GPT-4.1

GPT-4.1 is an OpenAI language model released on April 14, 2025, available exclusively via API (not in ChatGPT at launch). API snapshot: gpt-4.1-2025-04-14. It was designed for developers building agentic coding systems, with emphasis on instruction following and long-context tasks.

Key features

Context window of 1,047,576 tokens (1M); maximum output tokens: 32,768. Knowledge cutoff: June 2024. Supports tool use, fine-tuning, and multimodal input (text, image, documents).

Benchmark results

On SWE-bench Verified, it achieves 54.6% (conservative score 52.1%) — an improvement of 21.4 pp. over GPT-4o (33.2%). On Aider Polyglot diff, it scores 52.9% (2.9× better than GPT-4o). MMLU 90.2%, MMMU 74.8%, MathVista 72.2%, Video-MME (long) 72.0%. On the Needle in Haystack test at 1M token context — 100% recall.

Pricing and availability

Closed-weights model, available through the OpenAI API, Azure AI Foundry, and other hosting platforms. Pricing: $2/MTok input, $8/MTok output, cached input $0.50/MTok (75% discount). Batch API with 50% discount. No price premium for long context up to 1M tokens. Model retired from ChatGPT on February 13, 2026; still available via API.

Safety

OpenAI did not publish a separate system card, classifying the model as non-frontier. Independent research (Owain Evans/Oxford, ICML 2025; SplxAI) identified elevated misalignment risk after fine-tuning on unsafe code, as well as a tendency toward literal, more easily circumvented instruction following. In response, OpenAI published a dedicated prompting guide.