GPT-4.5 Preview
API ID gpt-4.5-preview
gpt-4.5-preview was a preview OpenAI model. It had a 128,000-token context window and up to 16,384 output tokens. It took image input and supported function calling and structured outputs.
- Status
- Retired
- Released
- —
- Shutdown
- Jul 14, 2025
- Replacement
- GPT-4.1gpt-4.1
gpt-4.5-preview stopped serving on July 14, 2025. Recommended replacement: gpt-4.1.
Benchmarks
Scores as each source published themComparing with GPT-4.1, the previous model in this line. A score is compared only with the same benchmark from the same source, at the same setting where both have it. Full comparison: prices, limits, features
Arena
| Arena | GPT-4.5 Preview | GPT-4.1 | Change | Rank | Date | Source |
|---|---|---|---|---|---|---|
TextBy category
| 1445 ±6 | 1415 ±4 | +30 | 88 of 415 14,547 votes | Oct 13, 2026 | source |
| Text: Coding | 1458 ±13 | 1456 ±7 | +2 within range | 143 of 410 1,939 votes | Oct 13, 2026 | source |
| Vision | 1225 ±12 | 1214 ±7 | +11 within range | 77 of 165 2,925 votes | Oct 9, 2026 | source |
Rated as gpt-4.5-preview-2025-02-27; GPT-4.1 as gpt-4.1-2025-04-14. Ratings come from people voting between two anonymous answers; Text and Vision are style-controlled, as on arena.ai. Data: Arena (arena.ai), CC BY 4.0.
Epoch AI
| Benchmark | GPT-4.5 Preview | GPT-4.1 | Change | Setting | Run | Source |
|---|---|---|---|---|---|---|
| GPQA Diamond | 68.7% | 66.9% | +1.8 pts | Feb 28, 2025 | source | |
| MATH Level 5 | 78.6% | 83.0% | −4.4 pts | Feb 28, 2025 | source | |
| OTIS Mock AIME 2024-2025 | 37.8% | 38.3% | −0.5 pts | Feb 28, 2025 | source |
Run independently by Epoch AI. Data: Epoch AI, Capabilities & benchmarking (epoch.ai/benchmarks), CC BY.
Alias
All aliasesCurrently points to gpt-4.5-preview-2025-02-27 (in place when tracking began (October 7, 2026)).
Features
Compare models| Feature | Value | Since | Source |
|---|---|---|---|
| Structured outputsThe model can be held to a JSON schema you supply, so every reply parses without retries or cleanup. | Supported | in place when tracking began (October 6, 2026) | source |
| Function callingAlso called tool use. You describe functions; the model replies with a function name and arguments for your code to run, then uses the result. | Supported | in place when tracking began (October 6, 2026) | source |
| Prompt cachingA prompt prefix you send again and again (system prompt, documents, tools) is cached, so later requests cost less and start faster. | Supported | in place when tracking began (October 6, 2026) | source |
| Batch APISend many requests as one job that runs in the background, usually finished within 24 hours, at a lower price than live requests. | Supported | in place when tracking began (October 6, 2026) | source |
| Image inputImages such as screenshots, photos and charts can go in the prompt for the model to read. | Supported | in place when tracking began (October 6, 2026) | source |
| Fine-tuningYou can train your own version of the model on your examples through the company's API, then call it like any other model. | Not supported | in place when tracking began (October 6, 2026) | source |
| Context windowThe most the model can work with in one request, in tokens: your prompt, conversation history and files. A token is about three-quarters of an English word. | 128,000 tokens | in place when tracking began (October 6, 2026) | source |
| Max output tokensThe longest reply the model can write in one response, in tokens. A token is about three-quarters of an English word. | 16,384 tokens | in place when tracking began (October 6, 2026) | source |
Not published by OpenAI: PDF input, Reasoning / thinking, Built-in web search, temperature parameter, top_p parameter, logprobs.
Price history
USD per 1M tokens, standard tierNo price history yet.
Prices
USD per 1M tokens, every type and band, by tierNo published prices.
Changelog
Changes we detected on OpenAI's pagesNo changes recorded yet.
Last checked · Sources: OpenAI deprecations, OpenAI pricing, OpenAI models