Output Tokens/sec

Speed

Generation speed — tokens produced per second

Higher is better

Rank	Model	Provider	Throughput
1	Celeris-1	Celeris	1827.0 tok/s
2	Inception: Mercury 2.5	Inception	731.3 tok/s
3	Google: Gemini 3.5 Flash Lite	Google	406.3 tok/s
4	Arcee AI: Trinity Large Thinking	Arcee AI	367.0 tok/s
5	NVIDIA: Nemotron 3.5 Lightning	NVIDIA	291.8 tok/s
6	gpt-oss-20b (low)	OpenAI	286.8 tok/s
7	inclusionAI: Ling 3.0 Flash FinHistorical	inclusionAI	280.6 tok/s
8	Nova Micro	Amazon	257.2 tok/s
9	NVIDIA Nemotron 3 Nano 30B A3B (Non-reasoning)	NVIDIA	248.3 tok/s
10	Claude Haiku 5.5 (Max)	Anthropic	241.0 tok/s
11	Muse Spark 1.3 (xhigh)	Meta	241.0 tok/s
12	Nemotron 3 Nano Omni 30B A3B Reasoning	NVIDIA	240.3 tok/s
13	Ministral 3 3B	Mistral AI	238.3 tok/s
14	gpt-oss-20b (high)	OpenAI	232.1 tok/s
15	gpt-oss-120b (low)	OpenAI	230.3 tok/s
16	Granite 4.2 3B	IBM	229.6 tok/s
17	NVIDIA Nemotron 3 Nano 30B A3B (Reasoning)	NVIDIA	229.3 tok/s
18	Nova 2.0 Lite (Non-reasoning)	Amazon	225.7 tok/s
19	Qwen3.5 Omni Flash	Alibaba	224.6 tok/s
20	Inkling (xhigh)	Thinking Machines	214.5 tok/s
21	gpt-oss-120b (high)	OpenAI	214.4 tok/s
22	inclusionAI: Ling 3.1 Flash	inclusionAI	211.5 tok/s
23	NVIDIA Nemotron Nano 9B V2 (Reasoning)	NVIDIA	210.3 tok/s
24	Nova 2.0 Lite (low)	Amazon	209.0 tok/s
25	NVIDIA Nemotron Nano 12B v2 VL (Non-reasoning)	NVIDIA	207.7 tok/s
26	Nova 2.0 Lite (high)	Amazon	203.6 tok/s
27	Nova 2.0 Lite (medium)	Amazon	199.5 tok/s
28	Thinking Machines: Inkling Small	Thinking Machines	196.4 tok/s
29	Mistral Small 4 (Reasoning)	Mistral AI	184.0 tok/s
30	DeepSeek V4.1 Flash (Reasoning, Max Effort)	DeepSeek	175.1 tok/s
31	Mistral Small 4 (Non-reasoning)	Mistral AI	171.2 tok/s
32	Claude Haiku 5.5 (Xhigh)	Anthropic	171.1 tok/s
33	Qwen3 Next 80B A3B (Reasoning)	Alibaba	170.6 tok/s
34	NVIDIA Nemotron Nano 9B V2 (Non-reasoning)	NVIDIA	169.3 tok/s
35	DeepSeek V4 Flash Vision (Reasoning, Max Effort)	DeepSeek	168.2 tok/s
36	Mistral: Mistral Medium 3.5	Mistral AI	166.7 tok/s
37	Qwen: Qwen3 Next 80B A3B InstructHistorical	Alibaba	163.7 tok/s
38	DeepSeek V4.1 Flash (Non-Reasoning)	DeepSeek	161.1 tok/s
39	Claude Haiku 5.5 (High)	Anthropic	158.9 tok/s
40	Cohere: Command A	Cohere	157.4 tok/s
41	GPT-5.5 Instant (June 2026)	OpenAI	154.8 tok/s
42	Claude Haiku 5.5 (Medium)	Anthropic	150.9 tok/s
43	Nemotron 3 Ultra 550B A55B (Reasoning)	NVIDIA	148.8 tok/s
44	Gemma 4 12B (Non-reasoning)	Google	147.4 tok/s
45	Qwen3.5 122B A10B (Non-reasoning)	Alibaba	146.9 tok/s
46	Gemma 4 12B (Reasoning)	Google	146.3 tok/s
47	Claude Haiku 5.5 (Low)	Anthropic	145.7 tok/s
48	Muse Glimmer (high)	Meta	144.8 tok/s
49	Qwen3.6 35B A3B (Non-reasoning)	Alibaba	144.5 tok/s
50	Nemotron 3 Super 120B A12B (Reasoning)	NVIDIA	144.2 tok/s

Access this data

JSON GET /api/v1

[
  {
    "rank": 1,
    "model": "Celeris-1",
    "provider": "Celeris",
    "score": "1827.011"
  },
  {
    "rank": 2,
    "model": "Inception: Mercury 2.5",
    "provider": "Inception",
    "score": "731.258"
  },
  {
    "rank": 3,
    "model": "Google: Gemini 3.5 Flash Lite",
    "provider": "Google",
    "score": "406.264"
  }
]

Curl

$ curl -s https://rightrank.ai/api/v1/rankings/output-tokens-sec.json | jq '.[:3]'

CLI brew install rightrank

$ rightrank top output-tokens-sec --limit 5
$ rightrank compare claude-opus-4.6 gpt-5.4-pro --category output-tokens-sec
$ rightrank value output-tokens-sec  # best performance per dollar

Claude Code Skill Skill

# In Claude Code, use the RightRank skill:
/rightrank best model for output-tokens-sec
/rightrank compare top 5 by value
/rightrank cheapest model scoring above 80 on output-tokens-sec