GPT-5.2
GPT-5.2 was the model that made ChatGPT good at office work, but it is now superseded. OpenAI released it on 11 December 2025 in three versions: Instant, Thinking and Pro. OpenAI's own tests showed big jumps over GPT-5.1, especially on professional tasks (GDPval 70.9% vs 38.8%) and abstract puzzles (ARC-AGI-2 52.9% vs 17.6%).
All GPT-5.2 models left ChatGPT in June 2026, and old chats now continue on GPT-5.5. The API version still runs at $1.75 per million input tokens and $14 per million output tokens, but OpenAI's model page now recommends GPT-6 Astra instead.
GPT-5.2 was a major upgrade in its day, and it still holds up for many API jobs, but there is little reason to choose it now. It roughly doubled GPT-5.1's GDPval score, a test of real job tasks such as financial models and presentations. It also scored 80.0% on SWE-bench Verified and 52.9% on ARC-AGI-2, all according to OpenAI. Box's CEO said it scored 7 points higher than GPT-5.1 on Box's own knowledge-work tests.
The launch was rushed. Reports described an internal "code red" at OpenAI after Google's Gemini 3 Pro took the lead. The Guardian also reported that GPT-5.2 was citing Grokipedia, an encyclopedia criticised for weak sourcing.
Who should use it: teams with a stable GPT-5.2 integration who need a few more weeks before migrating.
Who should not: anyone starting a new project. GPT-5.4 beat it on every benchmark in OpenAI's GPT-5.4 comparison table (GDPval 83.0%, OSWorld-Verified 75.0% vs 47.3%). GPT-6 Sol now costs less per output token ($10 vs $14).
Score breakdown
Best for
- Existing GPT-5.2 integrations during migration
- Baseline comparisons when testing newer models
- Spreadsheet and document tasks where GPT-5.2 prompts are already tuned
What we like
- Big jump in professional-task quality over GPT-5.1 (GDPval 70.9%, OpenAI figure)
- Strong maths and science scores: 100% on AIME 2025 and 92.4% on GPQA Diamond (OpenAI)
- 400K-token context and newer August 2025 knowledge cutoff
- Still callable in the API with a fixed snapshot
Watch out for
- Retired from ChatGPT in June 2026
- Costs more per output token ($14) than the newer GPT-6 Sol ($10)
- Pro version is very expensive at $168 per million output tokens
- Weak at computer use (47.3% on OSWorld-Verified) compared with GPT-5.4
Specs at a glance
| Developer | OpenAI |
|---|---|
| Release date | 11 December 2025 (ChatGPT paid plans and API) |
| Variants | GPT-5.2 Instant, GPT-5.2 Thinking, GPT-5.2 Pro, GPT-5.2-Codex |
| API snapshot | gpt-5.2-2025-12-11 |
| Context window | 400,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | 31 August 2025 |
| Input / output | Text and image in; text out |
| API pricing | $1.75 input, $0.175 cached input, $14 output per 1M tokens |
| GPT-5.2 Pro API pricing | $21 input, $168 output per 1M tokens |
| Predecessor / successor | GPT-5.1 / GPT-5.3-Codex and GPT-5.4 |
| ChatGPT retirement | Instant 3 June 2026, Thinking 5 June 2026; all gone by 12 June 2026 |
Benchmarks
Benchmarks are standard tests. Vendor-run results are marked as such; independent results are preferred where they exist.
| Benchmark | Score | Source | Note |
|---|---|---|---|
| GDPval (wins or ties vs experts) | 70.9% | OpenAI | GPT-5.2 Thinking; GPT-5.1 Thinking scored 38.8% |
| SWE-bench Verified | 80.0% | OpenAI | GPT-5.2 Thinking |
| SWE-bench Pro | 55.6% | OpenAI | |
| GPQA Diamond | 92.4% | OpenAI | |
| AIME 2025 | 100% | OpenAI | No tools |
| FrontierMath Tier 1–3 | 40.3% | OpenAI | |
| ARC-AGI-2 | 52.9% | OpenAI | GPT-5.1 Thinking scored 17.6% |
| OSWorld-Verified | 47.3% | OpenAI (GPT-5.4 launch post) | |
| BrowseComp | 65.8% | OpenAI (GPT-5.4 launch post) |
Pricing
| Plan / tier | Price | Notes |
|---|---|---|
| API input | $1.75 per 1M tokens | |
| API cached input | $0.175 per 1M tokens | 90% discount |
| API output | $14 per 1M tokens | |
| GPT-5.2 Pro (API) | $21 in / $168 out per 1M tokens | |
| ChatGPT | Not available | Retired in June 2026; old chats continue on GPT-5.5. |
GPT-5.2 vs GPT-5.1: benchmark gains
| Benchmark | GPT-5.1 Thinking | GPT-5.2 Thinking | Change |
|---|---|---|---|
| GDPval | 38.8% | 70.9% | +32.1 pts |
| SWE-bench Verified | 76.3% | 80.0% | +3.7 pts |
| SWE-bench Pro | 50.8% | 55.6% | +4.8 pts |
| GPQA Diamond | 88.1% | 92.4% | +4.3 pts |
| AIME 2025 | 94.0% | 100% | +6.0 pts |
| FrontierMath T1–3 | 31.0% | 40.3% | +9.3 pts |
| ARC-AGI-2 | 17.6% | 52.9% | +35.3 pts |
All figures are OpenAI's own, from the GPT-5.2 launch post. Independent labs may measure differently.
The three GPT-5.2 versions
- GPT-5.2 Instant: the fast everyday model for chat, learning and simple writing. Replaced as ChatGPT's default by GPT-5.3 Instant on 3 March 2026.
- GPT-5.2 Thinking: the reasoning model for complex work such as spreadsheets, financial models and multi-step projects. Replaced by GPT-5.4 Thinking on 5 March 2026.
- GPT-5.2 Pro: uses more computing power per answer for the hardest questions. Available on higher-tier paid ChatGPT plans and in the API at $21 / $168 per million tokens.
OpenAI also shipped GPT-5.2-Codex for its Codex coding agent. GPT-5.3-Codex replaced it in February 2026.
Retirement from ChatGPT
OpenAI kept GPT-5.2 in ChatGPT's Legacy menu for about three months after each replacement. GPT-5.2 Instant was set to retire on 3 June 2026 and GPT-5.2 Thinking on 5 June 2026. By 12 June 2026 all GPT-5.2 models, including Pro, were gone from ChatGPT, and existing chats moved to the matching GPT-5.5 model. The API model remains available, but its docs page now calls it a "previous flagship".
- Pin
gpt-5.2-2025-12-11in production until your migration tests pass, so nothing shifts under you. - Avoid GPT-5.2 Pro for routine work. At $168 per million output tokens, one long answer can cost more than a day of normal GPT-5.2 use.
- When you move to GPT-5.4 or later, re-check spreadsheet and slide outputs first. That is where GPT-5.2 prompts were most tuned.
- Use cached input for long, repeated instructions. It cuts input cost from $1.75 to $0.175 per million tokens.
Jargon explained
- ARC-AGI-2
- A set of visual logic puzzles that are easy for people but hard for AI. It tests whether a model can work out new rules instead of recalling facts.
- SWE-bench Verified
- A coding test where the model must fix real bugs in open-source Python projects. The score is the share it fixes correctly.
- GDPval
- OpenAI's test of real job tasks, such as reports and spreadsheets, judged against work by human experts.
- Snapshot
- A fixed, dated version of a model in the API. It does not change, unlike a moving alias such as 'latest'.
Alternatives to consider
Frequently asked questions
When was GPT-5.2 released?
11 December 2025, starting with paid ChatGPT plans and the API on the same day.
Is GPT-5.2 still in ChatGPT?
No. All GPT-5.2 models were removed from ChatGPT by 12 June 2026. Existing GPT-5.2 chats continue on GPT-5.5. The model is still in the API.
How much does the GPT-5.2 API cost?
$1.75 per million input tokens, $0.175 per million cached input tokens and $14 per million output tokens. GPT-5.2 Pro costs $21 input and $168 output per million tokens.
Is GPT-5.2 better than GPT-5.1?
Yes, clearly, on OpenAI's figures. The biggest gains were GDPval (70.9% vs 38.8%) and ARC-AGI-2 (52.9% vs 17.6%).
What should I use instead of GPT-5.2?
For most API work, GPT-6 Sol. For the hardest tasks, GPT-6 Astra. If you want the smallest change, GPT-5.4 was the direct successor.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- Introducing GPT-5.2 (OpenAI)
- GPT-5.2 model page (OpenAI API docs)
- GPT-5.2 (Wikipedia)
- Introducing GPT-5.4 (OpenAI)
- GPT-5.3 Instant (OpenAI)
- OpenAI upgrades GPT-5.5, as it plans to retire legacy ChatGPT models (BleepingComputer)
- ChatGPT release notes (OpenAI Help Center)