Skip to main content
BenchLM
Advertise

Model comparison

Claude Sonnet 4.5 vs GPT-5.2

Data verified

Head-to-head evidence from 9 shared benchmark results across 6 categories. Overall scores shown here use the public BenchAlign v5 ranking lane.

53.61/100
Margin
4.8pts
winning →
GPT-5.2
OpenAI
58.43/100
1 category wins4 category wins

Public leaderboard positions: Claude Sonnet 4.5 #96 (Estimated); GPT-5.2 #64 (Estimated). Intervals and evidence labels describe ranking uncertainty, not a guarantee for a specific workload.

Evidence parity. Claude Sonnet 4.5 and GPT-5.2 share 9 comparable benchmark results. 5 of 8 categories are comparable. 3 results are unique to Claude Sonnet 4.5; 19 to GPT-5.2.

Updated July 20, 2026
Shared results
9
Claude Sonnet 4.5 only
3
GPT-5.2 only
19
Comparable categories
5 / 8

Pick GPT-5.2 if you want the stronger benchmark profile. Claude Sonnet 4.5 only becomes the better choice if coding is the priority or you would rather avoid the extra latency and token burn of a reasoning model.

Confidence note. This is a partial-evidence comparison with 9 shared benchmark results across 6 evidence categories; 5 of 8 categories currently have scoreable aggregates for both models. Treat the verdict as directional until coverage is more balanced.

Why this result

GPT-5.2 is clearly ahead on the BenchAlign aggregate, 58.43 to 53.61. The gap is large enough that you do not need to squint at the spreadsheet to see the difference.

GPT-5.2's sharpest advantage is in reasoning, where it averages 52.9 against 13.6. The single biggest benchmark swing on the page is ARC-AGI-2, 13.6% to 52.9%. Claude Sonnet 4.5 does hit back in coding, so the answer changes if that is the part of the workload you care about most.

Claude Sonnet 4.5 is also the more expensive model on tokens at $3.00 input / $15.00 output per 1M tokens, versus $1.75 input / $14.00 output per 1M tokens for GPT-5.2. GPT-5.2 is the reasoning model in the pair, while Claude Sonnet 4.5 is not. That usually helps on harder chain-of-thought-heavy tests, but it can also mean more latency and more token spend in real use. GPT-5.2 gives you the larger context window at 400K, compared with 200K for Claude Sonnet 4.5.

Category breakdown

Exact category averages are shown below. Not measured means BenchLM does not have enough sourced public coverage for that model and category.

Category scores and score margins for Claude Sonnet 4.5 and GPT-5.2
CategoryClaude Sonnet 4.5ΔGPT-5.2
ReasoningClaude Sonnet 4.513.6Margin→ 39.3GPT-5.252.9
MathClaude Sonnet 4.511.2Margin→ 24.0GPT-5.235.2
KnowledgeClaude Sonnet 4.583.4Margin→ 9.0GPT-5.292.4
CodingClaude Sonnet 4.577.2Margin← 6.6GPT-5.270.6
AgenticClaude Sonnet 4.555.4Margin→ 0.3GPT-5.255.7
MultimodalClaude Sonnet 4.5Not measuredMarginNo overlapGPT-5.280.4

Decisive benchmark drivers

The largest measured benchmark gaps in this matchup, with exact reported values.

More
A · Claude Sonnet 4.5B · GPT-5.2

  1. ARC-AGI-2

    Reasoning
    Source ↗
    A 13.6%B 52.9%
    Winner: GPT-5.2Δ 39.3
    ARC-AGI-2: Claude Sonnet 4.5 scored 13.6%; GPT-5.2 scored 52.9%. GPT-5.2 wins this benchmark.

  2. FrontierMath v2 (Tiers 1-3)

    Math
    Source ↗
    A 13.495%B 40.700%
    Winner: GPT-5.2Δ 27.2
    FrontierMath v2 (Tiers 1-3): Claude Sonnet 4.5 scored 13.495%; GPT-5.2 scored 40.700%. GPT-5.2 wins this benchmark.

  3. FrontierMath v2 (Tier 4)

    Math
    Source ↗
    A 4.167%B 18.800%
    Winner: GPT-5.2Δ 14.6
    FrontierMath v2 (Tier 4): Claude Sonnet 4.5 scored 4.167%; GPT-5.2 scored 18.800%. GPT-5.2 wins this benchmark.

  4. OSWorld-Verified

    Agentic
    Source ↗
    A 61.4%B 47.3%
    Winner: Claude Sonnet 4.5Δ 14.1
    OSWorld-Verified: Claude Sonnet 4.5 scored 61.4%; GPT-5.2 scored 47.3%. Claude Sonnet 4.5 wins this benchmark.

  5. GPQA

    Knowledge
    Source ↗
    A 83.4%B 92.4%
    Winner: GPT-5.2Δ 9
    GPQA: Claude Sonnet 4.5 scored 83.4%; GPT-5.2 scored 92.4%. GPT-5.2 wins this benchmark.

Operational comparison

Runtime and commercial metrics are compared only when both models have a complete sourced value.

MetricClaude Sonnet 4.5GPT-5.2Comparison
Input / output priceUSD per 1M tokensClaude Sonnet 4.5$3 input / $15 outputGPT-5.2$1.75 input / $14 outputGPT-5.2 has the lower combined listed price.
Generation speedtokens per secondClaude Sonnet 4.5Not availableGPT-5.273 tok/sA complete speed comparison is not available.
First-answer latencyseconds to first tokenClaude Sonnet 4.5Not availableGPT-5.2130.34 sA complete latency comparison is not available.
Context windowmaximum listed tokensClaude Sonnet 4.5200KGPT-5.2400KGPT-5.2 lists the larger context window.

Benchmark Deep Dive

AgenticGPT-5.2 wins
BenchmarkClaude Sonnet 4.5GPT-5.2Result
Terminal-Bench 2.0Source 50%—Not comparable
OSWorld-VerifiedSource 61.4%47.3%Claude Sonnet 4.5 leads
VITA-BenchSource 17.0%—Not comparable
Gert LabsSource 48.51%46.54%Claude Sonnet 4.5 leads
JobBenchSource 27.7%34.3%GPT-5.2 leads
BrowseCompSource —65.8%Not comparable
τ²-bench resultsSource —84.8%Not comparable
CodingClaude Sonnet 4.5 wins
BenchmarkClaude Sonnet 4.5GPT-5.2Result
SWE-bench VerifiedSource 77.2%80%GPT-5.2 leads
SWE-bench ProSource —55.6%Not comparable
Vibe Code BenchSource —53.50%Not comparable
AA-SciCodeSource —52.1%Not comparable
ReasoningGPT-5.2 wins
BenchmarkClaude Sonnet 4.5GPT-5.2Result
ARC-AGI-2Source 13.6%52.9%GPT-5.2 leads
AA-LCRSource —72.7%Not comparable
CritPtSource —11.6%Not comparable
KnowledgeGPT-5.2 wins
BenchmarkClaude Sonnet 4.5GPT-5.2Result
GPQASource 83.4%92.4%GPT-5.2 leads
Artificial Analysis Intelligence IndexSource —42.2%Not comparable
AA-GPQA DiamondSource —90.3%Not comparable
AA-HLESource —35.4%Not comparable
AA-Omniscience IndexSource —-1.0%Not comparable
AA-Omniscience AccuracySource —43.8%Not comparable
AA-Omniscience Hallucination RateSource —79.7%Not comparable
MathGPT-5.2 wins
BenchmarkClaude Sonnet 4.5GPT-5.2Result
AIME 2025Source 87%—Not comparable
FrontierMath v2 (Tiers 1-3)Source 13.495%40.700%GPT-5.2 leads
FrontierMath v2 (Tier 4)Source 4.167%18.800%GPT-5.2 leads
AA AIME 2025Source —99.0%Not comparable
Multimodal
BenchmarkClaude Sonnet 4.5GPT-5.2Result
Design Arena WebsiteSource 12221227GPT-5.2 leads
MMMU-ProSource —79.5%Not comparable
MathVisionSource —83.0%Not comparable
CharXivSource —82.1%Not comparable
V*Source —75.9%Not comparable
Inst. Following
BenchmarkClaude Sonnet 4.5GPT-5.2Result
AA-IFBenchSource —75.4%Not comparable
Frequently Asked Questions (6)

Which is better, Claude Sonnet 4.5 or GPT-5.2?

GPT-5.2 is ahead on BenchLM's BenchAlign leaderboard, 58.43 to 53.61. The biggest single separator in this matchup is ARC-AGI-2, where the scores are 13.6% and 52.9%.

Which is better for knowledge tasks, Claude Sonnet 4.5 or GPT-5.2?

GPT-5.2 has the edge for knowledge tasks in this comparison, averaging 92.4 versus 83.4. Inside this category, GPQA is the benchmark that creates the most daylight between them.

Which is better for coding, Claude Sonnet 4.5 or GPT-5.2?

Claude Sonnet 4.5 has the edge for coding in this comparison, averaging 77.2 versus 70.6. Inside this category, SWE-bench Verified is the benchmark that creates the most daylight between them.

Which is better for math, Claude Sonnet 4.5 or GPT-5.2?

GPT-5.2 has the edge for math in this comparison, averaging 35.2 versus 11.2. Inside this category, FrontierMath v2 (Tiers 1-3) is the benchmark that creates the most daylight between them.

Which is better for reasoning, Claude Sonnet 4.5 or GPT-5.2?

GPT-5.2 has the edge for reasoning in this comparison, averaging 52.9 versus 13.6. Inside this category, ARC-AGI-2 is the benchmark that creates the most daylight between them.

Which is better for agentic tasks, Claude Sonnet 4.5 or GPT-5.2?

GPT-5.2 has the edge for agentic tasks in this comparison, averaging 55.7 versus 55.4. Inside this category, OSWorld-Verified is the benchmark that creates the most daylight between them.

Related Comparisons

Explore More

Anthropic Compare Pricing Methodology Find Your Best LLM Overall Rankings
Last updated: July 20, 2026

Choose a model with this week’s evidence

Join 2,000+ readers for ranking moves, pricing changes, and the claims that still need proof.

One email each week. Unsubscribe anytime.