GPT 4.1 in the API

OpenAI is rolling out GPT 4.1 as a new developer-focused model family in the API—three sizes built for different latency and cost needs—while adding a major capability jump: up to 1 million tokens of context across all three models, including GPT 4.1 Nano. OpenAI positions GPT 4.1 as a “powerhouse” for coding, complex instruction following, agent building, and long-context work, with GPT 4.1 Mini for faster responses on simpler tasks and GPT 4.1 Nano for high-volume workloads like autocomplete, classification, and extracting information from long documents.

On coding, GPT 4.1 is measured through SWE-bench, a benchmark where a model explores a Python repository, writes code, and adds tests. GPT 4.1 reaches 55% accuracy, up from 33% on the prior GPT-4.0 line, and it also outperforms smaller variants like o1 and o3-mini in that evaluation. OpenAI also highlights improvements beyond Python using Aider Polyglot, emphasizing better handling of diff-style edits—useful for reducing both latency and cost because unchanged tokens need not be regenerated. The company says GPT 4.1 closes the gap on whole-file versus diff performance and doubles its diff performance compared with GPT-4.0.

Instruction following is treated as a day-to-day reliability problem rather than a “nice to have.” OpenAI describes an internal instruction-following eval that mirrors how API developers structure prompts, with multi-part instruction sets covering formatting, ordered steps, overconfidence traps, and other categories. Results show GPT 4.1 performing strongly across difficulty levels, including hard cases where models must obey strict formatting and negative constraints. A concrete example: a trip-planning prompt requires the output to be a table with a specific number of rows and columns; OpenAI claims GPT 4.1 follows such constraints without the prompting tricks developers previously used.

Long-context performance is backed by multiple evaluations. OpenAI says the context window jumps from 128K tokens in prior models to 1M tokens (an 8x increase), then tests whether the model can actually retrieve information at any depth using a “needle in a haystack” setup. OpenAI reports that GPT 4.1, GPT 4.1 Mini, and even GPT 4.1 Nano can find the inserted text across the beginning, middle, end, and full 1M range. A more complex long-context benchmark (OpenAI MRCR, shared on Hugging Face) uses synthetic multi-turn conversations where the model must retrieve the correct item while avoiding confusion from similar content.

Multimodal capability also gets attention. On the Video MME benchmark—multiple-choice questions over 30–60 minute videos without subtitles—GPT 4.1 reaches 72% and is described as state of the art. For image and multimodal reasoning, OpenAI calls GPT 4.1 Mini the top choice.

The rollout includes practical demos in the OpenAI Playground: generating a full single-file Python web app that answers questions over a ~450,000-token NASA server log, and enforcing structured prompting rules via XML-wrapped queries that prevent the model from answering outside the provided log data. OpenAI ties improvements to a developer traffic sharing program that scrubs PII and uses opt-in data to create evals.

Pricing and deployment changes round out the announcement: GPT 4.1 is said to be 26% cheaper than GPT-4.0, GPT 4.1 Nano costs 12 cents per million blended tokens, and there is no pricing bump for long context. OpenAI also plans to deprecate GPT 4.5 in the API over roughly three months to reclaim GPUs. Early testers at Windsurf report a 60% improvement on end-to-end coding benchmarks and claim fewer degenerate behaviors, less unnecessary file access, and reduced verbosity. GPT 4.1 is available now in the API, with fine-tuning supported for GPT 4.1 and GPT 4.1 Mini and expected soon for Nano.