Coding improvements in new OpenAI GPT models

OpenAI recently launched three new models in the API: GPT‑4.1, GPT‑4.1 mini, and GPT‑4.1 nano. These models outperform GPT‑4o and GPT‑4o mini across the board, with major gains in coding and instruction following. They also have larger context windows—supporting up to 1 million tokens of context, and are able to better use that context with improved long-context comprehension. They feature a refreshed knowledge cutoff of June 2024.

Introducing GPT-4.1 in the API: Major coding improvements in new OpenAI GPT models

GPT‑4.1 performance highlights:

Coding:

• GPT‑4.1 scores 54.6% on SWE-bench Verified, improving by 21.4%abs over GPT‑4o and 26.6%abs over GPT‑4.5—making it a leading model for coding.
   

Instruction following:

• On Scale’s MultiChallenge benchmark, GPT‑4.1 scores 38.3%, a 10.5%abs increase over GPT‑4o.
   

Long context:

• On Video-MME, a benchmark for multimodal long context understanding, GPT‑4.1 sets a new state-of-the-art result—scoring 72.0% on the long, no subtitles category, a 6.7%abs improvement over GPT‑4o.
   

While benchmarks provide valuable insights, we trained these models with a focus on real-world utility. Close collaboration with the developer community enabled us to optimize these models for the tasks that matter most.

GPT 4.1 Family Intelligence by Latency

Cost and latency improvements:

The GPT‑4.1 model family offers exceptional performance at a lower cost, pushing forward at every point on the latency curve.

GPT‑4.1 mini:

• A significant leap in small model performance, surpassing GPT‑4o in many benchmarks. It reduces latency by nearly half and cost by 83%.
   

GPT‑4.1 nano:

• The fastest and cheapest model available. Ideal for low-latency tasks like classification or autocompletion, with a 1 million token context window.
   

Scores:

• MMLU: 80.1%
• GPQA: 50.3%
• Aider polyglot coding: 9.8%
   

Real-world applications and developer feedback:

The GPT‑4.1 models improve reliability and long context comprehension, making them ideal for powering agents that perform tasks independently on behalf of users.
Early testers noted that GPT‑4.1 can be more literal, so explicit and specific prompts are recommended.

Deprecation notice:

GPT‑4.5 Preview will be deprecated on July 14, 2025, as GPT‑4.1 offers improved performance at lower cost and latency. We will maintain the creativity, writing quality, humor, and nuance appreciated in GPT‑4.5 in future models.

Benchmark performance and real-world usage:

GPT‑4.1 demonstrates significant improvements across coding, instruction following, and long context handling. It excels in:

• Coding tasks: Agentically solving coding tasks, reliable code diffs, and frontend coding.
• Instruction following: Improved format compliance, multi-turn instructions, and reduced overconfidence.
• Long context processing: Efficient retrieval from up to 1 million tokens of input.
   

Real-world examples include improvements in coding benchmarks with Windsurf, accurate legal data extraction with Thomson Reuters, and fast, reliable code generation with Qodo.

Vision and multimodal capabilities:

The GPT‑4.1 family excels at image understanding and processing long videos without subtitles, making it suitable for multimodal applications. GPT‑4.1 series models are available now to all developers, with lower prices through efficiency improvements:

GPT-4.1 pricing:

• Input: $2.00
• Cached Input: $0.50
• Output: $8.00
• Blended Pricing: $1.84
   

GPT-4.1 mini pricing:

• Input: $0.40
• Cached Input: $0.10
• Output: $1.60
• Blended Pricing: $0.42
   

GPT-4.1 nano pricing:

• Input: $0.10
• Cached Input: $0.025
• Output: $0.40
• Blended Pricing: $0.12
   

GPT‑4.1 represents a major leap in practical AI application, addressing real-world developer needs from coding to long context comprehension. We look forward to seeing the innovations that the developer community builds using these models.