GPT-5.6 Sol Ultrafast: 14× Speed, 750 Tok/s — CS-4 Reportedly ~1,300, No Price Yet

GPT-5.6 Sol Ultrafast mode is the hardware-accelerated serving tier for the flagship — the same full GPT-5.6 Sol model run on Cerebras wafer-scale chips instead of GPU clusters, delivering up to 14× the speed of standard processing and up to 750 output tokens per second on the tier OpenAI announced in August. On 2026-08-18, Cerebras unveiled the next-generation CS-4 system that this tier grows onto, and an early, unverified report claims GPT-5.6 Sol on CS-4 already serves roughly 1,300 tokens per second. It is not a smaller model and not a distillation: only the hardware and scheduling changed, and the intelligence is preserved. What it does not yet have is a price — as of 2026-08-19, Ultrafast is a limited API preview with no published rate card, so the number you can actually budget against today is the standard tier on the live GPT-5.6 Sol model page: $5 per million input tokens and $30 per million output, passed through at zero markup. This article covers what the mode is, how fast the numbers really are — including the new CS-4 hardware and what is still unverified — what it costs (including the honest "no price yet" answer), and what to do until it reaches general availability.

What Ultrafast mode actually is

Ultrafast is a serving tier, not a new model. OpenAI announced it on August 13, 2026 as the first product of its January 2026 compute partnership with chipmaker Cerebras — a deal reported at roughly $10 billion over three years. Where standard GPT-5.6 Sol inference runs on GPU clusters and spends much of its time moving weights between memory and compute, Ultrafast loads the model's weights into 44 GB of on-chip SRAM on Cerebras's wafer-scale chips; a model Sol's size spans several wafers in layers, so the weights sit where the compute is. The result is a throughput ceiling set by silicon rather than by memory bandwidth.

The hardware story moved on 2026-08-18. At its Supernova event, Cerebras unveiled the CS-4, the next-generation rack-scale system built on its Nexus platform — three Wafer-Scale Engine chips per rack, up to 2× the token-generation speed of the prior CS-3, and, per Cerebras, more than 1,000 tokens per second on models above 10 trillion parameters. Those are Cerebras's claims for a system in early access, not yet generally available. Since the unveiling, a single post on X claims CS-4 already serves GPT-5.6 Sol at roughly 1,300 tokens per second — consistent in direction with Cerebras's own numbers, but confirmed by nobody so far. Treat it as an early signal: plausible, unverified, and worth re-checking before you plan capacity around it.

The part that matters for your code: the model id, the weights, the reasoning behavior, and the output are identical to standard GPT-5.6 Sol. OpenAI frames Ultrafast as "more useful work per second" rather than a quality tier, and the preview runs the full flagship — speed does not come from swapping in a cheaper model. What changes is how fast tokens come out.

Don't confuse Ultrafast with the existing fast mode. Fast mode is a separate, purchasable tier on standard hardware: up to roughly 2.5× output speed at a 2× per-token premium ($10 input / $60 output per 1M), with no quality change. Ultrafast is a different thing — Cerebras hardware, up to 14×, and for now no price at all.

How fast — the numbers that matter

The headline figure is 14×, and it needs a definition. OpenAI says Ultrafast generates up to 750 output tokens per second versus standard GPT-5.6 Sol processing. That is a throughput number for output tokens, not a flat "everything runs 14× faster" claim — end-to-end request time also includes input processing and the model's own reasoning, which is why 14× is labeled a maximum.

• Max output throughput — up to 750 output tokens per second, per OpenAI's announcement (2026-08-13).

• Vs standard processing — up to 14× faster, per the same announcement; actual gains vary with input length and task type.

• Humanity's Last Exam, 2,500 questions — OpenAI/Cerebras report GPT-5.6 Sol Ultrafast finishing in 11 hours 11 minutes versus 78 hours 27 minutes for Claude Fable 5, with comparable accuracy. Vendor-reported, and the comparison spans two vendors' hardware stacks — read it as directional, not a verdict.

• GDP-Val, end-to-end — Cerebras reports 5.6× faster overall with no measurable quality loss. Also vendor-reported.

• Fast mode baseline — the existing purchasable speed tier tops out at roughly 2.5× output speed for a 2× price premium.

• CS-4, the next hardware — Cerebras claims the CS-4 rack delivers more than 1,000 tokens per second on models above 10 trillion parameters, up to 2× CS-3's token generation. An unverified community report puts GPT-5.6 Sol on CS-4 at ~1,300 tokens per second — same direction, not yet confirmed by OpenAI or Cerebras.

One caution before you quote the 14×: independent coverage of the launch cites an ~11× generation-speed comparison against Claude Fable 5, while Cerebras's own figures imply ~7× for total test time on the same run — the difference is which portion of a workload you measure. The same discipline applies to the CS-4 speed signal: the ~1,300-token/s figure started as a single X post, and neither OpenAI nor Cerebras has confirmed it. All of these are vendor or community numbers announced this week, and none are independently verified yet. Treat them as claims, not benchmark verdicts.

What it costs — and the honest gap

Here is the state of the price question on 2026-08-19, stated plainly: Ultrafast has no published price. OpenAI has not announced one, and the preview does not appear on any public rate card. Reports that early-access slots are bundled or free are speculation, not policy — and until OpenAI prices the tier, any "GPT-5.6 Sol Ultrafast cost" number you find elsewhere is a guess.

What you can price today is the rest of the GPT-5.6 Sol line, verified against OpenAI's published rates on 2026-08-18:

• Standard GPT-5.6 Sol — $5.00 input / $30.00 output per 1M tokens. The baseline you can call right now.

• Fast mode — $10.00 / $60.00, a 2× premium for up to roughly 2.5× output speed. The closest thing to a speed tier you can actually buy.

• Prompt cache read — $0.50 per 1M (90% off fresh input); cache writes bill at $6.25 per 1M.

• Long-context tier — inputs past roughly 272K tokens bill at $10.00 / $45.00.

• Batch API — $2.50 / $15.00, a flat 50% off with a roughly 24-hour turnaround.

For context on the premium to expect: fast mode set the precedent at 2× the standard rate for 2.5× the speed. If Ultrafast follows a similar curve it will sit well above the standard $5 / $30 — and commentary around the preview expects exactly that, because Cerebras wafer capacity is scarce and expensive. But an expected premium is not a price. Plan against standard Sol or fast mode until a rate card exists.

How to use it — the preview reality

Access is the second honest gap. Ultrafast is available through the OpenAI API to a select group of customers on a waitlist basis, announced 2026-08-13, with OpenAI saying access will expand "as capacity grows." There is no general-availability date. The announced preview cohorts are coding, commerce, financial research, support, and other interactive workloads — teams whose agents make many calls per request and where response latency is the bottleneck. Jane Street, Rogo, and Podium are among the named early testers.

The usage pattern it is designed for: incident response (reading logs, traces, and diffs while an outage is live), financial research and fraud detection on moving data, real-time customer support and voice (where a half-second silence reads as broken), e-commerce checkout, and compressing overnight research batches into interactive sessions. If your workload is a single-turn chat, the 14× matters far less than it does for a 40-call agent loop.

What you can do today — the practical answer

While Ultrafast is in preview, the full GPT-5.6 Sol is live right now — and on OrcaRouter the standard tier is served at the provider rate with $0 per-token markup, as model ID openai/gpt-5.6-sol through the OpenAI-compatible endpoint at api.orcarouter.ai/v1. If you already have an OpenAI client, migration is a base-URL and model-id change; nothing else. The same key and endpoint carry 200+ models, so Sol sits next to GPT-5.6 Terra, GPT-5.6 Luna, Claude Opus 5, and the open-weight models you'd compare it against — and you can route by difficulty instead of re-integrating each one.

Two OrcaRouter specifics are worth naming on a speed page. BYOK: bring your own OpenAI key and OpenAI bills you directly at its own rates — OrcaRouter adds $0 per token and keeps your provider rate limits and credits intact. Guardrails: the PII Shield and content policy run before billing, so a blocked request returns a clean 400 and is never charged, and the Agent Firewall grades tool and MCP calls before they execute. When — and if — Ultrafast reaches general availability at a routable price, plugging it into the same endpoint is a model-id change; the setup you build today carries over.

The honest boundary — when Ultrafast isn't the answer

Four places the 14× story does not hold, so you don't architect or budget against a number that isn't settled:

14× is a maximum, not a guarantee. It is an output-throughput ceiling on Cerebras hardware; end-to-end latency still includes input processing, the model's reasoning time, and your own network. A reasoning-heavy prompt can spend most of its wall-clock thinking, where the headline speedup shrinks.

It has no price and no GA date. As of 2026-08-19 you cannot buy Ultrafast — only request preview access, and the CS-4 hardware behind it is in early access rather than generally available. Budget against standard Sol ($5 / $30) or fast mode ($10 / $60), and treat any Ultrafast price you see online as unverified.

The benchmarks are vendor-reported. The 11-hour HLE run and the 5.6× GDP-Val figure are OpenAI/Cerebras numbers from the announcement; independent estimates range from roughly 7× to 11× depending on what's measured. Add the CS-4 speed signal to that list: the ~1,300-token/s figure for GPT-5.6 Sol on CS-4 comes from a single X post and has no independent confirmation. None of these are independently verified yet.

It won't fix a bottleneck you don't have. If your agents are slow because of your own orchestration, tool latency, or input-heavy prompts, a faster output path moves the tail only slightly. The tier-matching decision — GPT-5.6 Sol at $5 / $30 versus GPT-5.6 Terra at $2 / $12 versus GPT-5.6 Luna at $0.20 / $1.20 — still moves your bill more than any speed tier does.

Bottom line. GPT-5.6 Sol Ultrafast is the fastest serving tier OpenAI has shipped for its flagship — the same weights on Cerebras hardware, up to 14× throughput and 750 output tokens per second on the announced tier. Cerebras's new CS-4 (unveiled 2026-08-18) reportedly pushes GPT-5.6 Sol toward ~1,300 tokens per second, but that figure is a single unverified X post. As of 2026-08-19, Ultrafast is a waitlist preview with no public price. If you are searching for a number to budget against, the honest answer is that there isn't one yet. Standard GPT-5.6 Sol at $5 / $30 is what you can call today, fast mode at $10 / $60 is the purchasable speed tier, and OrcaRouter passes both through at $0 markup on your own key. Build the routing now — and when Ultrafast gets a price and a date, it is one model-id change away.

Until Ultrafast gets a price, the full GPT-5.6 Sol is live on the GPT-5.6 Sol on OrcaRouter at $5 / $30 per million tokens, zero markup, ready for your own key.