Accelerating GPT-5.6 Sol Ultrafast

TL;DR Highlight

Running GPT-5.6 Sol on Cerebras hit up to 750 tokens/s with no quality loss

Who Should Read

Backend developers weighing speed against quality because coding agents or LLM API latency is slowing their workflows

Core Mechanics

Ultrafast is a new service tier
- It is a new service tier launching first in the OpenAI API, powered by Cerebras.
- It is a limited preview for select customers now, with access expanding over time.
Output speed up to 750 tokens/s
- Cerebras says this comes without quality compromise.
- Based on Artificial Analysis reported speeds, it is 11x faster than Fable 5 and 5x faster than Opus 4.8 on Fast mode.
HLE solved about 7x faster
- GPT-5.6 Sol Ultrafast took 11h 11m for 2,500 HLE questions, versus 78h 27m for Claude Fable 5.
- Accuracy was described as comparable, with both run at xhigh reasoning in Codex and Claude Code.
GDP-Val end-to-end 5.6x speedup
- The speedup is versus regular GPT-5.6 Sol, with no quality degradation claimed.
- It was measured at medium reasoning within Codex.
The speed comes from avoiding data movement

Community Reactions

Opinion that speed buys quality
- One commenter argued that repeated review and revision improves answers a lot, but people skip it because it takes time.
- They said near-instant responses could bake this loop into the harness by default.
Counterpoint: faster tokens don't speed everything
- Commenters noted 1-hour e2e tests, 10-minute typechecks, and slow grep over big codebases stay the same.
- They also suspected HLE runs without tool calls, so it differs from everyday work.
Opinion that parity and pricing are unclear
- One commenter said neither post flatly states performance equals regular Sol.
- Several commenters noted that no pricing was given.
Experience that quota will drain fast
- One commenter said Codex Fast is already 1.5x faster at 2x cost, so 14x faster could burn the quota in one session.
- Another said they'd gladly burn through a bigger subscription faster if per-token price stayed the same.

How to Apply

Add an automatic self-review step to your agent
- Within 30 minutes today, send a follow-up prompt after a task finishes asking the model to define criteria, review, and improve.
- Log the number of rounds and time taken until gains shrink.
Time each stage of your agent's work
- Log LLM generation time separately from tests, typecheck, and grep for one task.
- Mark tasks where LLM generation dominates as candidates for a faster tier.
Route tasks between fast and Standard tiers
- Send urgent work like incident analysis to the fast tier and bulk repetitive work to Standard.
- Keep the tier as a config value since Ultrafast is limited preview until you get access.

Try It

copy & run

# Paste into a Claude or ChatGPT chat right after a complex task finishes (self-review loop from the HN comments)

Review what you just wrote. First come up with a list of evaluation criteria suited to this task, then score your answer against them. Based on the results, write a better answer if warranted.

Task: {{task description}}

Next to try: repeat the same prompt until gains diminish, and in Claude Code say "Use sub-agents to iteratively review and iterate on your work until convergence."

Terminology

Ultrafast modeA new service tier in the OpenAI API that runs the same model on Cerebras hardware for more speed. Think of it as a fast lane.

Wafer-Scale EngineA chip as big as a whole silicon wafer. Instead of linking many small chips, everything sits on one piece.

SRAMFast memory on the chip itself. Like keeping documents spread on your desk instead of fetching them from a warehouse.

Humanity's Last Exam (HLE)A hard benchmark of 2,500 questions typically answerable only by PhD-level experts.

GDP-ValA benchmark of economically valuable knowledge work tasks.