaopen ↗	2.0 min	60.3
25	Granite 3.3 8Bno reasoningIBMopen ↗	1.9 min	39.2
26	Gemma 2 27BGoogleopen ↗	1.9 min	41.0
27	MiMo-V2.6-ProXiaomiopen ↗	1.9 min	64.0
28	GLM 4.7no reasoningZhipu AIopen ↗	1.9 min	48.9
29	Phi-4Microsoftopen ↗	1.8 min	38.9
30	Ling 3.0 TinyAnt Groupopen	1.8 min	48.7
31	Gemini 4 ArgonGoogle	1.8 min	63.6
32	GPT-5.3-CodexxhighOpenAI	1.8 min	63.8
33	Kimi K2.7 CodeMoonshot AIopen ↗	1.8 min	55.8
34	MiMo-V2.6-FlashXiaomiopen ↗	1.8 min	59.7
35	Qwen3 VL 235B A22B InstructthinkingAlibabaopen ↗	1.7 min	51.0
36	Claude Opus 4.6highAnthropic	1.7 min	62.7
37	Qwen3.7 PlusAlibaba	1.7 min	58.5
38	Qwen3 VL 8B InstructthinkingAlibabaopen ↗	1.7 min	42.7
39	Mistral Small 3Mistral AIopen ↗	1.7 min	39.3
40	Hermes 4 - Llama-3.1 405Bno reasoningNous Researchopen	1.7 min	43.4
41	Llama 3.1 70BMetaopen ↗	1.7 min	40.9
42	North Mini CodeCohereopen	1.7 min	45.9
43	Mistral Small 3.1Mistral AIopen	1.6 min	36.3
44	GPT-5.2-CodexOpenAI	1.6 min	57.3
45	Grok Build 0.1SpaceXAI	1.6 min	54.8
46	MiniMax M2.5MiniMaxopen ↗	1.6 min	53.6
47	Mistral Small 3.2Mistral AIopen ↗	1.6 min	41.6
48	Qwen3 235B A22B 2507thinkingAlibabaopen ↗	1.6 min	50.3
49	GLM 5V TurboZhipu AI	1.5 min	55.9
50	GLM 5.3 FlashZhipu AIopen ↗	1.5 min	62.7
51	MiniMax-M2.1MiniMaxopen ↗	1.5 min	53.8
52	Grok 3 minithinkingSpaceXAI	1.5 min	49.4
53	Gemma 3 12BGoogleopen ↗	1.5 min	36.9
54	Phi-4-MiniMicrosoftopen ↗	1.5 min	35.8
55	Mistral Medium 3Mistral AI	1.5 min	44.1
56	Claude Sonnet 4.5no reasoningAnthropic	1.4 min	50.3
57	Kimi K2Moonshot AIopen ↗	1.4 min	50.4
58	Qwen3.8 27BAlibabaopen ↗	1.4 min	54.6
59	Claude Sonnet 4.6no reasoningAnthropic	1.4 min	54.0
60	GPT-6.1 SollowOpenAI	1.4 min	64.9

Show 100 more

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