Claude vs GPT 2026: Opus 4.6 vs GPT-5.4 Benchmarks

Published: March 26, 2026 | Data updated: 2026-03-26 | ai-models

gpt-5-4claude-opus-4-6llm-benchmarkcoding-ai2026

Claude Opus 4.6 leads GPT-5.4 in coding (HumanEval: 97.8% vs 93.1%, SWE-bench Verified: 80.8% vs ~80%) and long-context recall (MRCR v2 at 1M tokens: 76% vs 18.5%), while GPT-5.4 wins on reasoning (ARC-AGI-2: 73.3% vs 68.8%), graduate-level science (GPQA Diamond: 92.8% vs 88.5–91%), and agentic OS tasks (OSWorld: 75.0% vs 72.7%). Neither model is a universal winner — the right choice depends on your workload.

TL;DR

Model Overview

Property GPT-5.4 Claude Opus 4.6
Developer OpenAI Anthropic
Release date March 5, 2026 February 5, 2026
Context window 1,050,000 tokens 1,000,000 tokens
Input price (per 1M tokens) $2.50 $5.00
Output price (per 1M tokens) $15.00 $25.00

GPT-5.4 launched one month after Claude Opus 4.6. Both offer approximately 1M token context windows, making long-document processing viable for both — though their recall accuracy at that limit differs significantly (see MRCR v2 below).

Full Benchmark Comparison: claude vs gpt 2026

{
  "benchmark_report": {
    "generated": "2026-03-26",
    "methodology": "Published vendor benchmarks cross-referenced against independent evaluations. All scores use pass@1 unless otherwise noted.",
    "models": {
      "gpt_5_4": {
        "name": "GPT-5.4",
        "developer": "OpenAI",
        "release_date": "2026-03-05",
        "context_window": 1050000,
        "pricing": {
          "input_per_1m": 2.50,
          "output_per_1m": 15.00
        }
      },
      "claude_opus_4_6": {
        "name": "Claude Opus 4.6",
        "developer": "Anthropic",
        "release_date": "2026-02-05",
        "context_window": 1000000,
        "pricing": {
          "input_per_1m": 5.00,
          "output_per_1m": 25.00
        }
      }
    },
    "results": [
      {
        "benchmark": "SWE-bench Verified",
        "category": "coding",
        "gpt_5_4": "~80%",
        "claude_opus_4_6": "80.8%",
        "winner": "Claude Opus 4.6"
      },
      {
        "benchmark": "SWE-bench Pro",
        "category": "coding",
        "gpt_5_4": "57.7%",
        "claude_opus_4_6": "~45%",
        "winner": "GPT-5.4"
      },
      {
        "benchmark": "HumanEval",
        "category": "coding",
        "gpt_5_4": "93.1%",
        "claude_opus_4_6": "97.8%",
        "winner": "Claude Opus 4.6"
      },
      {
        "benchmark": "ARC-AGI-2",
        "category": "reasoning",
        "gpt_5_4": "73.3%",
        "claude_opus_4_6": "68.8%",
        "winner": "GPT-5.4"
      },
      {
        "benchmark": "GPQA Diamond",
        "category": "science_qa",
        "gpt_5_4": "92.8%",
        "claude_opus_4_6": "88.5–91%",
        "winner": "GPT-5.4"
      },
      {
        "benchmark": "AIME 2025",
        "category": "math",
        "gpt_5_4": "100%",
        "claude_opus_4_6": "100%",
        "winner": "Tie"
      },
      {
        "benchmark": "OSWorld",
        "category": "agentic_os",
        "gpt_5_4": "75.0%",
        "claude_opus_4_6": "72.7%",
        "winner": "GPT-5.4"
      },
      {
        "benchmark": "Terminal-Bench 2.0",
        "category": "agentic_terminal",
        "gpt_5_4": "75.1%",
        "claude_opus_4_6": "65.4%",
        "winner": "GPT-5.4"
      },
      {
        "benchmark": "MRCR v2 at 1M tokens",
        "category": "long_context_recall",
        "gpt_5_4": "18.5%",
        "claude_opus_4_6": "76%",
        "winner": "Claude Opus 4.6"
      }
    ]
  }
}

Benchmark Table

Metric GPT-5.4 Claude Opus 4.6 Winner
SWE-bench Verified ~80% 80.8% Claude
SWE-bench Pro 57.7% ~45% GPT-5.4
HumanEval 93.1% 97.8% Claude
ARC-AGI-2 73.3% 68.8% GPT-5.4
GPQA Diamond 92.8% 88.5–91% GPT-5.4
AIME 2025 100% 100% Tie
OSWorld 75.0% 72.7% GPT-5.4
Terminal-Bench 2.0 75.1% 65.4% GPT-5.4
MRCR v2 at 1M tokens 18.5% 76% Claude

Tally: GPT-5.4 wins 5, Claude wins 3, 1 tie.

Coding Performance: HumanEval and SWE-bench

Claude Opus 4.6 holds a 4.7-point lead on HumanEval (97.8% vs 93.1%). On SWE-bench Verified — a benchmark of real GitHub issues — Claude edges GPT-5.4 by under 1 percentage point (80.8% vs ~80%). Both results indicate Claude has a marginal but consistent edge on standard and verified coding tasks.

However, SWE-bench Pro tells a different story. This harder variant, which includes more ambiguous or multi-step issues, gives GPT-5.4 a substantial lead: 57.7% vs ~45%. For production-level autonomous coding agents that must handle underspecified or complex tickets, GPT-5.4 demonstrates stronger generalization.

Practical implication: For code generation, completions, and well-scoped automated fixes, Claude Opus 4.6 is the better choice. For agentic coding pipelines where tasks are open-ended or require deeper planning, GPT-5.4 has a meaningful advantage.

Reasoning and Science Benchmarks

On ARC-AGI-2, GPT-5.4 scores 73.3% vs Claude’s 68.8% — a 4.5-point lead that reflects stronger abstract reasoning on novel problem structures. ARC-AGI-2 is designed to be unsolvable by memorization, making this a meaningful signal.

GPQA Diamond (graduate-level science questions) shows GPT-5.4 at 92.8% vs Claude at 88.5–91%. The range on Claude’s score reflects variation across evaluation runs reported by Anthropic; GPT-5.4’s single-point score reflects OpenAI’s published result. Either way, GPT-5.4 leads.

AIME 2025 is a perfect-score tie. Both models solve 100% of the 2025 AMC/AIME math competition problems, indicating equivalent ceiling performance on competition mathematics.

Long-Context Recall: The Clearest Performance Gap

The MRCR v2 benchmark at 1M tokens is where the performance gap is most dramatic. Claude Opus 4.6 scores 76%; GPT-5.4 scores 18.5%. That is a 57.5-point gap — the largest single-metric difference between these models across all evaluated benchmarks.

MRCR v2 measures multi-document retrieval and cross-reference recall at the full context limit. For applications that need to process entire codebases, legal corpora, or large research datasets in a single context window, Claude Opus 4.6 is not just better — it is categorically more reliable.

GPT-5.4’s 1.05M context window is slightly larger in raw token count, but if the model cannot reliably retrieve information from that window, the additional 50K tokens provide little practical benefit.

Agentic Task Performance: OSWorld and Terminal-Bench 2.0

For autonomous computer and terminal use, GPT-5.4 holds consistent leads:

The Terminal-Bench gap is the more significant finding. For AI agents that operate in CLI environments — DevOps automation, infrastructure scripting, or autonomous server management — GPT-5.4 shows meaningfully stronger execution reliability.

Pricing and Cost Analysis

Scenario GPT-5.4 cost Claude Opus 4.6 cost Difference
1M input + 100K output $2.50 + $1.50 = $4.00 $5.00 + $2.50 = $7.50 Claude costs ~88% more
10M input + 1M output $25.00 + $15.00 = $40.00 $50.00 + $25.00 = $75.00 Claude costs ~88% more
Long-context at scale Similar gap holds Similar gap holds GPT-5.4 wins on cost

At current list prices, Claude Opus 4.6 is roughly 1.88x more expensive per equivalent token volume. For high-throughput API use cases, this is a significant operational cost difference. Teams should factor in whether Claude’s specific advantages (HumanEval, SWE-bench Verified, long-context recall) justify the premium for their workload.

Use Case Recommendations

Choose Claude Opus 4.6 if:

Choose GPT-5.4 if:

When the choice is neutral:

FAQ

###Which model is better for coding in 2026, GPT-5.4 or Claude Opus 4.6?

For standard code generation and automated fixes on well-defined tasks, Claude Opus 4.6 leads (HumanEval: 97.8%, SWE-bench Verified: 80.8%). For complex, agentic coding agents handling underspecified issues, GPT-5.4 leads (SWE-bench Pro: 57.7% vs ~45%). The best choice depends on your task type.

###Is Claude Opus 4.6 worth the higher price compared to GPT-5.4?

At approximately 88% higher cost per token, Claude Opus 4.6 is worth the premium only in specific scenarios: long-context retrieval pipelines (MRCR v2: 76% vs 18.5%), standard coding tasks, and HumanEval-class code generation. For reasoning, agentic tasks, and science QA, GPT-5.4 offers better performance at lower cost.

###How do GPT-5.4 and Claude Opus 4.6 compare on long-context tasks?

This is the sharpest performance difference between the two models. Claude Opus 4.6 scores 76% on MRCR v2 at 1M tokens; GPT-5.4 scores 18.5%. While GPT-5.4 has a marginally larger context window (1.05M vs 1M tokens), Claude’s ability to accurately retrieve and cross-reference information at that scale is far superior.

###Which model wins on math benchmarks?

Both models score 100% on AIME 2025, making math performance a tie at the competition level. GPT-5.4 leads on GPQA Diamond (graduate science that includes physics and chemistry alongside math), suggesting stronger performance on multi-domain reasoning, not just pure mathematics.

###When did GPT-5.4 and Claude Opus 4.6 release?

Claude Opus 4.6 was released on February 5, 2026. GPT-5.4 was released on March 5, 2026 — one month later.

Last updated: 2026-03-26