OpenAI Previews Ultrafast: GPT-5.6 Sol at Up to 750 Tokens Per Second
On 13 August 2026, OpenAI shared an early look at Ultrafast — a new service tier in the OpenAI API that runs GPT-5.6 Sol at speeds of up to 750 output tokens per second, powered by Cerebras Systems and its Wafer-Scale Engine silicon architecture. The announcement arrived alongside a press release from Cerebras and positions Ultrafast as a limited preview available to a select group of customers, with expansion planned as Cerebras infrastructure capacity grows. No pricing has been announced as of the preview date. OpenAI describes the tier as delivering the same intelligence as GPT-5.6 Sol Standard — not a distilled or smaller variant — at Cerebras' substantially faster throughput.
The Speed Gap: 750 vs 53 Tokens Per Second
Understanding the significance of 750 tokens per second requires knowing the baseline: GPT-5.6 Sol on Standard processing generates approximately 53 tokens per second on OpenAI's existing infrastructure. Ultrafast is up to 14 times faster — a throughput difference that crosses a qualitative threshold for several classes of application. At 53 tokens per second, generating a 4,000-token response takes approximately 75 seconds. At 750 tokens per second, the same response completes in roughly 5 seconds. For interactive applications where a human is waiting, the difference between 5 and 75 seconds changes the product experience category: from a query that requires a context switch to one that feels close to real-time.
The effect compounds in agentic architectures. An agentic coding pipeline that makes ten sequential model calls — generating code, running tests, analysing errors, proposing fixes, and iterating — accumulates 750 seconds of inference wait at Standard throughput and approximately 50 seconds at Ultrafast throughput. For teams building multi-step agentic products, the 14x speed improvement is not merely a latency reduction: it changes which pipeline designs are viable within acceptable user-facing response budgets.
Cerebras Wafer-Scale Engine: The Hardware Behind Ultrafast
Cerebras Systems builds the Wafer-Scale Engine, an AI inference accelerator that differs fundamentally from conventional GPU-based hardware. Where standard AI accelerators are fabricated as individual chips joined by high-speed interconnects, a Wafer-Scale Engine spans an entire semiconductor wafer as a single chip — the largest chip in the world by transistor count. This architecture eliminates the chip-to-chip communication overhead that limits inference throughput on multi-GPU clusters, giving the WSE a specific advantage in single-stream, latency-sensitive inference: exactly the scenario that matters most for interactive AI products and sequential agentic pipelines.
The $10 Billion Partnership
The Ultrafast preview is the first end-user-facing output of a multiyear infrastructure partnership OpenAI and Cerebras announced in January 2026. The agreement covers deployment of up to 750 megawatts of Cerebras inference systems in stages through 2028. Reuters reported at the time that the deal was worth more than 10 billion US dollars, making it one of the largest AI infrastructure agreements on record. The Ultrafast limited preview represents the first stage of that Cerebras infrastructure reaching production customers, with capacity and access expanding as additional systems come online through 2028.
Intelligence Parity With Standard
A critical detail in the Ultrafast announcement: the model running on the Ultrafast tier is the same GPT-5.6 Sol as Standard mode, not a smaller or distilled variant optimised for speed at the expense of capability. Teams evaluating Ultrafast can expect identical benchmark performance, output quality, and model behaviour to Standard, with the difference being entirely in latency and throughput characteristics. This distinguishes Ultrafast from speed-optimised product tiers that achieve throughput gains by routing to a weaker model — a distinction that matters for applications where output quality is non-negotiable.
Use Cases Where 14x Speed Changes the Product
The throughput improvement does not affect all use cases equally. For single-turn question-answering or document summarisation where a human reads the full output before acting, the difference between a 5-second and 75-second response is noticeable but may not be architecturally significant. The use cases where the gap is most impactful are those where the model sits inside a loop: agentic coding assistants that call the model on each step of a multi-file refactoring task; automated test generation pipelines that process hundreds of functions per run; document intelligence workflows that chunk large files and call the model on each chunk sequentially; and real-time voice or streaming applications where token generation speed is the direct bottleneck on response latency.
What OpenAI Ultrafast Means for Indian Engineering Teams
For Indian software product companies building agentic features on top of the OpenAI API, the Ultrafast tier changes the architecture calculus for pipeline design. Agentic loops that chain model calls sequentially — a common pattern in coding agents, automated code review, and document processing workflows — can reduce end-to-end pipeline latency by close to an order of magnitude when the bottleneck is model inference speed. This is particularly relevant for Indian teams building B2B SaaS products where response time is a quality metric embedded in enterprise SLAs, or for teams building internal AI tools where long inference waits reduce adoption across non-technical users.
The absence of pricing information means production budget planning requires waiting for Cerebras capacity to expand and rates to be published. For teams currently building agentic products or planning a launch that depends on low-latency inference, the period between now and general availability is the right time to request preview access and benchmark real workloads: the latency improvement is large enough that it could justify re-architecting pipelines currently designed around Standard-mode constraints, and discovering that early in the design cycle is worth the evaluation effort.
The Bottom Line
On 13 August 2026, OpenAI previewed Ultrafast — a new API service tier powered by Cerebras' Wafer-Scale Engine running GPT-5.6 Sol at up to 750 output tokens per second, compared with approximately 53 tokens per second on Standard — a 14x throughput improvement with identical model intelligence. The tier is in limited preview with no pricing announced, expanding as Cerebras capacity grows under a multiyear agreement reportedly worth over 10 billion US dollars through 2028. For Indian engineering teams building latency-sensitive agentic pipelines, automated coding assistants, or AI-driven document workflows on the OpenAI API, GPT-5.6 Sol Ultrafast represents a qualitative change in what is achievable at production inference speeds and is worth evaluating as preview access expands.
Frequently Asked Questions
What is OpenAI Ultrafast and how fast is it?+
OpenAI Ultrafast is a new API service tier previewed on 13 August 2026 that runs GPT-5.6 Sol at up to 750 output tokens per second, powered by Cerebras Systems and its Wafer-Scale Engine silicon architecture. For comparison, GPT-5.6 Sol on Standard processing generates approximately 53 tokens per second, making Ultrafast up to 14 times faster. OpenAI states that the Ultrafast tier uses the same model as Standard — not a distilled or smaller variant — so output quality and intelligence are identical; only latency and throughput differ. The tier is in limited preview as of August 2026 with no pricing announced.
What is the Cerebras Wafer-Scale Engine and why does it enable faster inference?+
The Cerebras Wafer-Scale Engine (WSE) is an AI accelerator chip that differs from conventional GPU-based hardware. Where standard AI accelerators are fabricated as individual chips joined by high-speed interconnects, the WSE spans an entire semiconductor wafer as a single chip — the largest chip in the world by transistor count. This architecture eliminates the chip-to-chip communication overhead that limits inference throughput on multi-GPU clusters, giving the WSE a specific advantage in single-stream, latency-sensitive inference tasks. OpenAI and Cerebras announced a multiyear partnership in January 2026 covering deployment of up to 750 megawatts of Cerebras inference systems through 2028, reported by Reuters as worth more than 10 billion US dollars.
Which use cases benefit most from OpenAI Ultrafast?+
OpenAI Ultrafast's 14x speed improvement is most impactful in use cases where the model sits inside a loop or pipeline. Agentic coding assistants that make ten or more sequential model calls per task see cumulative latency reductions measured in minutes. Automated test generation pipelines processing hundreds of functions, document intelligence workflows that chunk large files and call the model on each chunk sequentially, and real-time voice or streaming applications where token generation speed is the direct bottleneck all benefit significantly. Single-turn question-answering or human-read document summarisation see a noticeable but less architecturally significant improvement.
When will OpenAI Ultrafast be generally available and what will it cost?+
As of the 13 August 2026 preview announcement, OpenAI has not published pricing or a general availability date for the Ultrafast tier. Access is in limited preview to a select group of customers, with OpenAI stating it will expand access as Cerebras infrastructure capacity grows under its multiyear agreement through 2028. Teams interested in evaluating Ultrafast should monitor the OpenAI API changelog and request preview access directly. Given that 750 megawatts of Cerebras inference systems are being deployed in stages, broader access and pricing transparency are expected as additional capacity comes online.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.
