GPT-4.1 API Release Notes / What's New
OpenAI has released a new series of API models—GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano—that outperform GPT-4o and GPT-4o mini in coding, instruction following, and long-context comprehension. These models introduce a expanded context window of up to 1 million tokens and a refreshed knowledge cutoff of June 2024.
Enhanced Coding Capabilities
GPT-4.1 is designed for high-performance software engineering, showing significant gains in agentic task solving and code editing reliability.
- Software Engineering Performance: GPT-4.1 scored 54.6% on SWE-bench Verified, a 21.4% absolute improvement over GPT-4o and 26.6% absolute improvement over GPT-4.5.
- Code Diff Reliability: The model more than doubles GPT-4o's score on Aider's polyglot diff benchmark and exceeds GPT-4.5 by 8% absolute. This allows developers to output only changed lines rather than rewriting entire files, reducing cost and latency.
- Frontend Development: In head-to-head comparisons, human graders preferred GPT-4.1's generated websites over GPT-4o's 80% of the time.
- Reduced Noise: Internal evaluations show that extraneous edits on code dropped from 9% with GPT-4o to 2% with GPT-4.1.
Improved Instruction Following and Reliability
GPT-4.1 demonstrates higher reliability across various complex instruction categories, including format following (XML, YAML, Markdown), negative instructions, ordered instructions, and ranking.
- Benchmark Gains: GPT-4.1 scored 38.3% on Scale's MultiChallenge benchmark (a 10.5% absolute increase over GPT-4o) and 87.4% on IFEval (compared to 81.0% for GPT-4o).
- Multi-turn Coherence: The model is trained to better retrieve information from past messages, improving coherence in deep conversations.
- Real-world Application: Early testers, such as Blue J, reported a 53% increase in accuracy on challenging real-world tax scenarios compared to GPT-4o. Hex reported a nearly 2x improvement on their most challenging SQL evaluation set.
Long-Context Understanding and Reasoning
All models in the GPT-4.1 family now support a context window of 1 million tokens, significantly increasing the capacity for processing large codebases or extensive document sets.
- Retrieval Accuracy: In "needle in a haystack" evaluations, GPT-4.1, mini, and nano consistently retrieved information accurately across all positions up to 1 million tokens.
- Complex Reasoning: To evaluate multi-hop reasoning, OpenAI introduced the Graphwalks dataset. GPT-4.1 achieved 61.7% accuracy on this benchmark, matching OpenAI o1 and exceeding GPT-4o.
- Multi-Round Coreference: OpenAI also open-sourced the OpenAI-MRCR eval, where GPT-4.1 outperforms GPT-4o at context lengths up to 128K and maintains strong performance up to 1 million tokens.
- Enterprise Impact: Thomson Reuters reported a 17% improvement in multi-document review accuracy for CoCounsel, while Carlyle saw 50% better retrieval from very large, dense documents.
Model Family Tiers and Performance
OpenAI has introduced three distinct tiers to balance intelligence, latency, and cost:
- GPT-4.1: The flagship model for the most complex tasks, offering the highest intelligence and the largest context window.
- GPT-4.1 mini: A high-efficiency model that matches or exceeds GPT-4o in intelligence evals while reducing latency by nearly half and reducing cost by 83%.
- GPT-4.1 nano: The fastest and cheapest model available, scoring 80.1% on MMLU and 50.3% on GPQA. It is optimized for low-latency tasks like classification and autocompletion.
Vision and Multimodal Capabilities
GPT-4.1 provides advanced image understanding, with GPT-4.1 mini frequently beating GPT-4o on image benchmarks. The model also sets a new state-of-the-art in long-video understanding, scoring 72.0% on the Video-MME (long, no subtitles) category, a 6.7% absolute improvement over GPT-4o.
Pricing and Availability
GPT-4.1 models are available via the API. GPT-4.5 Preview will be deprecated on July 14, 2025.
| Model | Input (per 1M) | Cached Input (per 1M) | Output (per 1M) | 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 |
Key Pricing Updates:
- Prompt Caching: The discount for repeated context has increased to 75% (from 50%).
- Long Context: Requests up to 1 million tokens are offered at no additional cost beyond standard per-token rates.
- Batch API: A 50% discount is available for use via the Batch API.
Sources
- OriginalIntroducing GPT-4.1 in the API