GPT-5.2 vs Gemini 3 Pro vs Opus 4.5: Benchmark Comparisons
OpenAI released GPT-5.2 in three tiers: Classic, Thinking, and Pro. Pro shipped at the same time as the others for the first time, and it is available in ChatGPT and via OpenRouter. Reports say the release accelerated after Google launched Gemini 3. This post compares the benchmark figures OpenAI published against Gemini 3 Pro and Claude Opus 4.5, and it notes where the numbers stop being comparable.
- GPT-5.2 Classic: the default mode for ChatGPT.
- GPT-5.2 Thinking: extended reasoning with
light, standard, extended, and heavy settings. - GPT-5.2 Pro: reasoning compute, called the juice level, up to 768, above the 128-256 range of earlier models. This tier sits behind the $200 ChatGPT plan.
Pietro, who tested the model, called it a serious leap forward in complex reasoning, math, coding, and simulations, and highlighted a one-shot build of a 3D graphics engine.
Context, Vision, and Reliability
Section titled “Context, Vision, and Reliability”GPT-5.2 scores near-perfect on OpenAI’s MRCv2 retrieval tests up to 256k tokens. Longer context means fewer chat resets on long tasks. On screenshot analysis, it identifies VGA, HDMI, and USB-C ports on a motherboard, which GPT-5.1 missed. OpenAI reports a hallucination rate of 0.8%, down 30-40% from earlier models.
The Reported Numbers
Section titled “The Reported Numbers”| Benchmark | GPT-5.2 (reported) | Gemini 3 Pro | Opus 4.5 |
|---|---|---|---|
| SWE-bench Pro | 55.6% | 43.3% | 52% |
| ARC-AGI v1 | +20% vs Gemini | baseline | +15% vs Opus |
| GPQA Diamond | top score claimed | not disclosed | not disclosed |
| CTF (security) | leading result among tested | not disclosed | not disclosed |
| GDP-Val | 71% win rate vs experts | not run | not run |
Three limits apply. First, nearly all figures are vendor-reported on vendor-chosen benchmarks. Second, some cells are qualitative because OpenAI did not publish competitor numbers. Third, a benchmark delta is not the same as a head-to-head result in your own workload. The table is a summary of what was published, not a verdict.
On the CTF benchmark, which uses realistic hacking scenarios at 12-shot pass@12, OpenAI reports GPT-5.2 leading the tested models. Internally, the company says the model replicates 55% of research engineers’ pull requests.
Cost and Business Tasks
Section titled “Cost and Business Tasks”On ARC-AGI, OpenAI reports efficiency rising from o1’s 88% at $4,500 per task to a higher score at $11 for GPT-5.2 Pro. That is a 390x cost drop in one year, as reported.
On business tasks, OpenAI reports GPT-5.2 beating experts 70.9% of the time, at under 1% of the cost and 11x the speed. Ethan Mollick of Wharton reviewed the GDP-Val results and noted that the model wins head-to-head on 4-8 hour expert tasks 71% of the time by judge ratings. In a hands-on test, GPT-5.2 Thinking produced a formatted PowerPoint in 19 minutes from a screenshot of notes. Excel and Sheets output reached polished financial models with professional formatting.
A Coding Demo
Section titled “A Coding Demo”In Cursor with the Codex extension, GPT-5.2 Pro built a terminal CLI agent from scratch. The agent uses pipx and scans network interfaces, routes, and Wi-Fi details. It asks the user for location and purpose, sends the data to GPT-5.2 via OpenRouter, and returns a risk rating. On a home setup it returned a risk rating of 3 out of 10, with HTTPS advice.
The Road Ahead
Section titled “The Road Ahead”Sam Altman said more ChatGPT updates are coming. Independent labs will need to run the same tests before the picture firms up. Until then, GPT-5.2’s published numbers are the most complete set OpenAI has released on project-level work, with the caveats above.
The Cost of the Race
Section titled “The Cost of the Race”The speed of the release cycle has a cost side. Pro mode runs hours of inference at up to 768 reasoning units, which draws significant compute. Inference at this scale has a measurable water and energy footprint, a point OpenAI has discussed in its own infrastructure reporting. Racing a competitor can also compress the safety work between releases. None of this changes the benchmark numbers. It is a separate question worth asking alongside them.