OpenAI's Ultrafast Really Is 14x Faster, But the Price Is Nowhere to Be Found and ChatGPT Doesn't Get It Either (2026-08)

Yesterday (2026-08-13) OpenAI dropped a very easy-to-understand number: 750 tokens per second, said to be 14 times faster than the original standard speed.

It sounds like “tomorrow, open ChatGPT and the answers will come spraying out.”

But after reading that official page front to back, then digging through the official pricing page, I found three things the press release didn’t mention — and that you’ll probably run into too: the price can’t be found on the official site, spots require filling out a form to apply, and ChatGPT isn’t part of this rollout at all. This article covers all three, and at the end tells you what’s actually relevant to you today — a different piece of news entirely.

1. Plain language first: what exactly is getting faster?

Let’s define a term first. A token is the smallest unit AI uses when reading and writing text — think of it as “one bite of food” for the AI. The speed at which AI outputs text is measured as “how many bites per second.”

What OpenAI just launched is called Ultrafast, officially defined as “a new service tier that lets GPT-5.6 Sol run up to 14 times faster than standard processing,” with “output of up to 750 tokens per second.” (Official source: OpenAI - Previewing Ultrafast, verified 2026-08-14)

Why does it get faster? It’s not that the model got smarter — it’s that the hardware it runs on changed. This time the compute comes from chip company Cerebras, using their “Wafer-Scale Engine” — in plain terms, instead of cutting a wafer into many chips like everyone else, they turn the entire wafer into one giant chip, so data doesn’t have to travel back and forth between multiple chips, which lowers latency. (Cerebras official press release, SUNNYVALE 2026-08-13, link, verified 2026-08-14)

This site did the math ourselves: what’s the baseline for that 14x?

The official announcement only says “14 times faster” but doesn’t say how slow it was to begin with. This kind of “give the multiplier, skip the baseline” writing is exactly the kind of thing this site loves to dig into.

So I went and checked third-party benchmarks myself. Independent speed-testing outfit Artificial Analysis currently lists GPT-5.6 Sol’s (via the official OpenAI API) output speed as:

(Source: Artificial Analysis, verified 2026-08-14. Per disclosure item 9.7: this is a third-party measurement, not an official OpenAI figure — this site is merely relaying the data.)

Then working backward from the official number: 750 ÷ 14 ≈ 53.6 tokens/sec — which nearly overlaps with the third-party measured 52.8–55.

Conclusion: this “14x” isn’t marketing inflation — the baseline checks out. Credit given where due.

2. Three things the press releases didn’t tell you

① The official pricing page has no price listed for Ultrafast

I read through OpenAI’s official API pricing page, and here’s the current GPT-5.6 family and service tier breakdown:

(Source: OpenAI official API pricing page, price verified: 2026-08-14)

See the key point? OpenAI already has an acceleration tier called Fast mode, priced at 2x. So what will Ultrafast — 14 times faster — cost?

The official announcement, Cerebras’s press release, and every report so far — none of them say. This site doesn’t estimate or extrapolate (per rule 1.2, no fabrication). We’ll update once it’s officially announced.

② You have to fill out a form to use it, and OpenAI picks its own customers

The official announcement states it’s being offered today “as a limited preview to a small group of customers,” to be expanded as capacity grows.

I clicked through to the official application form, and it states in black and white: “Capacity is limited. We’ll evaluate customers for inclusion based on workload fit and availability.” The form asks for name, work email, company name, and use case. (OpenAI Ultrafast interest form, verified 2026-08-14)

In plain terms: this isn’t first-come-first-served — they decide whether to grant access based on what you want to use it for.

③ Your ChatGPT will not get faster because of this news

The official announcement clearly states it’s “launching first on the OpenAI API.” The entire announcement makes no mention of ChatGPT, nor of any subscription plans.

In other words: whatever you’re paying for ChatGPT Plus today has nothing to do with Ultrafast. This is the misunderstanding this article most wants to save you from.

3. So what actually matters to you today?

If you’re an average user who doesn’t write code, what you should actually be paying attention to this week is the announcement from 8 days ago (2026-08-06):

OpenAI officially announced that GPT-5.6 Luna is now the default model for the Free and Go plans, free users get “unlimited text conversations,” and a new “Think” button (which lets the AI spend more time on hard problems). (Official source: OpenAI - Improving GPT-5.6 Sol in ChatGPT; media corroboration: TechCrunch 2026-08-06, MacRumors 2026-08-06. Verified 2026-08-14)

But “unlimited” has boundaries — the official announcement itself says so:

• Only plain text chat is unlimited
• File uploads, images, and other tool usage limits remain unchanged
• The official note adds it’s “subject to abuse guardrails”

Timeline: Paid users got the update on the announcement day; Free/Go users switch to Luna as default “this week”; the Think button starts “next week.” Counting from the 8/6 announcement date, the Think button lands right in the week of 8/10 — which is now. ⚠️ Per rule 9.4: this site cannot individually verify the actual rollout timing for each account — if you haven’t seen it yet, it just hasn’t reached you yet, it doesn’t mean OpenAI hasn’t done it.

For reference, here’s the Taiwan pricing currently in this site’s database, to help you decide whether to upgrade:

ChatGPT Taiwan official site monthly listed prices: Free NT$0 / Go NT$270 / Plus NT$690 / Pro from NT$3,300 Price verified: 2026-08-09 (this site accessed the official Traditional Chinese site directly via a Taiwan connection; record’s last_verified: 2026-08-14) Per rule 10.1 confirmation: the official FAQ explicitly lists Go, Plus, and Business as offering monthly billing, the above are monthly listed prices, not annual-billing monthly averages.

4. Honest disclosure on cross-verification (per rules 9.1/9.7)

• For the “750 tokens/sec” and “14x” figures, I found three sources: OpenAI’s official announcement, Cerebras’s official press release, and TechCrunch’s report.
• But to be honest: OpenAI and Cerebras are partners on this deal, so they don’t count as two “independent” sources; TechCrunch is also just relaying the same official figures. So at this stage, this speed figure has no independent third-party benchmark corroboration yet — this site can only label it as “official claim.”
• Access method disclosure: openai.com returns a 403 to automated scraping, so this site accessed the official pages via the r.jina.ai proxy to retrieve the content. The source content is still the official page, not a secondhand paraphrase, but the access path was not direct, noted here for transparency.
• Artificial Analysis’s speed figures are a third-party measurement, not officially published, and this site is merely relaying the data — the official announcement takes precedence.

5. Cat banter

Amo (spots the trap): “Fast mode already charges 2x. For something 14 times faster, do you really think they’ll dare put that on the pricing page? A speed service with no listed price usually isn’t because it’s cheap.”

Pimi (finds the upside): “But we did the baseline math ourselves — 53.6 lines up with the third-party measured 52.8–55, so there’s no inflation this time! The multiplier is real.”

Amo: “The multiplier’s real, the price is missing — that’s ‘get you wanting it first, ask about price later.’ And it’s not even on ChatGPT — regular people get zero benefit from this today.”

Pimi: “Okay fine, but unlimited chat on the free tier is still a good thing, right?”

Amo: “Only text is unlimited. Images, files, tools — all still locked. That’s an ‘unlimited’ with more fine print than headline.”

Quoted prices: GPT-5.6 Sol input US$5 / output US$30 per 1M tokens; Fast mode is 2x standard price. Price verified: 2026-08-14

6. The following is Ask Cat editorial opinion, not news fact

Our take: the Ultrafast news has zero short-term impact on 99% of readers in Taiwan, but the direction it signals is worth remembering.

Three reasons:

First, this is a war over “latency,” not “intelligence.” OpenAI itself says it plainly in the announcement — until now, getting real-time speed meant switching to a smaller model. What Ultrafast is trying to solve is “smart AND fast.” This signals that going forward, the competition among providers will gradually shift from “who scores higher on tests” to “who responds faster.” Because for things like real-time voice, customer service, and order placement, being half a second slower can mean losing the sale.

Second, not announcing a price is usually not good news. OpenAI’s existing acceleration tier, Fast mode, charges 2x — that’s a fact stated right on the official pricing page. Ultrafast is far faster, yet doesn’t even give a price range, and requires “evaluating workload fit” before granting access — this looks more like enterprise-level pricing negotiated per customer, not something you self-serve with a credit card. So if you’re a small team or independent developer in Taiwan, our advice is: there’s no need to adjust any plans right now — filling out the form probably won’t get you in anyway; wait until it’s on the pricing page.

Third, conversely, cheap models like Luna are the real battleground for people in Taiwan. On the official pricing page, Luna’s output price is 1/25th of Sol’s (US$1.20 vs US$30, verified 2026-08-14), and OpenAI chose it as the default model for the free tier — essentially betting that “for most people’s most questions, a cheap model is enough.” For the average reader, that matters a lot more than being 14 times faster.

One thing you can do right now

Open ChatGPT and look for a “Think” button near the input box.

If it’s there: congrats, your account has already gotten the update — click it on a hard question and let Luna think a bit longer before answering, for free. If it’s not there: no need to reinstall or unsubscribe, the official rollout is staged, just wait.

As for Ultrafast — hold off on your wallet, because there’s currently no price to spend it on.

Sources

Full-article price verification date: 2026-08-14. Prices and plans are subject to change at any time — the respective official websites’ announcements take precedence.