OpenAI has introduced a new generation of models—GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano—delivering marked improvements in code generation, instruction adherence, and long-context reasoning. Available exclusively through the API, GPT-4.1 represents a significant optimization over GPT-4o, offering better performance at reduced cost and latency.
On real-world coding tasks, GPT-4.1 achieves 54.6% on SWE-bench Verified, up from 33.2% with GPT-4o. It also improves diff-based editing, following structured changes with higher precision while reducing unnecessary edits from 9% to 2%. In frontend generation tests, human reviewers preferred GPT-4.1’s output 80% of the time.
Instruction-following gains are evident in hard-prompt scenarios, where GPT-4.1 scores 49.1% versus GPT-4o’s 29.2%, and 87.4% on IFEval. It is more literal and reliable when executing ordered or constrained instructions, handling formats like XML or YAML with better consistency.
GPT-4.1 supports up to 1 million tokens of context and outperforms previous models in benchmarks like OpenAI-MRCR and Graphwalks. These tests evaluate its ability to resolve coreferences across long texts and perform multi-hop reasoning across structured contexts—key capabilities for legal, financial, and software agent applications.
For developers prioritizing efficiency, GPT-4.1 mini offers lower latency and cost (83% cheaper than GPT-4o) while still outperforming it in many benchmarks. GPT-4.1 nano, the fastest and most affordable model in the family, achieves strong results on classification and code completion with minimal compute overhead.
The pricing model reflects these optimizations. GPT-4.1 input tokens cost $2.00 per million ($0.50 cached), while output costs $8.00. Mini and nano versions scale down both cost and latency proportionally. Prompt caching discounts have increased to 75%, further improving affordability for repeat queries.
For a detailed breakdown of benchmark scores, model behavior, and deployment examples from early adopters like Thomson Reuters, Blue J, Hex, and Carlyle, see OpenAI’s full GPT-4.1 model announcement.