OpenAI's new 'Ultrafast' mode makes its top AI model answer 14x faster
A limited preview lets select customers run GPT-5.6 Sol at up to 750 words per second, powered by custom Cerebras chips, with no reported drop in answer quality.
What happened: OpenAI has released a new processing mode called Ultrafast for its newest flagship model, GPT-5.6 Sol. The company says it runs about 14 times faster than the model's standard mode, generating up to 750 output tokens (roughly words) per second. The speed comes from a partnership with chipmaker Cerebras, whose custom hardware runs the mode. Ultrafast is currently in limited preview for a small set of customers, with OpenAI saying access will expand "as capacity grows." The company suggests it for incident response, customer support, financial analysis and e-commerce, where even a few seconds of delay matters.
Why it matters: AI builders have long had to pick between a smarter, slower model or a faster, weaker one, because bigger models take longer to move data through memory. Ultrafast claims to narrow that gap: OpenAI and Cerebras report it worked through a punishing 2,500-question academic benchmark in about 11 hours, versus roughly 78 hours for Anthropic's fastest current mode, without a reported accuracy loss. If that holds up outside curated benchmarks, it could bring frontier-level reasoning to tasks — live fraud checks, on-call engineering triage, real-time trading calls — that today rely on smaller, faster but less capable models purely because of speed limits.
How it works, plainly: Running a huge AI model on standard chips (GPUs) usually means constantly shuttling its "knowledge" — the model's weights — between fast on-chip memory and slower external storage, and that shuttling is what slows big models down. Cerebras builds giant wafer-sized chips that pack 44 gigabytes of memory directly onto the chip itself, so weights can stay put while text streams through a pipeline of chips without the usual traffic jam. Combined with OpenAI's own optimizations, that's what lets GPT-5.6 Sol hit up to 750 tokens per second in Ultrafast, instead of its normal pace.
The rollout: Ultrafast launched Thursday, August 13, 2026, inside the OpenAI API, but only for a "select group" of customers; there's no public timeline or pricing yet for wider access. It's pitched as a premium tier for the most time-sensitive, highest-stakes work, while routine tasks stay on standard processing. Rivals are moving in the same direction — Anthropic already offers a "fast mode" for Claude — though OpenAI and Cerebras's claim that Ultrafast is faster still comes from the companies selling it, not an independent outside test.