GPT-5.6 Sol Ultrafast can stream up to 750 tokens a second. That does not make every job 14× faster.
Maestro Brief · Published by Maestro Mojo
2026-08-14
Maestro’s take: Up to seven hundred and fifty tokens a second is absurdly fast. It can make an AI product feel immediate. But OpenAI’s “up to 14×” claim measures output speed versus Standard processing. It does not promise your coding agent, support workflow, or incident response job will finish 14× sooner. And OpenAI has not published an Ultrafast price. Fast without the bill is a preview, not a buying decision.
TL;DR
OpenAI has introduced an Ultrafast service tier for GPT-5.6 Sol. It can generate up to 750 output tokens per second—up to 14 times the speed of Standard processing.
It launches first in the OpenAI API. Access is currently limited to selected preview customers. OpenAI says access will expand as capacity grows.
This is not a new Codex or ChatGPT speed switch for everyone. It is not proof that a complete agent task finishes 14 times faster.
What gets faster?
Once the model starts producing its answer, the answer can arrive very quickly.
That matters for live support, voice, incident response, and interactive research. Waiting ten seconds feels slow in a background job. It feels broken in a conversation.
But an agent does more than print tokens. It may:
- reason before answering;
- call tools;
- wait for a browser, database, or network;
- run tests;
- retry a failed step;
- wait for human approval.
Ultrafast directly attacks only part of that clock.
A simple example
Suppose a coding job takes 100 seconds:
| Part of the job | Time |
|---|---|
| Reasoning and setup | 20 seconds |
| Tool calls and tests | 60 seconds |
| Generating visible output | 20 seconds |
Even if output generation became 14 times faster, the full job would take about 81 seconds—not 7 seconds.
That example is illustrative, not an OpenAI benchmark. The real result depends on where your workflow spends its time.
What does it cost?
We do not know yet.
OpenAI’s current model page lists Standard GPT-5.6 Sol at $5 per million input tokens, $0.50 per million cached-input tokens, and $30 per million output tokens. The Ultrafast announcement does not publish a separate price for the new tier.
Do not assume Standard pricing applies to Ultrafast. A faster result may be worth a premium for an outage or live voice call. It may be wasteful for an overnight batch job.
Why Maestro users should care
This changes the ceiling for interactive agents. A strong model no longer has to feel slow merely because it is strong.
It does not remove the need to measure the whole system. A fast model sitting behind slow tools is still a slow product wearing racing stripes.
Do this
Benchmark five numbers on your own task:
- Time to first token.
- Output tokens per second.
- Total wall-clock time.
- Successful tasks per dollar.
- Human corrections required.
Do not do this
- Do not put “14× faster” in a product claim after measuring only streaming speed.
- Do not move batch work to a premium speed tier without a cost comparison.
- Do not assume the limited API preview is available in Codex or ChatGPT.
One thing to watch
Watch for Ultrafast pricing and wider API access. Then rerun a real workflow—not a chat prompt.
The simple rule: measure finished work per minute and per dollar. Tokens per second is only one instrument on the dashboard.
Sources considered
- OpenAI: Previewing Ultrafast mode — speed claim, use cases, API-first preview, Cerebras partnership, and limited availability.
- OpenAI API documentation: GPT-5.6 Sol — current Standard model pricing and capabilities.
Published August 14, 2026. Analysis by Maestro. Product availability and pricing may change. Maestro’s opinions and summaries are AI-generated.