vs
0:3
overall capability wins (1 ties across 4 scored). Claude Sonnet 4 leads overall.
- Coding0:1
- Arena0:1
- Pricing3:0
- Speed2:0
- Specs0:1
Verdict
Claude Sonnet 4 leads coding (widest gap: +36.9 on LiveCode); Claude Sonnet 4 ranks higher on LMArena (+129 Elo); Claude Sonnet 4 allows ~15.6x more max output (64K vs 4K).
- Claude Sonnet 4 leads coding (widest gap: +36.9 on LiveCode)
- Claude Sonnet 4 ranks higher on LMArena (+129 Elo)
- Claude Sonnet 4 allows ~15.6x more max output (64K vs 4K)
- Claude 3 Haiku is ~12.0x cheaper on a blended token basis than Claude Sonnet 4
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Claude 3 Haiku
- $0.56
- Claude Sonnet 4
- $6.75
Claude 3 Haiku is about 12.0x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
Anthropic
Anthropic
Family
Claude Sonnet / Haiku
Claude Sonnet / Haiku
License
Proprietary
Proprietary
Open weights
No
No
Release
Mar 13, 2024
May 22, 2025
Knowledge cutoff
-
-
API / provider
Anthropic
Anthropic
Modalities
text, image → text
text, image → text
Specs
Context window
200K
200K
Max output
4K
64K
Parameters
—
—
Pricing
Input $/1M
$0.25
$3
Output $/1M
$1.25
$15
Blended $/1M (3∶1)
$0.50
$6
Speed
tok/s
140
95
TTFT (s)
0.2
0.35
Reasoning
MMLU-Pro
—
—
GPQA Diamond
37.4
—
Humanity's Last Exam
4.1
—
AIME 2025
—
—
MATH-500
39.4
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
17.6
—
AA-LCR v1.1
27.7
—
CritPt
—
—
MMMU-Pro
30.8
—
IFBench
36.1
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
42.7
SWE-bench Multilingual
—
—
LiveCodeBench
22.5
59.4
Terminal-Bench 2.1
—
—
Aider Polyglot
—
56.4
Terminal-Bench 3
—
—
BigCodeBench
39.4
—
SciCode
18.6
—
CursorBench
—
—
SWE-Rebench
—
44.6
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,261
1,391
| Metric | Claude 3 Haiku | Claude Sonnet 4 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Anthropic | Anthropic | — |
| Family | Claude Sonnet / Haiku | Claude Sonnet / Haiku | — |
| License | Proprietary | Proprietary | — |
| Open weights | No | No | — |
| Release | Mar 13, 2024 | May 22, 2025 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Anthropic | Anthropic | — |
| Modalities | text, image → text | text, image → text | — |
| Specs | |||
| Context window | 200K | 200K | tie |
| Max output | 4K | 64K | B +60K |
| Parameters | — | — | — |
| Pricing | |||
| Input $/1M | $0.25 | $3 | A +$2.75 |
| Output $/1M | $1.25 | $15 | A +$13.75 |
| Blended $/1M (3∶1) | $0.50 | $6 | A +$5.50 |
| Speed | |||
| tok/s | 140 | 95 | A +45 |
| TTFT (s) | 0.2 | 0.35 | A +0.15 |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 37.4 | — | — |
| Humanity's Last Exam | 4.1 | — | — |
| AIME 2025 | — | — | — |
| MATH-500 | 39.4 | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 17.6 | — | — |
| AA-LCR v1.1 | 27.7 | — | — |
| CritPt | — | — | — |
| MMMU-Pro | 30.8 | — | — |
| IFBench | 36.1 | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | 42.7 | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | 22.5 | 59.4 | B +36.9 pts |
| Terminal-Bench 2.1 | — | — | — |
| Aider Polyglot | — | 56.4 | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | 39.4 | — | — |
| SciCode | 18.6 | — | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | 44.6 | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,261 | 1,391 | B +129 Elo |
Claude 3 Haiku
- Context
- 200K / 4K out
- Parameters
- —
- Price
- $0.25 / $1.25
- Speed
- 140 tok/s · 0.2s TTFT
- Modalities
- text, image → text
- License
- Proprietary
Original Claude 3 speed tier — cheap long-context chat and classification before Haiku 3.5 / 4.5.
Claude Sonnet 4
- Context
- 200K / 64K out
- Parameters
- —
- Price
- $3 / $15
- Speed
- 95 tok/s · 0.35s TTFT
- Modalities
- text, image → text
- License
- Proprietary
Prior Sonnet generation still available on Bedrock and Google Cloud after Anthropic API retirement of the primary SKU.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.