Claude Haiku 4.5 vs Phi 4
Updated October 21, 2026. 33 benchmarks have a result for both. Coding is ranked on the Vibe Coding Index; every other board is shown as published.
Decision reading
Only Claude Haiku 4.5 has a coding score, 35.6 on the Vibe Coding Index; Phi 4 has none, so the two cannot be ranked on coding.
Phi 4 has the edge where output tokens cost 35.7x less.
- Shared results
- 33
- Claude Haiku 4.5 only
- 61
- Phi 4 only
- 12
- Like-for-like categories
- 7 of 9
Shared results can support a head-to-head reading. Results only one model has describe coverage, not superiority.
Which one for your work
Each card is decided by one number both models carry. Where only one has it, the card says so and makes no call.
Coding
Vibe Coding Index, six public coding benchmarks
No call
Only Claude Haiku 4.5 has a coding score, so no call is made.
One side not measured
Intelligence
AA Intelligence Index
Claude Haiku 4.5
Claude Haiku 4.5 leads, 16.9 against 5.9.
Same board, both measured
Agentic work
AA Agentic Index
No call
Only Claude Haiku 4.5 has a AA Agentic Index result, 8.0.
Directional only
Reasoning
GPQA Diamond
No call
Only Claude Haiku 4.5 has a GPQA Diamond result, 72.2%.
Directional only
Math
AIME
No call
Only Claude Haiku 4.5 has a AIME result, 82.7%.
Directional only
Knowledge
MMLU Pro
No call
Only Claude Haiku 4.5 has a MMLU Pro result, 78.7%.
Directional only
Web design
Design Arena: websites
No call
Only Claude Haiku 4.5 has a Design Arena: websites result, 1122.
Directional only
Long documents
Listed context window
Claude Haiku 4.5
Claude Haiku 4.5 takes 200K tokens against 16K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Phi 4
Claude Haiku 4.5 $0.0035, Phi 4 $0.0001 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
No call
Only Claude Haiku 4.5 has the number, $0.56.
Directional only
At a glance
- Coding index
- Claude Haiku 4.5 35.6 from 4 benchmarks, Phi 4 not benchmarked.
- Widest category
- Coding, 36.2 points apart on average over 2 shared benchmarks.
- Widest single gap
- OTIS Mock AIME: Claude Haiku 4.5 66.7% against Phi 4 13.8%, 52.9 points apart on a 0 to 100 scale.
- Output price
- Claude Haiku 4.5 $5.00 against Phi 4 $0.14 per million output tokens, 35.7x apart.
- Context window
- Claude Haiku 4.5 takes 200K tokens, Phi 4 16K.
Benchmark by benchmark
First the score: boards each model leads, per category, where both have a result. Then every board, a category to a fold; bars are each result on a 0 to 100 scale, so compare the two in a row, not one row with another.
| Category | Claude Haiku 4.5 | Bars | Phi 4 | Ties |
|---|---|---|---|---|
| Coding | 2 | 0 | ||
| Intelligence | 2 | 0 | ||
| Agentic work | 1 | 0 | ||
| Reasoning | 4 | 0 | ||
| Math | 4 | 0 | ||
| Knowledge | 4 | 0 | ||
| General | 16 | 0 | ||
| All shared boards | 33 | 0 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index20
| Benchmark | Claude Haiku 4.5 | Phi 4 | Leader |
|---|---|---|---|
| SWE-bench Verified | 66.6% | Not measured | Only Claude Haiku 4.5 measured |
| SWE-bench Multilingual | 64.7% | Not measured | Only Claude Haiku 4.5 measured |
| LiveBench coding | 48.1% | Not measured | Only Claude Haiku 4.5 measured |
| LMArena WebDev | 1330 | Not measured | Only Claude Haiku 4.5 measured |
| AA Coding Index | 43.9 | Not measured | Only Claude Haiku 4.5 measured |
| LiveCodeBench (Vals AI) | 41.2% | Not measured | Only Claude Haiku 4.5 measured |
| SWE-bench Verified (Vals AI) | 66.6% | Not measured | Only Claude Haiku 4.5 measured |
| Vibe Code Bench (Vals AI) | 11.4% | Not measured | Only Claude Haiku 4.5 measured |
| Code migration (Vals AI) | 7.5% | Not measured | Only Claude Haiku 4.5 measured |
| ProgramBench (Vals AI) | 0.0% | Not measured | Only Claude Haiku 4.5 measured |
| Terminal-Bench 2.1 (Vals AI) | 43.8% | Not measured | Only Claude Haiku 4.5 measured |
| LiveCodeBench (Artificial Analysis run) | Not measured | 23.1% | Only Phi 4 measured |
| Terminal-Bench Hard (Artificial Analysis run) | Not measured | 3.8% | Only Phi 4 measured |
| Terminal-Bench | 35.5% | Not measured | Only Claude Haiku 4.5 measured |
| WeirdML | 45.4% | Not measured | Only Claude Haiku 4.5 measured |
| LMArena Text: coding | 1454 | 1231 | Claude Haiku 4.5 |
| LMArena Text: software and IT | 1439 | 1228 | Claude Haiku 4.5 |
| LMArena WebDev: React | 1326 | Not measured | Only Claude Haiku 4.5 measured |
| LMArena WebDev: HTML | 1318 | Not measured | Only Claude Haiku 4.5 measured |
| Neither model has a result on the other 19 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index20
| Benchmark | Claude Haiku 4.5 | Phi 4 | Leader |
|---|---|---|---|
| AA Intelligence Index | 16.9 | 5.9 | Claude Haiku 4.5 |
| Epoch Capabilities Index | 142.2 | 130.5 | Claude Haiku 4.5 |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index10
| Benchmark | Claude Haiku 4.5 | Phi 4 | Leader |
|---|---|---|---|
| AA Agentic Index | 8.0 | Not measured | Only Claude Haiku 4.5 measured |
| LiveBench research | 42.2% | Not measured | Only Claude Haiku 4.5 measured |
| LiveBench browser | 40.0% | Not measured | Only Claude Haiku 4.5 measured |
| LiveBench long agentic | 11.1% | Not measured | Only Claude Haiku 4.5 measured |
| Finance Agent (Vals AI) | 31.0% | Not measured | Only Claude Haiku 4.5 measured |
| Tax Agent Bench (Vals AI) | 5.9% | Not measured | Only Claude Haiku 4.5 measured |
| Legal Research Bench (Vals AI) | 10.6% | Not measured | Only Claude Haiku 4.5 measured |
| Excel Modeling Benchmark (Vals AI) | 22.7% | Not measured | Only Claude Haiku 4.5 measured |
| τ²-Bench Telecom | Not measured | 0.0% | Only Phi 4 measured |
| BALROG | 31.2% | 11.6% | Claude Haiku 4.5 |
| DeepResearch Bench | 45.5% | Not measured | Only Claude Haiku 4.5 measured |
| Vending-Bench 2 | $459 | Not measured | Only Claude Haiku 4.5 measured |
| Neither model has a result on the other 27 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index40
| Benchmark | Claude Haiku 4.5 | Phi 4 | Leader |
|---|---|---|---|
| GPQA Diamond | 72.2% | Not measured | Only Claude Haiku 4.5 measured |
| LiveBench reasoning | 61.7% | Not measured | Only Claude Haiku 4.5 measured |
| MMMU Pro (Vals AI) | 46.1% | Not measured | Only Claude Haiku 4.5 measured |
| GPQA Diamond (Artificial Analysis run) | Not measured | 57.5% | Only Phi 4 measured |
| Humanity's Last Exam (Artificial Analysis run) | Not measured | 3.8% | Only Phi 4 measured |
| GPQA Diamond (Epoch AI run) | 71.2% | 56.1% | Claude Haiku 4.5 |
| ARC-AGI | 47.7% | Not measured | Only Claude Haiku 4.5 measured |
| ARC-AGI-2 | 4.0% | Not measured | Only Claude Haiku 4.5 measured |
| Chess puzzles | 8.0% | Not measured | Only Claude Haiku 4.5 measured |
| LMArena Text: hard prompts | 1421 | 1220 | Claude Haiku 4.5 |
| LMArena Text: expert | 1443 | 1203 | Claude Haiku 4.5 |
| LMArena Text: hard prompts in English | 1435 | 1230 | Claude Haiku 4.5 |
| Neither model has a result on the other 9 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index40
| Benchmark | Claude Haiku 4.5 | Phi 4 | Leader |
|---|---|---|---|
| AIME | 82.7% | Not measured | Only Claude Haiku 4.5 measured |
| MGSM (Vals AI) | 92.2% | Not measured | Only Claude Haiku 4.5 measured |
| AA Math Index | Not measured | 18.0 | Only Phi 4 measured |
| AIME 2025 (Artificial Analysis run) | Not measured | 18.0% | Only Phi 4 measured |
| OTIS Mock AIME | 66.7% | 13.8% | Claude Haiku 4.5 |
| MATH level 5 | 96.4% | 64.9% | Claude Haiku 4.5 |
| LMArena Text: math | 1398 | 1246 | Claude Haiku 4.5 |
| LMArena Text: mathematical work | 1430 | 1254 | Claude Haiku 4.5 |
| Neither model has a result on the other 5 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index40
| Benchmark | Claude Haiku 4.5 | Phi 4 | Leader |
|---|---|---|---|
| MMLU Pro | 78.7% | Not measured | Only Claude Haiku 4.5 measured |
| LegalBench (Vals AI) | 81.2% | Not measured | Only Claude Haiku 4.5 measured |
| CaseLaw (Vals AI) | 56.5% | Not measured | Only Claude Haiku 4.5 measured |
| CorpFin (Vals AI) | 60.6% | Not measured | Only Claude Haiku 4.5 measured |
| TaxEval (Vals AI) | 67.5% | Not measured | Only Claude Haiku 4.5 measured |
| MortgageTax (Vals AI) | 62.2% | Not measured | Only Claude Haiku 4.5 measured |
| MedQA (Vals AI) | 79.6% | Not measured | Only Claude Haiku 4.5 measured |
| MedCode (Vals AI) | 32.7% | Not measured | Only Claude Haiku 4.5 measured |
| MMLU Pro (Artificial Analysis run) | Not measured | 71.4% | Only Phi 4 measured |
| SimpleQA Verified | 13.2% | Not measured | Only Claude Haiku 4.5 measured |
| MMLU | Not measured | 84.8% | Only Phi 4 measured |
| LMArena Text: science | 1406 | 1206 | Claude Haiku 4.5 |
| LMArena Text: medicine and healthcare | 1394 | 1202 | Claude Haiku 4.5 |
| LMArena Text: legal and government | 1400 | 1241 | Claude Haiku 4.5 |
| LMArena Text: business and finance | 1400 | 1199 | Claude Haiku 4.5 |
| Neither model has a result on the other 2 boards here | |||
Web designDirectional onlyPublic boards, shown as published, kept out of the index9 boards, none shared
| Benchmark | Claude Haiku 4.5 | Phi 4 | Leader |
|---|---|---|---|
| Design Arena: websites | 1122 | Not measured | Only Claude Haiku 4.5 measured |
| Design Arena: UI components | 1100 | Not measured | Only Claude Haiku 4.5 measured |
| Design Arena: game dev | 1097 | Not measured | Only Claude Haiku 4.5 measured |
| Design Arena: data viz | 1124 | Not measured | Only Claude Haiku 4.5 measured |
| Design Arena: 3D | 1072 | Not measured | Only Claude Haiku 4.5 measured |
| Design Arena: all code categories | 1115 | Not measured | Only Claude Haiku 4.5 measured |
| Design Arena: SVG | 1026 | Not measured | Only Claude Haiku 4.5 measured |
| Design Arena: ASCII art | 1139 | Not measured | Only Claude Haiku 4.5 measured |
| LiveBench design | 59.9% | Not measured | Only Claude Haiku 4.5 measured |
GeneralLike-for-likePublic boards, shown as published, kept out of the index160
| Benchmark | Claude Haiku 4.5 | Phi 4 | Leader |
|---|---|---|---|
| LiveBench data analysis | 59.3% | Not measured | Only Claude Haiku 4.5 measured |
| Public Benefits Bench (Vals AI) | 54.3% | Not measured | Only Claude Haiku 4.5 measured |
| SAGE (Vals AI) | 31.8% | Not measured | Only Claude Haiku 4.5 measured |
| MedScribe (Vals AI) | 85.2% | Not measured | Only Claude Haiku 4.5 measured |
| Vals Multimodal Index | 42.9% | Not measured | Only Claude Haiku 4.5 measured |
| IFBench | Not measured | 23.5% | Only Phi 4 measured |
| Long context reasoning (AA-LCR) | Not measured | 0.0% | Only Phi 4 measured |
| DTBench | 73.6% | Not measured | Only Claude Haiku 4.5 measured |
| LMCA | 30.9% | Not measured | Only Claude Haiku 4.5 measured |
| Creative writing (Lech Mazur) | Not measured | 62.6% | Only Phi 4 measured |
| ExploitBench | 13.7% | Not measured | Only Claude Haiku 4.5 measured |
| LMArena Text | 1397 | 1217 | Claude Haiku 4.5 |
| LMArena Text: instruction following | 1414 | 1201 | Claude Haiku 4.5 |
| LMArena Text: creative writing | 1371 | 1181 | Claude Haiku 4.5 |
| LMArena Text: English | 1414 | 1231 | Claude Haiku 4.5 |
| LMArena Text: non-English | 1377 | 1197 | Claude Haiku 4.5 |
| LMArena Text: Chinese | 1418 | 1212 | Claude Haiku 4.5 |
| LMArena Text: German | 1374 | 1222 | Claude Haiku 4.5 |
| LMArena Text: Japanese | 1338 | 1158 | Claude Haiku 4.5 |
| LMArena Text: Korean | 1346 | 1151 | Claude Haiku 4.5 |
| LMArena Text: French | 1411 | 1224 | Claude Haiku 4.5 |
| LMArena Text: Spanish | 1421 | 1234 | Claude Haiku 4.5 |
| LMArena Text: Russian | 1381 | 1209 | Claude Haiku 4.5 |
| LMArena Text: Polish | 1343 | Not measured | Only Claude Haiku 4.5 measured |
| LMArena Text: multi-turn | 1409 | 1206 | Claude Haiku 4.5 |
| LMArena Text: longer queries | 1427 | 1217 | Claude Haiku 4.5 |
| LMArena Text: writing and language | 1386 | 1198 | Claude Haiku 4.5 |
| LMArena Text: entertainment and media | 1369 | 1162 | Claude Haiku 4.5 |
| LMArena Document | 1420 | Not measured | Only Claude Haiku 4.5 measured |
| Neither model has a result on the other 17 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | Claude Haiku 4.5 | Phi 4 | Leader |
|---|---|---|---|
| Neither model has a result on 5 boards here | |||
What each workload costs
Three fixed token mixes at each model's listed price, so per-token rates become a bill.
Chat turn
1K input + 500 output tokens
Phi 4 costs less
Repository review
50K input + 3K output tokens
Phi 4 costs less
Agent loop
200K input + 10K output tokens
Phi 4 costs less
Listed OpenRouter rates, without cache or batch discounts. A workload that does not fit a model's context window would need splitting.
Specification differences
The leading value in a row is marked; a row with a missing value names no leader. Our two rows are marked in red.
| Specification | Anthropic Claude Haiku 4.5 | Microsoft Phi 4 |
|---|---|---|
| Maker | Anthropic | Microsoft |
| Weights | No public weights listed | Open weights |
| Released | October 15, 2025 | January 10, 2025 |
| Context window | 200K | 16K |
| Input $/Mtok | $1.00 | $0.07 |
| Output $/Mtok | $5.00 | $0.14 |
| Coding index | 35.6 | Not benchmarked |
| Independent measurements behind it | 4 | 0 |
| Cost per solved task | $0.56 | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is Claude Haiku 4.5 better than Phi 4 for coding?
Only Claude Haiku 4.5 has a coding score, 35.6 on the Vibe Coding Index; Phi 4 has none, so the two cannot be ranked on coding. Phi 4 has the edge where output tokens cost 35.7x less.
Which is cheaper, Claude Haiku 4.5 or Phi 4?
Claude Haiku 4.5 costs $1.00 per million input tokens and $5.00 per million output tokens; Phi 4 costs $0.07 and $0.14. Cost per solved task is known for Claude Haiku 4.5 only.
Which has the bigger context window, Claude Haiku 4.5 or Phi 4?
Claude Haiku 4.5, with 200K tokens against 16K.
How many benchmarks is the Claude Haiku 4.5 vs Phi 4 comparison based on?
33 benchmarks have a result for both models as of October 21, 2026. Claude Haiku 4.5's coding index rests on 4 independent benchmarks and Phi 4's on 0. Where only one model has a result, the page says so instead of naming a winner.