Claude Sonnet 4 vs Claude Sonnet 4.5
Wins 1 of 7 areas
Facts
Wins 5 of 7 areas
Coding · Agents · Reasoning · Images and charts · Long documents
Claude Sonnet 4.5 is the stronger all-rounder.Claude Sonnet 4 is better at facts.
Scores updated · 59 tests both models report · How we compare
Where each one wins
Tests won in each of the seven areas where both have results. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- CodingWriting and fixing software06Claude Sonnet 4.56 of 7 tests · 1 tie
- ReasoningHard problems that need careful thinking04Claude Sonnet 4.54 of 5 tests · 1 tie
- AgentsCarrying out multi-step tasks on its own04Claude Sonnet 4.54 of 4 tests
- Images and chartsUnderstanding pictures, charts and video02Claude Sonnet 4.52 of 2 tests
- Long documentsFinding answers in very long texts01Claude Sonnet 4.51 of 1 test
- FactsGetting facts right instead of making them up21Claude Sonnet 42 of 3 tests
- Following instructionsDoing exactly what it is asked11Even1 each
Long documents rests on a single test.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where Claude Sonnet 4.5 pulls ahead
- Command-line tasks in a real terminalTerminal-Bench 2.1+19.5points ahead
- Completes tasks by operating a computer desktopOSWorld-Verified+19.2points ahead
- Recent programming contest problemsLiveCodeBench v6+15.5points ahead
Where Claude Sonnet 4 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+20points ahead
- Keeps track of context across a multi-turn chatMulti-Challenge+1.8points ahead
Every test, side by side
All 59 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingClaude Sonnet 4.5
- Terminal-Bench 2.1Claude Sonnet 4.5 by 19.536.355.8+19.5
- LiveCodeBench v6Claude Sonnet 4.5 by 15.548.564+15.5
- SWE-bench MultilingualClaude Sonnet 4.5 by 10.156.967+10.1
- SciCodeClaude Sonnet 4.5 by 5.74045.7+5.7
- SWE-bench VerifiedClaude Sonnet 4.5 by 4.572.777.2+4.5
- Terminal-Bench HardClaude Sonnet 4.5 by 4.531.135.6+4.5
- SWE-bench Protie42.743.6tie
AgentsClaude Sonnet 4.5
- OSWorld-VerifiedClaude Sonnet 4.5 by 19.242.261.4+19.2
- BrowseCompClaude Sonnet 4.5 by 11.912.224.1+11.9
- GDPValClaude Sonnet 4.5 by 10.89.720.5+10.8
- τ-Bench V3 · BankingClaude Sonnet 4.5 by 7.816.724.5+7.8
ReasoningClaude Sonnet 4.5
- SimpleBenchClaude Sonnet 4.5 by 8.845.554.3+8.8
- ARC-AGI-2Claude Sonnet 4.5 by 7.75.913.6+7.7
- Humanity's Last ExamClaude Sonnet 4.5 by 7.110.717.8+7.1
- GPQA DiamondClaude Sonnet 4.5 by 5.777.783.4+5.7
- CritPttie0.31.1tie
FactsClaude Sonnet 4
- AA-Omniscience · Non-hallucinationClaude Sonnet 4 by 2070.950.9+20
- AA-Omniscience · AccuracyClaude Sonnet 4.5 by 10.222.732.9+10.2
- Vectara HHEM hallucination ratelower is betterClaude Sonnet 4 by 1.710.312+1.7
Images and chartsClaude Sonnet 4.5
Following instructionsEven
- IFBenchClaude Sonnet 4.5 by 2.654.757.3+2.6
- Multi-ChallengeClaude Sonnet 4 by 1.857.155.3+1.8
Other results35 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- HMMT 2025Claude Sonnet 4.5 by 58.715.974.6+58.7
- τ²-BenchClaude Sonnet 4.5 by 52.845.298+52.8
- ARC-AGI-1Claude Sonnet 4.5 by 23.74063.7+23.7
- FinSearchComp-globalClaude Sonnet 4.5 by 18.84260.8+18.8
- AIME 2025Claude Sonnet 4.5 by 16.570.587+16.5
- Artificial Analysis Coding IndexClaude Sonnet 4.5 by 14.537.652.1+14.5
- Terminal-BenchClaude Sonnet 4.5 by 14.535.550+14.5
- τ²-Bench Telecom (AA run)Claude Sonnet 4.5 by 13.564.678.1+13.5
- BrowseComp-ZHClaude Sonnet 4.5 by 13.329.142.4+13.3
- HLE (with tools)Claude Sonnet 4.5 by 11.720.332+11.7
- τ²-Bench (Retail)Claude Sonnet 4.5 by 11.27586.2+11.2
- TAU-bench (airline)Claude Sonnet 4.5 by 106070+10
- Multi-SWE-BenchClaude Sonnet 4.5 by 8.635.744.3+8.6
- HarmBenchClaude Sonnet 4 by 69892+6
- TAU-bench (retail)Claude Sonnet 4.5 by 5.780.586.2+5.7
- ArtifactsBenchClaude Sonnet 4.5 by 4.257.361.5+4.2
- frontiermath_tier_4_v1Claude Sonnet 4.5 by 4.204.2+4.2
- theagentcompanyClaude Sonnet 4.5 by 43741+4
- MMLU-ProClaude Sonnet 4.5 by 3.883.787.5+3.8
- vectara_answer_rateClaude Sonnet 4 by 398.695.6+3
- GAIA (text only)Claude Sonnet 4.5 by 2.968.371.2+2.9
- MMMLUClaude Sonnet 4.5 by 2.686.589.1+2.6
- LiveCodeBenchClaude Sonnet 4.5 by 2.568.571+2.5
- bbqClaude Sonnet 4.5 by 2.396.698.9+2.3
- MMLU-ReduxClaude Sonnet 4.5 by 293.695.6+2
- AA IntelligenceClaude Sonnet 4.5 by 1.818.920.7+1.8
- vectara_avg_summary_lengthClaude Sonnet 4 by 18 rating points145.8127.8+18 rating
- vectara_factual_consistencyClaude Sonnet 4 by 1.789.788+1.7
- AIR-Bench 2024Claude Sonnet 4.5 by 1.588.389.8+1.5
- xbench-DeepSearchClaude Sonnet 4.5 by 1.464.666+1.4
- anthropic_red_teamClaude Sonnet 4 by 199.298.2+1
- AA-Omnisciencetie0.2-0.1tie
- XSTesttie96.696.4tie
- AA Agentic Indextie17.617.5tie
- simple_safety_teststie100100tie
Questions people ask
Which is better, Claude Sonnet 4 or Claude Sonnet 4.5?
Claude Sonnet 4.5 wins five of the seven areas where both have results: coding, agents, reasoning, images and charts and long documents. Claude Sonnet 4 wins facts. They are level on following instructions.
Which is better for coding?
Claude Sonnet 4.5. It wins 6 of the 7 coding tests both models report; Claude Sonnet 4 wins none, and 1 is a tie.
Which is cheaper?
Claude Sonnet 4 costs $3.00 per million input tokens and $15.00 per million output tokens; Claude Sonnet 4.5 costs $3.00 and $15.00. For a million tokens read plus a million written, they cost about the same.
How do you compare the two?
We use the 59 benchmark tests both models have published scores on. The verdict counts the 24 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 35 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.