GPT-5.4 vs GPT-5.3-Codex vs Claude Opus 4.6 for Coding

Quick verdict: pick by workload

The benchmark trap: these numbers are not apples-to-apples

SWE-Bench Verified and SWE-Bench Pro Public are different signals

Claude Opus 4.6’s strongest case comes from SWE-Bench Verified. The cited reports put it at 79.2%, 79.4%, or 80.8% on that benchmark variant ^[3][5][7][9].

GPT-5.3-Codex is harder to summarize because the provided reports use different SWE-Bench lines. One GPT-5.4 analysis lists GPT-5.3-Codex at 56.8% on SWE-Bench Pro, while two Opus-vs-Codex comparisons list GPT-5.3-Codex at 78.2% on SWE-Bench Pro Public ^[3][6][7]. That is a warning against casual ranking, not a reason to average the scores. Multiple sources explicitly caution that SWE-Bench Verified and SWE-Bench Pro Public are not directly comparable ^[6][7][10].

GPT-5.4’s cleanest OpenAI-on-OpenAI coding edge in these sources is small: 57.7% on SWE-Bench Pro versus 56.8% for GPT-5.3-Codex in the same GPT-5.4-focused analysis ^[3]. Another summary also flags the 57.7% GPT-5.4 SWE-Bench Pro Public figure while warning that the broader Claude-vs-GPT comparison is not apples-to-apples ^[10].

Terminal-Bench results include the agent harness

Terminal-Bench 2.0 is especially easy to misread because the public leaderboard lists agent/model pairs, not isolated base-model scores ^[1]. In that leaderboard, GPT-5.3-Codex appears at 78.4% with SageAgent, 77.3% with Droid, and 75.1% with Simple Codex ^[1]. Claude Opus 4.6 appears at 79.8% with ForgeCode, 75.3% with Capy, and 62.9% with Terminus 2 ^[1].

That spread is large enough to change the apparent winner. The GPT-5.4-focused comparison reports GPT-5.3-Codex ahead of Claude Opus 4.6 on Terminal-Bench 2.0, 77.3% versus 65.4% ^[3]. But the public leaderboard has a ForgeCode/Claude Opus 4.6 entry at 79.8%, above the SageAgent/GPT-5.3-Codex entry at 78.4% ^[1]. The practical conclusion is that terminal-agent evaluations must hold the harness constant before making a model claim.

Model-by-model read

Claude Opus 4.6: strongest Verified-style bug-fixing signal

If your proxy for coding quality is SWE-Bench Verified, Claude Opus 4.6 is the best-supported starting point in these sources. Its reported Verified scores cluster around 79% to 81%: 79.2% in the GPT-5.4 analysis, 79.4% in Opus-vs-Codex comparisons, and 80.8% in other benchmark roundups ^{[3][5][6][7][9]}.

That does not prove Opus 4.6 wins every coding workload. Its Terminal-Bench story is mixed: comparison reports cite 65.4%, while the public leaderboard shows 79.8% when Opus 4.6 is paired with ForgeCode and 62.9% with Terminus 2 ^[1][3][7][9]. Opus 4.6 is the safest first test for Verified-style repository repair, but not a universal coding champion.

GPT-5.3-Codex: the OpenAI terminal-agent standout

GPT-5.3-Codex has the strongest OpenAI case when the workload resembles Terminal-Bench-style agentic shell work. It is reported at 77.3% on Terminal-Bench 2.0 in comparison reports, and the public leaderboard lists GPT-5.3-Codex at 78.4% with SageAgent, 77.3% with Droid, and 75.1% with Simple Codex ^[1][3][7][9].

Its SWE-Bench interpretation needs more care. Some reports list GPT-5.3-Codex at 78.2% on SWE-Bench Pro Public, while others list 56.8% on SWE-Bench Pro ^[3][6][7][9]. Because the cited sources warn that these variants are not directly interchangeable, GPT-5.3-Codex should be judged in the same SWE-Bench variant and evaluation setup you plan to use ^[6][7][10].

GPT-5.4: a modest coding bump with a tool-use angle

GPT-5.4 does not look like a coding blowout in the provided benchmark set. The main same-source comparison gives it a narrow SWE-Bench Pro lead over GPT-5.3-Codex, 57.7% versus 56.8%, while also showing a lower Terminal-Bench 2.0 result, 75.1% versus 77.3% ^[3].

The more distinctive GPT-5.4 datapoint is tool use. The GPT-5.4 analysis says tool search reduces MCP token usage by 47% by loading tool definitions on demand instead of putting all definitions into context ^[3]. For tool-heavy coding agents, that may be a real systems advantage, but it should be measured separately from benchmark accuracy.

How to compare them without fooling yourself

1. Pick the benchmark variant before picking the winner. SWE-Bench Verified, SWE-Bench Pro, and SWE-Bench Pro Public should not be collapsed into one score table ^[6][7][10].
2. Keep the agent harness constant for terminal tasks. The public Terminal-Bench 2.0 leaderboard shows that the same model can land at meaningfully different accuracies depending on the agent pairing ^[1].
3. Separate coding accuracy from tool efficiency. GPT-5.4’s reported 47% MCP token reduction is useful evidence for tool-heavy systems, but it is not the same claim as a SWE-Bench or Terminal-Bench win ^[3].
4. Treat mixed-source rankings as directional. The provided sources support different winners under different benchmarks, which is exactly why a single universal ranking would overstate the evidence ^{[1][3][6][7][10]}.

Bottom line

Start with Claude Opus 4.6 for SWE-Bench Verified-style bug fixing, keep GPT-5.3-Codex in any terminal-agent bakeoff, and test GPT-5.4 when you need the latest OpenAI model or want to evaluate its tool-search efficiency ^{[1][3][5][7][9]}. The safest overall verdict is not that one model dominates coding. It is that the winner changes with the benchmark variant, the agent harness, and the workload you actually plan to run ^{[1][6][7][10]}.