# Speed Is Now a Paid Tier at OpenAI > Source: > Published: 2026-08-14 15:09:51+00:00 [AI](https://sourcefeed.dev/c/ai)Article # Speed Is Now a Paid Tier at OpenAI Cerebras wafer-scale silicon pushes GPT-5.6 Sol to 750 tokens per second, and changes how inference gets priced. [Mariana Souza](https://sourcefeed.dev/u/mariana_souza) For two years, [Cerebras](https://www.cerebras.ai) has been the company that serves open-weight models absurdly fast while the frontier labs shrugged. That era ended this week. OpenAI and Cerebras are previewing an "Ultrafast" mode for GPT-5.6 Sol that pushes output to 750 tokens per second — roughly 14x the standard tier's ~53 tokens per second — with both companies insisting there's no quality haircut. It's the first time a closed frontier model has shipped on wafer-scale silicon as a product you can (eventually) buy, and it confirms something the pricing pages have been hinting at all year: inference speed is now a product tier, sold like cloud storage classes. ## The January deal, bearing fruit This didn't come out of nowhere. In January, [OpenAI](https://openai.com) signed a deal with Cerebras reported at more than $10 billion — around 750 megawatts of wafer-scale compute through 2028 — as part of its sprawl away from Nvidia-only inference (AMD and Broadcom being the other prongs). At the time it read as a capacity hedge and a lifeline for Cerebras, which badly needed revenue diversification away from G42 ahead of its IPO. Ultrafast mode is the first consumer-visible product of that deal, and it recasts the arrangement: this isn't overflow capacity, it's a differentiated tier that GPU racks can't easily replicate. The reason it can't be replicated is also the reason it'll stay scarce. Cerebras's Wafer-Scale Engine keeps model weights in on-chip SRAM — 44 GB per wafer — instead of streaming them from HBM every forward pass, which is the bandwidth wall that caps GPU inference speed. But 44 GB per wafer means a frontier-scale model has to be pipelined across many wafers, and wafers are not something you rack by the thousands on short notice. That's why Ultrafast is launching as a limited preview with a signup form rather than a checkbox in the API console. The physics that makes it fast makes it supply-constrained, and supply-constrained means premium-priced for the foreseeable future. ## Read the benchmarks as marketing, the arithmetic as real The launch numbers deserve sorting. The 750 tokens/second figure is a straightforward throughput claim and consistent with what Cerebras has demonstrated publicly on large open models, so I believe it. The competitive comparisons — 11x faster than Anthropic's Fable 5, 5x faster than Opus 4.8 in [fast mode](https://www.anthropic.com) — are vendor-run measurements of competitors' products; treat them as directional. And the headline stunt, running all 2,500 questions of Humanity's Last Exam serially in 11 hours versus a claimed 78+ hours for Fable 5, is a benchmark nobody runs that way in practice. But the stunt illustrates the thing that actually matters: latency compounds. A single chat completion at 53 tokens/second is tolerable. An agent that makes ten sequential model calls, each emitting a couple thousand reasoning-plus-output tokens, is not — that's six-plus minutes of wall clock at standard speed and under 30 seconds at 750 tokens/second. Reasoning models made this worse, because thinking tokens are output tokens; every chain-of-thought a model burns through is billed in seconds the user spends staring at a spinner. Speed at the token level is the difference between an agent you watch work and an agent you come back to later. Those are different product categories, not different performance grades. ## What you should actually do with this Almost nothing, yet — and that's not a knock. Access is preview-only, and Ultrafast pricing hasn't been published. The signal from OpenAI's existing tiers is that speed costs real money: per reporting from the-decoder, the Fast tier already runs roughly double standard rates for about 2.5x throughput, and Ultrafast sits above it. So the move for now is architectural, not procedural: **Split your traffic by latency sensitivity.** Batch classification, overnight evals, and document pipelines should stay on standard or batch pricing forever; paying a speed premium there is burning money. The candidates for a fast tier are interactive agent loops, voice, and anything where a human is blocked waiting.**Instrument wall-clock per task, not tokens per second.** If your agent's critical path is dominated by tool execution, retrieval, or network hops, a 14x model speedup buys you little. Measure where the seconds actually go before budgeting for the premium.**Watch feature parity at GA.** Neither company has said whether Ultrafast supports the full API surface — prompt caching, structured outputs, fine-grained tool use. Preview tiers historically launch narrow. Don't commit an architecture to it until the parity story is public. ## Who gets squeezed The obvious loser is [Groq](https://groq.com), whose entire pitch — SRAM-heavy silicon, blistering tokens per second — just got adopted by the biggest lab in the market, applied to a frontier model Groq will never host. Fast inference on open weights is still a real business, but "we're the fast ones" stops being a moat when speed becomes a first-party tier at OpenAI and Anthropic alike. The GPU-based speed specialists — the speculative-decoding shops — face the same compression from the other direction. The subtler consequence is for developers' cost models. We've spent two years treating price-per-token as the axis of model choice, with speed as an accident of load. Vendors have now discovered that latency is willingness-to-pay segmentation, and they're going to price it that way. Expect the gap between batch and interactive pricing to widen, and expect "which speed tier" to become a routing decision your orchestration layer makes per request, the way it already picks model sizes. My read: this is a genuine shift, not a demo. The deal size, the dedicated 750MW, and the fact that OpenAI is productizing it rather than press-releasing it all say Ultrafast is infrastructure, not stunt. But the SRAM math guarantees scarcity, and scarcity plus undisclosed pricing means the right posture today is preparation: get your latency-sensitive paths measured and your routing layer ready, and let someone else pay preview prices to find the sharp edges. ## Sources & further reading - [Accelerating GPT-5.6 Sol Ultrafast](https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai)— cerebras.ai - [Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed](https://openai.com/index/previewing-ultrafast/)— openai.com - [GPT-5.6 Sol goes 14x faster as OpenAI launches Ultrafast mode powered by Cerebras](https://the-decoder.com/gpt-5-6-sol-goes-14x-faster-as-openai-launches-ultrafast-mode-powered-by-cerebras/)— the-decoder.com - [Cerebras scores OpenAI deal worth over $10 billion ahead of AI chipmaker's IPO](https://www.cnbc.com/2026/01/14/cerebras-scores-openai-deal-worth-over-10-billion.html)— cnbc.com [Mariana Souza](https://sourcefeed.dev/u/mariana_souza)· Senior Editor Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon. ## Discussion 0 No comments yet Be the first to weigh in.