OpenAI Ultrafast Mode: GPT-5.6 Sol at 750 tok/s (2026)

OpenAI's Ultrafast mode is a new API service tier previewed on August 13, 2026, available only for the GPT-5.6 Sol model. It runs inference on Cerebras's Wafer-Scale Engine (WSE) and is officially rated at up to 750 output tokens per second, up to 14x faster than the standard tier. The announcement hit the top of Hacker News on August 14, but access remains a limited preview for select customers, and pricing for the mode has not been disclosed as of this writing.

A speed jump of several times to over an order of magnitude matters most for workloads that generate tokens repeatedly and at volume, rather than for a single chat reply. That includes multi-step agent execution where a model calls tools and verifies results across dozens of turns, large-scale batch evaluation or data generation across hundreds or thousands of prompts, and the perceived latency of interactive UIs. In agentic workloads especially, per-request output wait time accumulates directly into total task duration, so faster generation tends to translate almost linearly into shorter end-to-end runs. For GPT-5.6's broader pricing context, see our coverage of the GPT-5.6 API price cut.

Every multiplier in this table comes from OpenAI's or Cerebras's own published figures; independent third-party verification has not yet surfaced. Numbers attributed to other vendors reflect those vendors' own measurement conditions, so treat cross-vendor comparisons as directional rather than apples-to-apples.

In typical GPU inference, generating each token requires reading model weights out of HBM (high-bandwidth memory) into the compute cores, and that memory bandwidth effectively caps generation speed. Cerebras's Wafer-Scale Engine takes a different structural approach: it treats an entire silicon wafer as a single chip, carrying 44GB of SRAM per chip. Keeping model weights resident in this on-chip SRAM eliminates the off-chip memory round trip that normally happens on every generated token, sidestepping the memory-bandwidth bottleneck altogether — this is the mechanism OpenAI and Cerebras describe as the source of the speedup. For another example of Cerebras-hosted inference, see our piece on running Gemma 4 on Cerebras Inference.

OpenAI and Cerebras report completing the 2,500-question Humanity's Last Exam in 11 hours on Ultrafast, versus a stated 78 hours for Claude Fable 5 on the same benchmark — roughly a 7x difference. On GDP-Val, they claim an end-to-end 5.6x speedup with no quality degradation. All of these figures, however, are vendor-run and vendor-published benchmarks, and it is not clear from public information how much detail on evaluation conditions — prompts, sampling settings, scoring methodology — has been disclosed to third parties. The claim that speed gains come without accuracy loss is an important one, but it is reasonable to treat these numbers as preliminary until independent labs or the community can reproduce them.

Ultrafast is positioned as one of the service tiers available when calling GPT-5.6 Sol through the API. As of this writing it is a limited, application-based preview, with OpenAI planning to expand access to more customers over time. Exactly what changes in existing API code — specific endpoints or request parameters — to switch tiers is not confirmed in public information, and we avoid speculating on undocumented API details here. Anyone evaluating Ultrafast should check the primary sources directly: OpenAI's own announcement (Previewing Ultrafast) and Cerebras's official blog post.

The standard tier for GPT-5.6 Sol is priced at $5 per million input tokens and $30 per million output tokens. Pricing for Ultrafast itself has not been published as of this writing, and it's unclear how large a premium it will carry over the standard tier. Service tiers backed by dedicated hardware for guaranteed speed generally command a premium industry-wide, but that is a general pattern, not a guarantee of what Ultrafast will eventually cost. For background on GPT-5.6's pricing history, see our comparison of GPT-5.6 and Grok 4.5 API pricing.

Cerebras and Groq both use dedicated hardware to sidestep the memory-bandwidth bottleneck, but their implementations differ. If you're exploring faster inference for local deployments rather than hosted APIs, our local LLM inference engine comparison covers that separate landscape.

Ultrafast looks best suited today to larger customers who both have generation speed as a genuine bottleneck — in high-volume batch pipelines or multi-step agent workflows — and can secure access to the limited preview. Given that pricing remains undisclosed, that the quality claims haven't yet been independently verified, and that a general-availability timeline hasn't been announced, most small and midsize development teams would be reasonable to wait for a published price, wider availability, and independent benchmark reproductions before adopting it. A practical stance in the meantime is to keep using the standard GPT-5.6 tier for its cost-speed balance while tracking how Ultrafast develops.