OpenAI used DevDay to productize something builders have been paying for indirectly all year: wall-clock latency. On September 29, 2026, @OpenAI posted: “This is Ultrafast. Our premium speed tier, Ultrafast offers up to 8x faster token generation (300 tokens per second) in Codex and up to 6x in the API.” A follow-up said the tier is available today for GPT-6 Astra in Codex, ChatGPT Work, and the API, with GPT-6.1 Sol coming soon. Codex and ChatGPT Work access is tied to a new top seat, Pro 500 — highest usage limits (25x Plus) plus Ultrafast — listed on chatgpt.com/pricing. OpenAI also said it is reopening Pro 200 subscriptions with frontier models including Astra and GPT-6.1 Sol.
Sam Altman, in the same news cycle: “Ultrafast is so fast I do not ever want to go back.” The line is a preference, not a benchmark. The number that matters for a coding loop is the official Codex claim — 300 tokens per second, up to 8x — and the API claim — up to 6x. This is explainx.ai's builder read of the speed SKU, the subscription gate, and the reported API ladder you still need to confirm yourself.
The official Ultrafast video is on that post: OpenAI's Ultrafast clip on X. Watch the side-by-side before you treat 8x as a feeling rather than a marketing ceiling.
TL;DR — what actually shipped?
| Question | Direct answer |
|---|---|
| What is Ultrafast? | A premium speed tier, not a new model. Same Astra (today) or Sol (soon), faster decode. |
| Official Codex speed? | Up to 8x, 300 tokens per second, per OpenAI's September 29 post. |
| Official API speed? | Up to 6x faster token generation. OpenAI did not publish a tok/s number for the API in that tweet. |
| Which model today? | GPT-6 Astra in Codex, ChatGPT Work, and the API. |
| Which model later? | GPT-6.1 Sol, “coming soon.” |
| How do I get it in Codex / Work? | Pro 500 (highest usage, 25x Plus, includes Ultrafast). Check chatgpt.com/pricing. |
| Is Pro 200 dead? | No. OpenAI said it is reopening Pro 200 with frontier models including Astra and GPT-6.1 Sol. |
| Is this the August 750 tok/s preview? | No. That was GPT-5.6 Sol on Cerebras at 750 tok/s, no public price. DevDay's Codex figure is 300 tok/s on Astra. |
| API dollars? | Reported, not scraped by us. Stats Wire quoted a 4-rung Astra ladder ending at $60 / $300 for Ultrafast. Verify on openai.com/api/pricing. |
| Where does this sit in DevDay? | One SKU inside a 20+ announcement day. Hub: OpenAI DevDay 2026. Agents: Dots. |
What is Ultrafast, and what is it not?
Ultrafast is a speed tier: you pay (or subscribe) for faster token generation on a model you already know. It is not a smarter Astra, not a new weights drop, and not a replacement for choosing Sol versus Astra on quality or list price.
That distinction matters in agent loops. A 40-step Codex session multiplies every slow token. Buying 8x decode can shrink wall-clock more than swapping models — and it can also blow a budget if the API rung is several times Standard. Speed and intelligence are now separately billed axes. The DevDay recap already flagged Ultrafast as the latency SKU; this post is the deeper billing and comparison note.
What Ultrafast is not:
- Not the August Cerebras 750 tok/s product restated. August branded a GPT-5.6 Sol preview at 750 tokens per second for select API customers, with no ChatGPT or Codex access and no published price. Today's official Codex number is 300 tok/s (8x) on GPT-6 Astra. If you mash 750 and 300 into one “Ultrafast is 750” sentence, you are mixing generations and models.
- Not automatically on Plus or a leftover Pro 200 seat. OpenAI's follow-up tied Codex and Work Ultrafast to Pro 500.
- Not a guarantee you will measure 300 tok/s on your repo. 300 is OpenAI's up to Codex figure. Prefill, tool calls, and network still sit outside decode TPS.
How fast is 300 tokens per second in a real Codex session?
Three hundred tokens per second is a decode rate. After the first token, a 1,500-token function body is about five seconds of generation if you actually hit 300 tok/s and nothing else blocks. The same body at 37.5 tok/s (300 ÷ 8) is forty seconds. That is the 8x story in one arithmetic pass — and it is why Altman's “I do not ever want to go back” line resonates with people who sit in the loop.
The API claim is up to 6x, not 300 tok/s. Do not paste the Codex number onto API invoices. If you need a tok/s figure for an HTTP client, measure it on your key and region; OpenAI's tweet did not publish an API tok/s.
Interactive coding, voice, and tight agent cycles are the obvious buyers. Batch evals and overnight refactors are not. If the job can wait, Batch or Flex is the cheaper shape — even before you look at a reported Ultrafast multiplier.
Why did people mix 750 tok/s and 300 tok/s?
Because OpenAI reused the Ultrafast name.
In August, explainx.ai covered GPT-5.6 Sol Ultrafast on Cerebras: up to 750 tokens per second, 14x that model's normal speed, API-only preview, pricing undisclosed. That post is still the right source for the 5.6 Sol / Cerebras preview. It is the wrong source for DevDay Astra Codex 300 tok/s.
| Generation | Model | Published speed | Surfaces then | Price then |
|---|---|---|---|---|
| August 13, 2026 preview | GPT-5.6 Sol | Up to 750 tok/s (Cerebras) | Select API customers | Not disclosed |
| September 29, 2026 DevDay | GPT-6 Astra | 300 tok/s in Codex (8x); API up to 6x | Codex, ChatGPT Work, API | Subscription gate + reported API ladder |
If someone says “Ultrafast is 750,” ask which model. If they say “Ultrafast is 300,” they are quoting today's Codex Astra post. Both can be true in their own weeks. They are not interchangeable.
What is Pro 500, and did the $500 rumor just win?
OpenAI's follow-up: to use Ultrafast in Codex and ChatGPT Work, they introduced Pro 500 — highest usage limits (25x Plus) and Ultrafast. That is the official product name. It is not “Pro Max” on the pricing page unless OpenAI later aliases it.
explainx.ai tracked the leak as ChatGPT Pro Max at $500. The rumor had the price band and the fastest Work/Codex angle. DevDay confirmed a $500-class top seat under the name Pro 500. Treat “Pro Max” as the leak label; treat Pro 500 as the SKU to search on chatgpt.com/pricing.
25x Plus is a usage multiplier, not a speed multiplier. Ultrafast is the speed add-on. You can imagine a high-limit seat that is still slow, or a fast seat that still caps out — OpenAI bundled both on Pro 500 for Codex and Work.
OpenAI also said it is reopening Pro 200 with frontier models including Astra and GPT-6.1 Sol. That matches the Pro 200 reopen / half API-dollar thread from the same week: $200 is back as a frontier seat, not as the Ultrafast Codex/Work gate. If you only need Astra/Sol on a $200 login and you do not need 300 tok/s in the IDE, Pro 200 may be the plan you actually buy. If Codex latency is the bottleneck and you are already hitting Plus-class caps, Pro 500 is the plan aimed at you.
Enterprise was mentioned in the broader DevDay recap as another Ultrafast path. Confirm with your admin and the live pricing page; this post is consumer-SKU first.
Reported Astra API ladder — verify on the pricing page
explainx.ai did not independently scrape openai.com/api/pricing. An X account, Stats Wire, quoted this Astra Ultrafast input/output ladder (dollars per million tokens, as reported):
| Speed rung (reported) | Input / M | Output / M | vs reported Standard |
|---|---|---|---|
| Batch / Flex | $5 | $25 | Cheaper, slower / deferred |
| Standard | $10 | $50 | Baseline |
| Fast | $20 | $100 | 2x Standard |
| Ultrafast | $60 | $300 | 6x Standard |
Same model. If those numbers are live, max speed is about 6x Standard dollars — which lines up with OpenAI's “up to 6x” API speed line more cleanly than with the Codex 8x / 300 tok/s line. Community replies called that multiplier Denial-of-Wallet: a tight loop can spend like a different product without changing the model name in your logs.
Until you open the official page, treat the table as reported. Screenshots rot. Region and SKU names move. If your finance team needs a number for a forecast, copy it from OpenAI, not from this post.
Worked example if the reported Ultrafast rates hold: 10 million input and 10 million output tokens on Ultrafast is $600 + $3,000 = $3,600. The same tokens on reported Standard is $100 + $500 = $600. That is a $3,000 delta for speed on one chunk of traffic. Route only the interactive slice to Ultrafast; keep evals and rewrites on Batch/Flex or Standard.
August's Cerebras preview had no price. Do not back-solve 750 tok/s into today's $60/$300 rumor. Different model, different week, different commercial story.
What people are asking
Do I need Pro 500 if I only call the API?
Not for the HTTP path. OpenAI said Ultrafast is in the API today for Astra. The Pro 500 sentence was about Codex and ChatGPT Work. API users still need a key, a rate limit, and a confirmed price on the API page. A $500 ChatGPT seat is the wrong instrument if you never open Codex.
Does Plus or Team get Ultrafast in the IDE?
OpenAI's posted access path for Codex and Work is Pro 500. Do not assume Team or Plus inherited it. Recheck chatgpt.com/pricing after you read this — SKUs change in a day.
When does GPT-6.1 Sol get Ultrafast?
“Coming soon” is all OpenAI committed to in the follow-up. Sol's job on DevDay was price-performance, not 300 tok/s. If your default model is Sol, plan as if Ultrafast is Astra-only until the Sol toggle appears.
Is 300 tok/s faster than Gemini Flash or Cerebras open-weight demos?
Different models, different hardware stories. The August Sol preview's 750 tok/s is a higher published decode number on a previous OpenAI model via Cerebras. DevDay's 300 tok/s is the official Codex Astra number. Comparing 300 to a third-party 1,800 tok/s open-weight demo is usually category error. Compare your task's wall-clock and your invoice.
Will Dots eat Ultrafast quota?
Dots are always-on Astra agents with a cloud computer. Background work can be token-heavy without feeling “interactive.” If Dots and Ultrafast Codex share a usage bucket on Pro 500, a Dot farm can starve the IDE. OpenAI has not published a split we can cite here. Watch the usage page for a week before you staff five Dots and a 300 tok/s Codex habit on one seat.
Should I rewrite my harness to stream Ultrafast?
Only if human-in-the-loop latency is the bottleneck. Tool-call chatter, sandbox boots, and test runs often dominate. Profile a real session before you pay 6x on every completion. Prefix-stable prompts still matter for cache; Ultrafast does not erase a cache-miss tax.
What this changes for what you build or pay
- Split interactive vs batch in the router. Interactive Codex and user-facing agents are the Ultrafast candidates. Nightly jobs are not.
- Do not use 750 tok/s in a 2026-09 Astra forecast. That number is the 5.6 Sol Cerebras preview. Use 300 tok/s (Codex, up to) and 6x (API, up to) unless you are still on that preview SKU.
- Confirm dollars on OpenAI's pages. Subscription: chatgpt.com/pricing. API: openai.com/api/pricing. The Stats Wire ladder is a lead, not a contract.
- Pick Pro 200 vs Pro 500 on two axes. Model access and usage (Pro 200 reopen) versus Codex/Work Ultrafast + 25x Plus (Pro 500). The half API-dollar $200 math is a separate argument from speed.
- Keep a second lab in the eval set. GPT-6.1 Sol at $2/$10 is the cheap quality rung. Ultrafast Astra is the expensive latency rung. They can live in one product as two routes.
If you train teams on this stack, explainx.ai's workshops and agent-skills material is built for the same decision: when to spend on the model versus the loop.
Related reading
- OpenAI DevDay 2026 — full announcement map
- GPT-6.1 Sol launch, pricing, and benchmarks
- Dots: always-on ChatGPT agents
- GPT-6 Astra launch benchmarks and pricing
- GPT-5.6 Sol Ultrafast on Cerebras (750 tok/s, no price)
- ChatGPT Pro Max $500 rumor — now Pro 500
- ChatGPT Pro $200 reopen and usage math
- Official: ChatGPT pricing · OpenAI API pricing · @OpenAI Ultrafast post
Speed claims, Pro 500 / Pro 200 availability, and API rates are as of September 30, 2026, sourced to OpenAI's September 29 posts and to a reported third-party quote of the Astra ladder. explainx.ai did not scrape the live pricing HTML. Confirm Codex, Work, and API SKUs on OpenAI's pages before you change production routing or a company card.