ResearchLLM Verifications Leaderboard

Data as of · 78 models, 80 questions · reviewed by Suhas Bhairav · methodology

MicrosoftOpen weightsReleased 10 Jan 2025

Phi 4 benchmarks, speed and API pricing

Reviewed by Suhas Bhairav · Tested

Compare

Available through OpenRouter on

  • DeepInfra$0.056 in · $0.112 out
Intelligence41.2% correct#54 of 72
Speed103tok/s#35 of 72
Cost$0.00012per question#6 of 72List $0.07 in · $0.14 out per 1M
Verbosity699tokens per question#23 of 72

Summary

  • Answers 33 of 80 verification questions correctly (41.2%), placing #54 of 72 models tested on the current suite.
  • Generates 103 output tokens per second, faster than 52% of the models the leaderboard measures.
  • Starts responding after 0.5 s (median time to first token).
  • Costs $0.00012 per question as billed, #6 cheapest of 12 models priced under $0.15 per 1M tokens.

Specifications

Reasoning
No
Input
text
Output
text
Context window
16k tokens
Knowledge cutoff
30 Jun 2024
OpenRouter ID
microsoft/phi-4
Last tested
26 Oct 2026

Peer group: 13 models with a blended list price under $0.15 per 1M tokens (3 input : 1 output). Charts below compare Phi 4 with all models; cost and speed charts focus on this peer group.

Intelligence

Verification score: the share of 80 questions from open datasets answered correctly. See the benchmark datasets.

Score vs. cost per question

Phi 4 circled; dashed line joins models no cheaper model beats

Most attractive quadrant0255075100$0.00001$0.0001$0.001$0.01$0.10Cost per question (USD, log) →Verification score (%) →Qwen3.8 2.4T A95BCommand AReka Flash 3Gemma 4 26B A4BGemma 4 31BQwen3.5 397B A17BHermes 4 405BMuse WonderGrok 4.7o3gpt-oss-20bGranite 4.0 MicroNova Micro 1.0Mercury 2.5Phi 4Granite 4.2 8BHy3GLM 5.3 FlashSolar Pro 4Trinity Large ThinkingMiniMax M3Mistral Large 4Claude Opus 5.5GPT-6.1 SolGemini 4 ArgonGPT-5.4GPT-6 Lunagpt-oss-120b

Score vs. end-to-end response time

Phi 4 circled · accurate and quick sits top left

Most attractive quadrant02550751000 s25 s50 s75 s100 sEnd-to-end response time (s) →Verification score (%) →Qwen3.8 2.4T A95BCommand AReka Flash 3Llama 4 ScoutGemma 4 26B A4BLing 3.1 FlashHermes 4 405BMuse WonderGranite 4.0 MicroSolar Pro 4DeepSeek V4 Pro 0813Mistral Large 4Claude Opus 5.5Llama 4 MaverickKimi K3.1 Pro

Score vs. tokens per question

Phi 4 circled · output tokens, reasoning included

Most attractive quadrant02550751001001,00010,000Output tokens per question (log) →Verification score (%) →Reka Flash 3Qwen3.7 PlusGemma 4 26B A4BGemma 4 31BQwen3.5 397B A17BHermes 4 405BClaude Fable 5.1GPT-6 AstraMuse WonderGrok 4.7o3gpt-oss-20bGranite 4.0 MicroNova Micro 1.0Mercury 2.5Phi 4Granite 4.2 8BHy3GLM 5.3 FlashSolar Pro 4Qwen3.6 35B A3BInkling SmallNemotron 3 UltraClaude Opus 5.5Gemini 4 ArgonGPT-5.4GPT-6 SolMiMo-V2.6-Pro

Verification score

All models tested on the current suite

  1. 1Claude Opus 5.585.0
  2. 2Gemini 4 Argon83.8
  3. 3GPT-5.3-Codex82.5
  4. 3Gemini 3.8 Flash82.5
  5. 5Claude Fable 5.181.2
  6. 6Grok 4.780.0
  7. 7Muse Wonder78.8
  8. 7Claude Sonnet 5.578.8
  9. 9Gemini 3.1 Pro Preview77.5
  10. 9Muse Spark Open77.5
  11. 54Phi 441.2

61 more models · see the full leaderboard

6 of 78 tested models have no verification score

Cost

Average cost of one question, split into input, reasoning and answer tokens at each model's list price. Reasoning tokens are billed as output.

Cost per question, as billed

All models · lower is better

  1. 1Granite 4.0 Micro$0.000051
  2. 2gpt-oss-20b$0.000066
  3. 3Granite 4.2 8B$0.000077
  4. 4Nova Micro 1.0$0.000087
  5. 5Mercury 2.5$0.00011
  6. 6Phi 4$0.00012
  7. 7gpt-oss-120b$0.00013
  8. 8GLM 5.3 Flash$0.00015
  9. 9Ministral 3 3B 2512$0.00019
  10. 10GPT-6 Luna$0.00025

62 more models · see the full leaderboard

6 of 78 tested models have no cost per question
  • LFM2.5-2.6B: no verification score (80 of 80 questions failed on the host's side)
  • Ling 3.0 Flash VL: no verification score (not yet run on the current question suite)
  • Qwen3.5-122B-A10B: no verification score (not yet run on the current question suite)
  • Qwen3 Next 80B A3B Thinking: no verification score (not yet run on the current question suite)
  • Nova Premier 1.0: no verification score (not yet run on the current question suite)
  • GPT-5.5 Pro: no verification score (not yet run on the current question suite)

Speed and latency

Median of 3 streamed runs of the same prompt, measured 25 Oct 2026.

Latency vs. output speed

Phi 4 circled · quick to start and fast sits top left

Most attractive quadrant02004006008000.0 s5.0 s10.0 s15.0 s20.0 sTime to first token (s) →Output speed (tok/s) →Nemotron 3.5 Super VLQwen3.5 397B A17BMuse WonderMinistral 3 3B 2512Trinity Large ThinkingGemini 3.5 Flash LiteGemini 3.1 Pro PreviewGemini 3.8 FlashGemini 4 Argon

Results by category

Questions answered correctly in each category

Science12/30
Engineering6/15
Medicine4/11
Economics5/9
Law2/6
Humanities & social science4/9

Compare Phi 4

Head to head with its closest rivals, other Microsoft models and the top models overall.

Hosts and prices

Every host serving this model through OpenRouter, from OpenRouter's catalogue

HostInput / 1MOutput / 1MContextMax outputQuantization
DeepInfra$0.056$0.11216k15kbf16

Phi 4: frequently asked questions

How accurate is Phi 4?
Phi 4 answered 33 of 80 questions correctly (41.2%) in the 2026-10-27 run, ranking #54 of 78 models.
How fast is Phi 4?
Phi 4 generated a median of 103 output tokens per second in the 2026-10-27 run, ranking #35 of 78. Median time to first token was 0.5 s.
What is the latency of Phi 4?
Phi 4 had a median time to first token of 0.5 s in the 2026-10-27 run, ranking #9 of 78 models.
How much does Phi 4 cost?
Phi 4 cost $0.00012 per question on average as billed in the 2026-10-27 run, reasoning tokens included.
What is Phi 4 API pricing?
Phi 4 is listed at $0.07 per million input tokens and $0.14 per million output tokens on OpenRouter.
What is the context window of Phi 4?
Phi 4 accepts up to 16,384 tokens of context, according to OpenRouter's model catalogue.
Is Phi 4 open source?
Yes. Phi 4 is an open-weights model: Microsoft publishes the weights, so you can run it yourself as well as through an API.
Who makes Phi 4?
Phi 4 is made by Microsoft. The LLM Verifications Leaderboard tested it through OpenRouter using the model ID microsoft/phi-4.
Where can I use Phi 4 via API?
Through OpenRouter, Phi 4 is served by 1 host: DeepInfra.