Speed: output tokens per second

Median output throughput measured across healthy providers over the last 30 minutes of live traffic (OpenRouter).

As of 5 Oct 2026, BenchLeader ranks Celeris-1 first on the Speed: output tokens per second board at 1499, ahead of Mercury 2 (922) and Mercury 2.5 (712).

642 models · data as of 5 Oct 2026

1. 1Celeris-11499
2. 2Mercury 2922
3. 3Mercury 2.5712
4. 4Gemini 3.5 Flash Lite360
5. 5Trinity Large (thinking)355
6. 6Gemini 2.5 Flash-Lite (thinking)345
7. 7Ling 3.0 Flash325
8. 8Grok 4.20 Multi-Agent318
9. 9Gemini 3.1 Flash Lite306
10. 10Gemini 3.7 Flash (high)303
11. 11Nemotron 3.5 Lightning301
12. 12Nova Micro296
13. 13Ling 3.0 Flash Fin290
14. 14Nemotron 3 Nano Omni 30B A3b (thinking)270
15. 15Gemini 3.8 Flash (high)249

SearchAll providersAlibabaAmazonAnt GroupAnthropicArcee AIByteDanceCelerisCohereDeepSeekGoogleIBMInceptionMBZUAIMeituanMetaMicrosoftMiniMaxMistral AIMoonshot AIMultiverse ComputingNous ResearchNVIDIAOpenAIPerplexityRekaSpaceXAIStepFunTencentThinking MachinesUpstageWriterXiaomiZhipu AIOpen weightsNew modelsBest per model (248)248 of 248

#	Model	tok/s ↓	First token ↕
1	CCeleris-1Celeris	1499	0.6 s
2	Mercury 2Inception	922	0.3 s
3	Mercury 2.5Inception	712	1.5 s
4	Gemini 3.5 Flash LiteGoogle	360	0.7 s
5	Trinity LargethinkingArcee AIopen	355	0.6 s
6	Gemini 2.5 Flash-LitethinkingGoogle	345	0.5 s
7	Ling 3.0 FlashAnt Groupopen ↗	325	1.0 s
8	Grok 4.20 Multi-AgentSpaceXAI	318	6.6 s
9	Gemini 3.1 Flash LiteGoogle	306	0.9 s
10	Gemini 3.7 FlashhighGoogle	303	2.5 s
11	Nemotron 3.5 LightningNVIDIAopen ↗	301	0.2 s
12	Nova MicroAmazon	296	0.3 s
13	Ling 3.0 Flash FinAnt Groupopen	290	1.1 s
14	Nemotron 3 Nano Omni 30B A3bthinkingNVIDIAopen	270	–
15	Gemini 3.8 FlashhighGoogle	249	2.5 s
16	Muse Spark 1.2xhighMeta	242	3.9 s
17	Nemotron 3 Nano 30B A3Bno reasoningNVIDIAopen ↗	237	0.8 s
18	gpt-oss-20blowOpenAIopen ↗	224	0.4 s
19	Granite 4.2 3BIBMopen ↗	221	–
20	Qwen3.5 Omni FlashAlibaba	220	1.9 s
21	Muse Spark 1.1Meta	220	3.6 s
22	Gemini 2.5 FlashthinkingGoogle	217	0.8 s
23	DeepSeek V4 FlashmaxDeepSeekopen ↗	216	1.4 s
24	Gemini 3.5 FlashhighGoogle	216	1.6 s
25	DeepSeek V4.1 Flashno reasoningDeepSeekopen ↗	215	1.0 s
26	Command A+Cohereopen	215	0.3 s
27	Deepseek v4 Flash VisionmaxDeepSeek	215	1.3 s
28	GPT-5.4 MinixhighOpenAI	212	0.9 s
29	o3-miniOpenAI	207	4.3 s
30	Qwen3 7maxAlibaba	203	1.1 s
31	Gemini 3 FlashthinkingGoogle	203	1.2 s
32	Solar Mini4Upstage	202	1.1 s
33	Step 3.7 FlashStepFunopen ↗	201	3.6 s
34	Ministral 3 3BMistral AIopen	194	0.7 s
35	Nemotron Nano 12B v2 VLno reasoningNVIDIAopen	194	1.1 s
36	Mistral Small 4thinkingMistral AIopen	192	0.3 s
37	Gemini 3.6 FlashhighGoogle	189	1.3 s
38	Nova 2.0 LitelowAmazon	189	–
39	TInkling SmallThinking Machinesopen ↗	180	0.5 s
40	Qwen3-Next 80B A3BthinkingAlibabaopen ↗	178	2.1 s
41	gpt-oss-120bhighOpenAIopen ↗	174	0.4 s
42	Nova LiteAmazon	174	0.4 s
43	Mistral Medium 3.5Mistral AIopen	165	0.6 s
44	GPT-5.4 Nanono reasoningOpenAI	165	0.7 s
45	Nemotron 3 Ultra 550B A55BthinkingNVIDIAopen ↗	164	0.3 s
46	Nemotron 3 Super 120B A12bthinkingNVIDIAopen ↗	162	1.5 s
47	Qwen3.5-35B-A3Bno reasoningAlibabaopen ↗	156	2.1 s
48	nvidia-nemotron-nano-9b-v2no reasoningNVIDIAopen	156	2.2 s
49	DeepSeek V4 Prono reasoningDeepSeekopen ↗	153	2.0 s
50	Muse Spark 1.3maxMeta	152	4.2 s
51	GPT-5 nanomediumOpenAI	151	2.4 s
52	Solar Pro 3Upstage	146	0.6 s
53	Qwen3 30B A3B 2507thinkingAlibabaopen ↗	145	1.9 s
54	o4-minihighOpenAI	143	4.3 s
55	Quasar 438BMultiverse Computing	143	–
56	GPT-4.1 minihighOpenAI	142	0.7 s
57	Qwen3.6 35B A3Bno reasoningAlibabaopen ↗	141	2.1 s
58	Grok 4.3highSpaceXAI	141	1.6 s
59	Qwen3.5-122B-A10Bno reasoningAlibabaopen ↗	141	2.3 s
60	Muse GlimmerhighMetaopen ↗	140	0.8 s

Show 100 more

Cite as: BenchLeader, “Speed: output tokens per second”, https://www.benchleader.com/leaderboards/speed, data as of 5 Oct 2026.