GPT-4.1 Nano: The 1M-Token Small Model and Its October Shutdown

Small models are usually sold on one property: they are cheap. GPT-4.1 nano was unusual because it was cheap and also carried the same million-token context window as its much larger sibling. That combination made it the default for a specific kind of job, and it is the reason a lot of pipelines still call it.

Those pipelines have a deadline. OpenAI announced on 22 April 2026 that gpt-4.1-nano and its dated snapshot shut down on 23 October 2026, six days after this article was published, with gpt-5.6-luna named as the replacement. This guide covers what nano was actually good at, so you know what to protect when you move.

What nano was built for

OpenAI released the GPT-4.1 family on 14 April 2025 as three API models: GPT-4.1, GPT-4.1 mini and GPT-4.1 nano. Nano was positioned as the fastest and cheapest of the three, and OpenAI's launch material pointed it at classification and autocompletion rather than open-ended reasoning.

The current model page describes it the same way: strong at instruction following and tool calling, low latency, and no reasoning step. That last part matters. Nano answers immediately rather than spending hidden tokens thinking first, which keeps latency and cost predictable on short, high-volume calls.

OpenAI's published scores for nano were 80.1% on MMLU and 50.3% on GPQA, and 9.8% on Aider's polyglot coding benchmark. Read those together. The first two say it knows a fair amount for its size. The third says plainly that it is not a coding model, and nobody should have been routing multi-file edits to it.

A million tokens on the smallest model

OpenAI lists the context window at 1,047,576 tokens, with up to 32,768 output tokens and a knowledge cutoff of June 2024. Input can be text or images; output is text.

The large window on a small model enabled one genuinely useful pattern: cheap first-pass triage over a big document. Drop a long contract, a log dump or a large transcript into nano, ask it to find the sections relevant to a question, then send only those sections to a stronger model. That is retrieval done by a model rather than by an embedding index, and at nano prices it was often cheaper than building the index.

The caveat is the usual one. An advertised window is not the same as reliable attention across it, and small models tend to degrade faster as prompts grow. Context length vs effective context covers how to measure where that drop-off sits for a given model, which is worth doing before trusting any long-context triage step.

The cost arithmetic

OpenAI's standard price is $0.10 per million input tokens and $0.40 per million output tokens, with cached input at $0.025. The Batch API halves the standard rate to $0.05 and $0.20, and OpenRouter lists the same figures for its :batch endpoint, which is the same model on a discounted asynchronous queue rather than a different model. Prices are as listed at the time of writing.

A worked example makes the scale clear. Classifying one million support tickets at around 400 input tokens and 10 output tokens each is 400 million input tokens and 10 million output tokens. At standard rates that is about $40 plus $4. Through the Batch API it is roughly half that. The cache discount helps further when every request shares a long instruction prefix, because the prefix is billed at a quarter of the input rate once it is cached.

Numbers like that are why nano became a routing tier rather than a model anyone chose for its intelligence. Model routing and fallbacks covers the pattern: cheap model first, escalate when confidence is low.

Nano against GPT-4o mini

The obvious comparison is GPT-4o mini, which filled the same small-and-cheap slot a year earlier. On paper nano wins on both axes that usually decide a routing tier:

• Price. Nano lists at $0.10 and $0.40 against mini's $0.15 and $0.60 per million tokens.
• Window. Nano accepts roughly a million tokens; mini stops at 128,000.
• Freshness. Nano's knowledge runs to June 2024; mini's to October 2023.

Mini still has one advantage now: time. OpenAI's October 2026 notice gives gpt-4o-mini-2024-07-18 until 14 January 2027, nearly three months after nano disappears. Moving a nano workload onto mini to buy time is possible, but it is a sideways step onto another retiring model, and you would repeat the migration in January.

Within its own family, pick nano over GPT-4.1 mini only when the task is narrow enough that a label or a short extraction is the whole output. Anything that needs judgement across several steps was a mini or full GPT-4.1 job from the start.

Migrating a nano workload without regressions

Classification pipelines fail quietly. A replacement model that is smarter overall can still assign a different label to the borderline cases your downstream code was tuned around. Treat the move as a measured change rather than a model ID swap:

1. Freeze a labelled set. Pull a few hundred real inputs, including the ambiguous ones, with the labels nano produced and the labels a human agrees with.
2. Run the replacement in shadow. Send the same inputs to gpt-5.6-luna or whichever model you choose, and diff the outputs against both nano and the human labels.
3. Check the format, not just the answer. If you relied on structured outputs or a fixed label vocabulary, confirm the replacement respects the same schema without extra prose.
4. Compare cost per thousand items. A newer model may emit more tokens per answer, which matters more than its list price on a high-volume job.

Fine-tuned nano models are covered by the same notice: ft-gpt-4.1-nano-2025-04-14 also shuts down on 23 October. If you fine-tuned nano, budget time to rebuild the training run on a supported base, not just to change a string. Model deprecation and migration has a fuller checklist for this kind of forced move.