April 14, 2025
A new generation of GPT models offering major enhancements in coding, instruction-following, and long-context capabilities—plus our first-ever nano model.
Explore it in the Playground • Share
We’re excited to introduce three new models via the API: GPT‑4.1, GPT‑4.1 mini, and GPT‑4.1 nano. These additions outpace GPT‑4o and GPT‑4o mini on all fronts, showing particularly strong advances in coding and instruction-following skills. They also offer significantly larger context windows—up to 1 million tokens—and are better at leveraging that context, thanks to improved long-context comprehension. Additionally, they feature an updated knowledge cutoff of June 2024.
Although benchmarks illustrate substantial gains, the driving goal behind GPT‑4.1 was to excel in real-world applications. We collaborated closely with the developer community to fine-tune these models for practical use cases that matter most to you.
In short, the GPT‑4.1 family offers outstanding performance at lower cost—continuing to push the boundaries of capability and efficiency.
The GPT‑4.1 lineup delivers strong results across different latency tiers:
These improvements make GPT‑4.1 models notably more robust for powering “agentic” systems that can handle tasks independently. Combined with primitives like the Responses API, developers can build agents that more effectively tackle software engineering, large-document analysis, customer support, and other complex operations.
Note: GPT‑4.1 is only available via the API. In ChatGPT, many instruction-following, coding, and intelligence upgrades have gradually been integrated into the latest GPT‑4o release. We will continue to carry over more GPT‑4.1 innovations in upcoming ChatGPT updates.
We will begin deprecating GPT‑4.5 Preview in the API. GPT‑4.5 Preview will be shut down on July 14, 2025, giving developers time to transition. GPT‑4.1 provides similar or better performance, plus reduced latency and cost. GPT‑4.5 was initially a research preview for large, compute-intensive models, and we’re carrying forward its strengths—creativity, writing quality, humor, nuance—into future models.
Below, we break down GPT‑4.1’s improvements on standard benchmarks, along with real-world use cases from early adopters like Windsurf, Qodo, Hex, Blue J, Thomson Reuters, and Carlyle. These examples highlight how GPT‑4.1 performs on tasks ranging from coding to analyzing lengthy legal documents.
GPT‑4.1 exhibits substantial gains in:
On SWE-bench Verified (real-world software engineering tasks), GPT‑4.1 successfully completes 54.6% of tasks, well above GPT‑4o’s 33.2%. This reflects its better grasp of navigating codebases, producing functional code, and passing tests.
| Model | SWE-bench Verified Accuracy |
|---|---|
| GPT-4.1 | 54.6% |
| GPT-4o (2024-11-20) | 33.2% |
| GPT-4.5 | 38.0% |
| GPT-4.1 mini | 23.6% |
| GPT-4o mini | 8.7% |
| OpenAI o1 (high) | 41.0% |
| OpenAI o3-mini (high) | 49.3% |
We excluded 23 out of 500 tasks that couldn’t be run under our infrastructure. Assigning 0% to these tasks drops GPT‑4.1’s accuracy slightly, from 54.6% to about 52%.
GPT‑4.1 also excels at code diffs. On Aider’s polyglot diff benchmark—requiring a model to produce accurate edits in different programming languages—GPT‑4.1 more than doubles GPT‑4o’s score and even outperforms GPT‑4.5 by an 8% margin. This benchmark highlights both coding skill and strict adherence to patch formats.
| Model | Whole-file Format | Diff Format |
|---|---|---|
| GPT-4.1 | 51.6% | 52.9% |
| GPT-4o (2024-11-20) | 30.7% | 18.2% |
| GPT-4.5 | 38.0% | 32.0% |
| GPT-4.1 mini | 34.7% | 31.6% |
| GPT-4.1 nano | 9.8% | 6.2% |
| GPT-4o mini | 3.6% | 2.7% |
| OpenAI o1 (high) | 64.6% | 61.7% |
| OpenAI o3-mini (high) | 66.7% | 60.4% |
By outputting only the changed lines (rather than the full file), developers can save time and cost. For those who prefer rewriting entire files, GPT‑4.1’s output token limit is 32,768 tokens, double GPT‑4o’s 16,384.
GPT‑4.1 significantly outperforms GPT‑4o in creating user interfaces and web apps. In controlled tests where paid human evaluators assessed head-to-head website generation, GPT‑4.1’s results were judged superior 80% of the time.
Example Prompt:
“Make a flashcard web application…”
(Details omitted for brevity.)
| GPT‑4o Output | GPT‑4.1 Output |
|---|---|
| Basic UI and partial features | More polished UI, better animations, better code structure |
Internal tests showed that GPT‑4.1 commits unnecessary edits only 2% of the time, down from 9% with GPT‑4o. This leads to more focused and efficient coding sessions.
GPT‑4.1 is markedly more reliable at following instructions, reflected in multiple benchmarks and internal evaluations.
We developed an internal test covering:
GPT‑4.1 shows a dramatic jump in success on the “hard” subset of prompts, from 29% with GPT‑4o to 49%.
| Model | Accuracy (Hard Subset) |
|---|---|
| GPT-4.1 | 49% |
| GPT-4o (2024-11-20) | 29% |
| GPT-4.5 | 54% |
| GPT-4.1 mini | 45% |
| GPT-4.1 nano | 32% |
| GPT-4o mini | 27% |
| OpenAI o1 (high) | 51% |
| OpenAI o3-mini (high) | 50% |
MultiChallenge from Scale tests how well a model handles multi-turn conversations while retaining details from earlier prompts. GPT‑4.1 improves by 10.5 percentage points over GPT‑4o, scoring 38.3%.
| Model | MultiChallenge |
|---|---|
| GPT-4.1 | 38% |
| GPT-4o (2024-11-20) | 28% |
| GPT-4.5 | 44% |
| GPT-4.1 mini | 36% |
| GPT-4.1 nano | 15% |
| GPT-4o mini | 20% |
| OpenAI o1 (high) | 45% |
| OpenAI o3-mini (high) | 40% |
IFEval tests whether a model can comply with various instructions (content length, tone, avoiding specific words, etc.). GPT‑4.1 scores 87.4%, up from 81.0% with GPT‑4o.
| Model | IFEval Accuracy |
|---|---|
| GPT-4.1 | 87% |
| GPT-4o (2024-11-20) | 81% |
| GPT-4.5 | 88% |
| GPT-4.1 mini | 84% |
| GPT-4.1 nano | 75% |
| GPT-4o mini | 78% |
| OpenAI o1 (high) | 92% |
| OpenAI o3-mini (high) | 94% |
All three GPT‑4.1 models—standard, mini, and nano—support up to 1 million tokens of context, a significant jump from GPT‑4o’s 128,000 tokens. This means GPT‑4.1 can effectively handle entire codebases or large document sets while maintaining a strong grasp of relevant details and ignoring distractors.
In our internal “needle in a haystack” test (finding a single piece of hidden info anywhere in up to 1 million tokens), GPT‑4.1 consistently retrieves the correct detail at any position in the text. Its accuracy remains stable across the entire context window.
We are introducing OpenAI-MRCR, an evaluation that tests how well a model can locate and disambiguate multiple “needles” hidden among numerous near-duplicate requests. GPT‑4.1 surpasses GPT‑4o in performance for context lengths up to 128K tokens and continues to perform well at 1 million tokens, though the task remains challenging.
Graphwalks is a dataset requiring multi-hop reasoning within large context windows. It involves BFS (breadth-first search) queries through a huge directed graph. GPT‑4.1 matches the performance of OpenAI o1 (high) at 61.7% accuracy, significantly beating GPT‑4o at 42%.
We’ve reworked our inference stack to reduce time-to-first-token. Prompt caching can further minimize wait times while reducing your costs. Under typical conditions, GPT‑4.1’s p95 latency for 128,000 tokens is around 15 seconds for the first token, and roughly 30 seconds for 1 million tokens. GPT‑4.1 mini and nano are faster still, with GPT‑4.1 nano often returning the first token in under five seconds at 128,000 tokens.
The GPT‑4.1 family excels at image-based tasks, with GPT‑4.1 mini often outperforming GPT‑4o on vision-oriented benchmarks.
| Benchmark | GPT-4.1 | GPT-4o (2024-11-20) | GPT-4.5 | GPT-4.1 mini | GPT-4.1 nano | GPT-4o mini |
|---|---|---|---|---|---|---|
| MMMU | 74.8% | 68.7% | 75.2% | 72.7% | 55.4% | 56.3% |
| MathVista | 72.2% | 61.4% | 72.3% | 73.1% | 56.2% | 56.5% |
| CharXiv-Reasoning | 56.7% | 52.7% | 55.4% | 56.8% | 40.5% | 36.8% |
In Video-MME tasks with 30–60 minute videos and no subtitles, GPT‑4.1 reaches 72.0%, up from GPT‑4o’s 65.3%, a new best-in-class result for video comprehension.
We’re pleased to offer these new GPT‑4.1 models at reduced rates compared to GPT‑4o, thanks to optimizations in our inference infrastructure. Additionally, GPT‑4.1 nano is now our cheapest and fastest model. We also increased the prompt caching discount to 75% (up from 50%) for all GPT‑4.1 models, and there’s no added cost for long-context usage beyond the standard per-token price.
| Model | Input | Cached Input | Output | Blended Pricing* |
|---|---|---|---|---|
| gpt-4.1 | $2.00 | $0.50 | $8.00 | $1.84 |
| gpt-4.1-mini | $0.40 | $0.10 | $1.60 | $0.42 |
| gpt-4.1-nano | $0.10 | $0.025 | $0.40 | $0.12 |
* Blended pricing is estimated from typical ratios of input, cached input, and output tokens.
All three models are available at an additional 50% discount when used in our Batch API.
GPT‑4.1 delivers significant improvements in real-world applications—especially in coding tasks, instruction-following, and handling extensive context. Its advanced capabilities empower developers to build more sophisticated, reliable applications and AI-powered “agents.” We can’t wait to see how the community leverages GPT‑4.1’s strengths to create the next generation of intelligent systems.
Below is a comprehensive table summarizing GPT‑4.1 performance across academic tests, coding evals, instruction-following tasks, long-context benchmarks, vision, and function-calling evaluations.
| Category | GPT-4.1 | GPT-4.1 mini | GPT-4.1 nano | GPT-4o (2024-11-20) | GPT-4o mini | OpenAI o1 (high) | OpenAI o3-mini (high) | GPT-4.5 |
|---|---|---|---|---|---|---|---|---|
| AIME '24 | 48.1% | 49.6% | 29.4% | 13.1% | 8.6% | 74.3% | 87.3% | 36.7% |
| GPQA Diamond 1 | 66.3% | 65.0% | 50.3% | 46.0% | 40.2% | 75.7% | 77.2% | 69.5% |
| MMLU | 90.2% | 87.5% | 80.1% | 85.7% | 82.0% | 91.8% | 86.9% | 90.8% |
| Multilingual MMLU | 87.3% | 78.5% | 66.9% | 81.4% | 70.5% | 87.7% | 80.7% | 85.1% |
| Category | GPT-4.1 | GPT-4.1 mini | GPT-4.1 nano | GPT-4o (2024-11-20) | GPT-4o mini | OpenAI o1 (high) | OpenAI o3-mini (high) | GPT-4.5 |
|---|---|---|---|---|---|---|---|---|
| SWE-bench Verified | 54.6% | 23.6% | – | 33.2% | 8.7% | 41.0% | 49.3% | 38.0% |
| SWE-Lancer | $176K (35.1%) | $165K (33.0%) | $77K (15.3%) | $163K (32.6%) | $116K (23.1%) | $160K (32.1%) | $90K (18.0%) | $186K (37.3%) |
| SWE-Lancer (IC-Diamond) | $34K (14.4%) | $31K (13.1%) | $9K (3.7%) | $29K (12.4%) | $11K (4.8%) | $29K (9.7%) | $17K (7.4%) | $41K (17.4%) |
| Aider’s polyglot: whole | 51.6% | 34.7% | 9.8% | 30.7% | 3.6% | 64.6% | 66.7% | – |
| Aider’s polyglot: diff | 52.9% | 31.6% | 6.2% | 18.2% | 2.7% | 61.7% | 60.4% | 44.9% |
| Category | GPT-4.1 | GPT-4.1 mini | GPT-4.1 nano | GPT-4o (2024-11-20) | GPT-4o mini | OpenAI o1 (high) | OpenAI o3-mini (high) | GPT-4.5 |
|---|---|---|---|---|---|---|---|---|
| Internal (hard subset) | 49.1% | 45.1% | 31.6% | 29.2% | 27.2% | 51.3% | 50.0% | 54.0% |
| MultiChallenge | 38.3% | 35.8% | 15.0% | 27.8% | 20.3% | 44.9% | 39.9% | 43.8% |
| MultiChallenge (o3-mini grader) | 46.2% | 42.2% | 31.1% | 39.9% | 25.6% | 52.9% | 50.2% | 50.1% |
| COLLIE | 65.8% | 54.6% | 42.5% | 50.2% | 52.7% | 95.3% | 98.7% | 72.3% |
| IFEval | 87.4% | 84.1% | 74.5% | 81.0% | 78.4% | 92.2% | 93.9% | 88.2% |
| Multi-IF | 70.8% | 67.0% | 57.2% | 60.9% | 57.9% | 77.9% | 79.5% | 70.8% |
| Category | GPT-4.1 | GPT-4.1 mini | GPT-4.1 nano | GPT-4o (2024-11-20) | GPT-4o mini | OpenAI o1 (high) | OpenAI o3-mini (high) | GPT-4.5 |
|---|---|---|---|---|---|---|---|---|
| OpenAI-MRCR: 2 needle (128k) | 57.2% | 47.2% | 36.6% | 31.9% | 24.5% | 22.1% | 18.7% | 38.5% |
| OpenAI-MRCR: 2 needle (1M) | 46.3% | 33.3% | 12.0% | – | – | – | – | – |
| Graphwalks BFS (<128k) | 61.7% | 61.7% | 25.0% | 41.7% | 29.0% | 62.0% | 51.0% | 72.3% |
| Graphwalks BFS (>128k) | 19.0% | 15.0% | 2.9% | – | – | – | – | – |
| Graphwalks Parents (<128k) | 58.0% | 60.5% | 9.4% | 35.4% | 12.6% | 50.9% | 58.3% | 72.6% |
| Graphwalks Parents (>128k) | 25.0% | 11.0% | 5.6% | – | – | – | – | – |
| Category | GPT-4.1 | GPT-4.1 mini | GPT-4.1 nano | GPT-4o (2024-11-20) | GPT-4o mini | OpenAI o1 (high) | GPT-4.5 |
|---|---|---|---|---|---|---|---|
| MMMU | 74.8% | 72.7% | 55.4% | 68.7% | 56.3% | 77.6% | 75.2% |
| MathVista | 72.2% | 73.1% | 56.2% | 61.4% | 56.5% | 71.8% | 72.3% |
| CharXiv-Reasoning | 56.7% | 56.8% | 40.5% | 52.7% | 36.8% | 55.1% | 55.4% |
| CharXiv-D | 87.9% | 88.4% | 73.9% | 85.3% | 76.6% | 88.9% | 90.0% |
| Category | GPT-4.1 | GPT-4.1 mini | GPT-4.1 nano | GPT-4o (2024-11-20) | GPT-4o mini | OpenAI o1 (high) | OpenAI o3-mini (high) | GPT-4.5 |
|---|---|---|---|---|---|---|---|---|
| ComplexFuncBench | 65.5% | 49.3% | 57.0% | 66.5% | 38.6% | 47.6% | 17.6% | 63.0% |
| Taubench Airline | 49.4% | 36.0% | 14.0% | 42.8% | 22.0% | 50.0% | 32.4% | 50.0% |
| Taubench Retail | 68.0% | 55.8% | 22.6% | 60.3% | 44.0% | 70.8% | 57.6% | 68.4% |
Thank you for exploring GPT‑4.1! We look forward to seeing how you harness these new models to build powerful applications and advanced AI-driven systems.