OpenAI drops GPT-4o with massive 64K output context

Anyone who has tried asking an AI to generate an entire code repository or a deep-dive industry report knows the annoyance of hitting an invisible wall mid-sentence, forcing you to type please continue.

OpenAI just rolled out an experimental version of its flagship model, dubbed GPT-4o Long Output, that unlocks up to 64K output tokens per single request.

Available right now to participants in OpenAI's alpha program under the model name gpt-4o-64k-output-alpha, the release is explicitly designed to unlock new use cases that demand massive single-shot completions.

Longer outputs come with a steeper bill

While getting tens of thousands of tokens in one go sounds ideal, it comes with a noticeable catch on the pricing sheet.

OpenAI noted that long completions are significantly more expensive from an inference standpoint. As a result, the per-token price tag has gone up to match those computational costs.

Running gpt-4o-64k-output-alpha will set developers back $6.00 per million input tokens and $18.00 per million output tokens.

In a related development on the infrastructure front, Microsoft also rolled out the GPT-4o mini API with vision support across Global and East US regional standard deployments on Azure, giving developers image and video processing capabilities in their apps.

Why it matters

We think this experimental release tackles an asymmetric limitation that has frustrated power users for months. While frontier models have offered massive input context windows for a while, output generation has almost always been bottlenecked at around 4K tokens, forcing engineering teams to build clunky recursive prompting pipelines just to assemble a long document or full codebase.

Allowing a single call to output 64K tokens eliminates a massive layer of orchestration headaches. If you are building automated coding agents, technical documentation generators, or deep research tools, this changes what single-shot prompts can actually deliver.

However, the economics and latency will be the real hurdle. At $18 per million output tokens, a fully maxed-out 64K completion costs over a dollar per call, which quickly adds up if the output drifts off-target. Long generations also introduce practical engineering issues around network timeouts and waiting for tokens to stream. Still, it signals that OpenAI is finally looking to match output bandwidth with input capacity.