Models / Phi-4

Microsoft

Phi-4

Phi-4 is best for sticking to the facts; a strong pick at a modest price. The cheapest model of similar intelligence is Gemma 3 4B, at $0.04 per million tokens against $0.09.

Results 16 Nov 2026 · 28 results · 5 sources

Test Phi-4 on your prompt → It runs beside the models you could use instead, with every answer checked and costed.

Price per million tokens
$0.07 in · $0.14 out
Typical host on OpenRouter, 10 Nov 2026 · compare prices
Speed
45 tokens a second; on its test prompt the answer starts after 1.0 s, thinking time included
Artificial Analysis, on its usual API
Intelligence Index
5.9
Artificial Analysis
Spring Prompt overall
Not ranked yet
needs our benchmarks and two groups of results
Developer
Microsoft, based in the US
Weights published
Released
12 Dec 2024 (Artificial Analysis)
16,384 tokens of context

Where it stands among the models you could choose

Available models grouped by price, intelligence, speed and developer location, plus models with published weights. Retired models are excluded.

AmongIntelligence IndexSpring Prompt overallPrice, blendedSpeed
Models available today 175th of 185Claude Opus 5.5 – $0.09 blended, 5th percentilecheapest: Mistral Nemo 61st of 65Nemotron 3.5 Super VL
Models priced under $0.25 per million tokens 32nd of 41Claude Haiku 5.5 – sets this group 19th of 19Nemotron 3.5 Super VL
Models with similar intelligence sets this group – $0.09 blended, 12th percentilecheapest: Gemma 3 4B 13th of 13Nova Micro 1.0
Models writing under 50 tokens a second 9th of 9MiMo-V2.6-Pro – $0.09 blended, 0th percentilethe cheapest sets this group
Models from developers based in the US 85th of 92Claude Opus 5.5 – $0.09 blended, 7th percentilecheapest: Gemma 3 4B 33rd of 33Nemotron 3.5 Super VL
Open-weights models 95th of 104MiMo-V2.6-Pro – $0.09 blended, 7th percentilecheapest: Mistral Nemo 40th of 43Nemotron 3.5 Super VL

The name under each rank is the leader of that group; green is the top quarter of the group and red the bottom quarter, by rank. Price is shown as a percentile: the share of the group that costs less, so lower is cheaper. "Sets this group" marks the measure the group is defined by. Spring Prompt overall is our score out of 100 across every source (how it works). Price is the developer's list price (or the typical OpenRouter host where we haven't read one) for three input tokens to one output; intelligence and speed are from Artificial Analysis (data sourced from Artificial Analysis); "similar intelligence" means within 4 points on its Intelligence Index. Developer locations are where each company is based, not where a model is served. Groups under 3 models are not ranked.

How good is it, and for what?

Each use case ranks the available models its benchmarks measured, then those priced under $0.25 per million tokens. The bar shows its percentile in that field, best to the right. Open a row for the results behind it.

Product listingsTurning a sparse product feed and photos into listings that can go live Not measured
  • CatalogBench: has not measured this model.
Decks from an analysisTurning a finished analysis into a deck you could present as it is Not measured
  • DeckBench: has not measured this model.
User surveysPlanning a user survey and reading its results without being misled Not measured
  • SurveyBench: has not measured this model.
Marketing planningPlanning a year of ad spend without overspending Not measured
  • ROASBench: has not measured this model.
Agents and tool useMulti-step tasks with tools: support desks, coding agents, function calls 115th of 159 available · 14th of 37 at its price
BenchmarkPhi-4Rank among availableBest available
Berkeley Function Calling Leaderboard (BFCL) V4BFCL: overall accuracy · reported 28.8% 27th of 36 Claude Opus 4.5 77.5%
  • tau2-bench: has not measured this model.
  • OpenHands Index: has not measured this model.
  • Microsoft STATE-Bench: has not measured this model.
  • Artificial Analysis: has not measured this model.
  • Vending-Bench 2: has not measured this model.
Professional workReal tasks from banking, consulting and law, business documents and freelance projects Not measured
  • APEX-Agents: has not measured this model.
  • GDP.pdf: has not measured this model.
  • Remote Labor Index: has not measured this model.
Reasoning and knowledgeHard questions across science, maths and general knowledge 169th of 187 available · 31st of 41 at its price
BenchmarkPhi-4Rank among availableBest available
Artificial AnalysisArtificial Analysis: Artificial Analysis Intelligence Index · reported 5.9 175th of 185 Claude Opus 5.5 57.6
Artificial AnalysisArtificial Analysis: Humanity's Last Exam · reported 3.8% 172nd of 186 Claude Opus 5.5 61.4%
Artificial AnalysisArtificial Analysis: GPQA Diamond · reported 57.5% 138th of 168 GPT-6 Astra 96.3%
WritingWhat people prefer in blind comparisons, and judged writing quality 161st of 169 available · 33rd of 39 at its price
BenchmarkPhi-4Rank among availableBest available
Arena (formerly LMArena)Arena Text: overall · reported 1,256 147th of 153 Gemini 4 Argon 1,525
UGI LeaderboardUGI: writing score · reported 25.7 114th of 124 Gemini 3.8 Flash 78.6
Sticking to the factsSummarising without inventing things, and factual answers 4th of 139 available · 2nd of 30 at its price
BenchmarkPhi-4Rank among availableBest available
Vectara Hallucination LeaderboardVectara: hallucination rate · reported 3.7% 3rd of 71 GPT-5.4 nano 3.1%
  • Arena (formerly LMArena): has not measured this model.
  • SimpleQA Verified (Epoch AI): has not measured this model.
Speed under pressureGood decisions against a real clock (fast chess) Not measured
  • BulletBench: has not measured this model.

Nearest alternatives

Against Phi-4 at $0.09 blended, Intelligence Index 5.9. "Similar intelligence" means within 4 points or better.

Where to run it

1 host on OpenRouter at the standard service tier, with a median of $0.09 blended per million tokens; quantisation where the host states it: BF16. Cheapest first.

HostQuantisationInputOutputBlended
DeepInfraBF16$0.07$0.14$0.09

Other service tiers, not counted above: DeepInfra flex $0.06 in, $0.11 out. Flex and batch are cheaper but slower or queued; priority and fast cost more.

Prices per million tokens from OpenRouter's public endpoint list, checked 10 Nov 2026.

Against the alternatives

The models you would most likely weigh it against: the leaders of the groups above.

Scroll sideways to see every alternative.

Phi-4Claude Haiku 5.5
most intelligent at under $0.25 per million tokens
Gemma 3 4B
cheapest with similar intelligence
Nova Micro 1.0
fastest with similar intelligence
Claude Opus 5.5
most intelligent available today
Price per million tokens $0.09 $0.20$0.04$0.06$8.00
Tokens a second 45 240–26096
Intelligence Index 5.9 43.44.85.957.6
Spring Prompt overall – 40––82
Rank among available models, by use case
Product listings – 20th––13th
Decks from an analysis – 18th––7th
User surveys – 7th––4th
Marketing planning – 30th––9th
Agents and tool use 115th of 159 28th158th152nd3rd
Professional work – –––3rd
Reasoning and knowledge 169th of 187 25th181st175th1st
Writing 161st of 169 –157th166th5th
Sticking to the facts 4th of 139 –32nd19th3rd
Speed under pressure – 21st–––

Shaded figures are better than Phi-4's. A dash means no result.

Where it has been measured

Cheapest of the 9 models writing under 50 tokens a second. 32nd most intelligent of the 41 models priced under $0.25 per million tokens. Among models available today, in the top quarter (by percentile) for sticking to the facts; in the bottom quarter for reasoning and knowledge, and writing.

Every result

28 published results from 5 sources, each in the source's own units, with the configuration that produced it.

Arena (formerly LMArena) · reported by Arena (formerly LMArena) · 6 results
MeasureValueRankConfigurationDated
Arena Text: business, management and financeArena rating, higher is better 1,256
1,244–1,268
308 of 406 Phi-4 2 Oct 2026
Arena Text: creative writingArena rating, higher is better 1,210
1,201–1,220
316 of 411 Phi-4 2 Oct 2026
Arena Text: expert promptsArena rating, higher is better 1,268
1,251–1,285
286 of 363 Phi-4 2 Oct 2026
Arena Text: instruction followingArena rating, higher is better 1,245
1,239–1,252
311 of 413 Phi-4 2 Oct 2026
Arena Text: overallArena rating, higher is better 1,256
1,252–1,261
317 of 413 Phi-4 2 Oct 2026
Arena Text: writing, literature and languageArena rating, higher is better 1,229
1,222–1,237
316 of 412 Phi-4 2 Oct 2026

Not shown: Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.

Artificial Analysis · reported by Artificial Analysis · 9 results
MeasureValueRankConfigurationDated
Artificial Analysis: Artificial Analysis Intelligence Indexindex score, higher is better 5.9 399 of 413 Phi-4 20 Oct 2026
Artificial Analysis: GPQA Diamond% of questions, higher is better 57.5% 303 of 360 Phi-4 20 Oct 2026
Artificial Analysis: Humanity's Last Exam% of questions, higher is better 3.8% 389 of 419 Phi-4 20 Oct 2026
Artificial Analysis: IFBench% of instructions, higher is better 23.5% 284 of 288 Phi-4 20 Oct 2026
Artificial Analysis: Long-context reasoning (AA-LCR)% of questions, higher is better 0.0% 390 of 407 Phi-4 20 Oct 2026
Artificial Analysis: Output speedtokens per second, higher is better 45.1 122 of 137 Phi-4 20 Oct 2026
Artificial Analysis: Terminal-Bench Hard% of tasks, higher is better 3.8% 244 of 282 Phi-4 20 Oct 2026
Artificial Analysis: Time to first answer tokenseconds, lower is better 1.03 28 of 137 Phi-4 20 Oct 2026
Artificial Analysis: τ²-bench telecom% of tasks, higher is better 0.0% 281 of 286 Phi-4 20 Oct 2026

Not shown: Not business work, and a blend: read the parts for any one task.

Berkeley Function Calling Leaderboard (BFCL) V4 · reported by Berkeley Function Calling Leaderboard (BFCL) V4 · 8 results
MeasureValueRankConfigurationDated
BFCL: irrelevance detection% correct, higher is better 87.6% 18 of 109 Phi-4 (bfcl prompt) 16 Dec 2025
BFCL: memory% correct, higher is better 24.7% 31 of 109 Phi-4 (bfcl prompt) 16 Dec 2025
BFCL: multi-turn tasks% correct, higher is better 3.9% 84 of 109 Phi-4 (bfcl prompt) 16 Dec 2025
BFCL: overall accuracy% correct, higher is better 28.8% 70 of 109 Phi-4 (bfcl prompt) 16 Dec 2025
BFCL: relevance detection% correct, higher is better 50.0% 98 of 109 Phi-4 (bfcl prompt) 16 Dec 2025
BFCL: single-turn calls (curated)% correct, higher is better 69.6% 88 of 109 Phi-4 (bfcl prompt) 16 Dec 2025
BFCL: single-turn calls (user-contributed)% correct, higher is better 60.7% 85 of 109 Phi-4 (bfcl prompt) 16 Dec 2025
BFCL: web search% correct, higher is better 4.5% 65 of 109 Phi-4 (bfcl prompt) 16 Dec 2025

Not shown: Not general refusal or safety behaviour.

UGI Leaderboard · reported by UGI Leaderboard · 3 results
MeasureValueRankConfigurationDated
UGI: requested-length error% off the requested word count, lower is better 19.0% 171 of 379 Phi-4 16 Nov 2025
UGI: style adherencescore from 0 to 1, higher is better 0.34 218 of 379 Phi-4 16 Nov 2025
UGI: writing scorescore out of 100, higher is better 25.7 309 of 379 Phi-4 16 Nov 2025

Not shown: Not other format limits such as character counts or bullet counts.

Vectara Hallucination Leaderboard · reported by Vectara Hallucination Leaderboard · 2 results
MeasureValueRankConfigurationDated
Vectara: answer rate% of documents, higher is better 80.7% 104 of 108 Phi-4 22 Sep 2026
Vectara: hallucination rate% of summaries, lower is better 3.7% 4 of 108 Phi-4 22 Sep 2026

Not shown: Not a quality score: a low rate usually means content filters were triggered, and hallucination rates are measured on answered documents only.

Will Phi-4 work for your product?

Benchmarks aren't your task. Paste the prompt your product uses and run it beside the models you could use instead, with every answer checked and costed.

New here? Your email starts a free trial. No credit card needed.