Compare / head-to-head
GPT-5.2-CodexvsGPT-5.4
GPT-5.4 leads 10 of 11 shared benchmarks. GPT-5.2-Codex is 1.1x cheaper per token. GPT-5.4 has the larger context window (1.1M tokens).
Benchmarks from cited public sources; pricing from official pages; status from official provider feeds.
Shared benchmarks
1 – 10
GPT-5.4 leads
Cheaper per token
GPT-5.2-Codex
1.1x cheaper, input + output
Larger context
GPT-5.4
1.1M tokens
Providers
OpenAI
same provider
GPT-5.2-Codex
OpenAI · released 2025-12-18
textvision
Quality
Benchmark matrix
11 shared · 2 only GPT-5.2-Codex · 31 only GPT-5.4| Benchmark | GPT-5.2-Codex | GPT-5.4 | Δ | Edge |
|---|---|---|---|---|
| Reported by both · 11 | ||||
| LiveBench Agentic Coding (LiveBench) | 49.4% ↗ | 53.8% ↗ | -4.4 pt | GPT-5.4 |
| LiveBench Coding (LiveBench) | 83.6% ↗ | 77.5% ↗ | +6.1 pt | GPT-5.2-Codex |
| LiveBench Data Analysis (LiveBench) | 78.2% ↗ | 79.3% ↗ | -1.1 pt | GPT-5.4 |
| LiveBench Instruction Following (LiveBench) | 66.4% ↗ | 70.2% ↗ | -3.8 pt | GPT-5.4 |
| LiveBench Language (LiveBench) | 73.7% ↗ | 82.6% ↗ | -8.9 pt | GPT-5.4 |
| LiveBench Mathematics (LiveBench) | 88.8% ↗ | 94.1% ↗ | -5.3 pt | GPT-5.4 |
| LiveBench Reasoning (LiveBench) | 77.7% ↗ | 88.1% ↗ | -10.4 pt | GPT-5.4 |
| LMArena WebDev (LMArena) | 1338.9 ↗ | 1395.1 ↗ | -56.2 | GPT-5.4 |
| OSWorld-Verified | 38.2% ↗ | 75% ↗ | -36.8 pt | GPT-5.4 |
| SWE-Bench Pro | 56.4% ↗ | 57.7% ↗ | -1.3 pt | GPT-5.4 |
| Terminal-Bench 2.0 | 64% ↗ | 75.1% ↗ | -11.1 pt | GPT-5.4 |
| Only GPT-5.2-Codex reports · 2 | ||||
| Cybersecurity Capture The Flag Challenges | 67.4% ↗ | not reported | — | — |
| SWE-Lancer IC Diamond | 76% ↗ | not reported | — | — |
| Only GPT-5.4 reports · 31 | ||||
| GPQA Diamond | not reported | 92.8% ↗ | — | — |
| AIME 2026 | not reported | 99.2% ↗matharena ⚠ | — | — |
| Humanity's Last Exam (no tools) | not reported | 39.8% ↗ | — | — |
| LMArena Elo | not reported | 1452.6 ↗ | — | — |
| APEX-Agents (Mercor) | not reported | 52.4% ↗ | — | — |
| BrowseComp | not reported | 82.7% ↗ | — | — |
| Chess Puzzles (Epoch AI run) | not reported | 44% ↗epoch run | — | — |
| DeepSWE v1.1 (Datacurve) | not reported | 51.8% ↗ | — | — |
| EBR-bench (Epoch AI run) | not reported | 25.4% ↗epoch run | — | — |
| FrontierMath Tier 4 v2 (Epoch AI run) | not reported | 49% ↗epoch run | — | — |
| FrontierMath Tiers 1-3 v2 (Epoch AI run) | not reported | 78.6% ↗epoch run | — | — |
| Furniture Assembly (Epoch AI run) | not reported | 37.5% ↗epoch run | — | — |
| GDPval (wins or ties) | not reported | 83% ↗ | — | — |
| GSO Opt@1 (GSO) | not reported | 30.4% ↗ | — | — |
| Humanity's Last Exam (with tools) | not reported | 52.1% ↗ | — | — |
| LMArena Agent (LMArena) | not reported | 0.0105 ↗ | — | — |
| LMArena Vision (LMArena) | not reported | 1283.8 ↗ | — | — |
| MCP Atlas | not reported | 67.2% ↗ | — | — |
| MirrorCode (Epoch AI run) | not reported | 15.6% ↗epoch run | — | — |
| MMMU-Pro (no tools) | not reported | 81.2% ↗ | — | — |
| MMMU-Pro (with tools) | not reported | 82.1% ↗ | — | — |
| Mystery Game Puzzles (Epoch AI run) | not reported | 37% ↗epoch run | — | — |
| OTIS Mock AIME 2024-2025 (Epoch AI run) | not reported | 97.8% ↗epoch run | — | — |
| SAGE (Vals AI) | not reported | 43.3% ↗ | — | — |
| SimpleQA Verified | not reported | 45.1% ↗epoch run | — | — |
| SWE-bench Verified (Epoch AI run) | not reported | 76.9% ↗epoch run | — | — |
| tau2-bench Banking Knowledge (Sierra) | not reported | 39.4% ↗ | — | — |
| Tau2-bench Telecom | not reported | 98.9% ↗ | — | — |
| Toolathlon | not reported | 54.6% ↗ | — | — |
| Vending-Bench 2 (Andon Labs) | not reported | 6144.18 ↗ | — | — |
| WeirdML (Håvard Tveit Ihle) | not reported | 77.7% ↗ | — | — |
Scores tagged "epoch" or "matharena" are independent runs, used only where the lab has not published its own; ⚠ marks rows MathArena flags as released after the competition. Higher is better on every row. Δ is GPT-5.2-Codex minus GPT-5.4 in the benchmark's own unit. "Not reported" means the lab has not published that figure; it is not a zero. ↗ opens the source.
Specs & pricing
Side by side
Official pricing pages and model cards| Spec | GPT-5.2-Codex | GPT-5.4 | Edge |
|---|---|---|---|
| Context window | 400K tokens | 1.1M tokens | GPT-5.4 |
| Max output | 128K tokens | 128K tokens | Tie |
| Input price / 1M | $1.75 ↗ | $2.5 ↗ | GPT-5.2-Codex |
| Output price / 1M | $14 ↗ | $15 ↗ | GPT-5.2-Codex |
| Cached input / 1M | $0.175 ↗ | $0.25 ↗ | GPT-5.2-Codex |
| Input + output / 1M Lower is cheaper. List prices; batch, tool and regional fees excluded. | $15.75 | $17.5 | GPT-5.2-Codex |
| Modalities | text · vision | text · vision | Tie |
| Released | 2025-12-18 | 2026-03-05 | — |
| Cited benchmark scores | 13 | 50 | — |
Reliability
OpenAI status
Live from /statusMore matchups
GPT-5.2-Codex vs …
Models sharing the most benchmarksMore matchups
GPT-5.4 vs …
Models sharing the most benchmarksBuilt by Respan
Which one wins on your data?
Public benchmarks are a starting point. Run GPT-5.2-Codex and GPT-5.4 on your own prompts with Respan evals, or route to either through one gateway key with automatic failover.