OpenAI Ultrafast: 300 tokens/s, pricing and supported models
OpenAI Ultrafast generates up to 300 tokens per second in Codex, 8x faster than standard mode. We break down the pricing, model support, and when it's worth it.
👁 120 reads
OpenAI Ultrafast is the new fastest and most expensive tier for OpenAI models, announced at DevDay on September 29, 2026. In Codex it generates up to 300 tokens per second, which OpenAI says is up to 8 times faster than standard mode, and up to 6 times faster in the API. It currently works with GPT-6 Astra, and in the API costs six times more than standard tier. We examined the pricing and documentation to figure out when this premium makes sense.
Quick summary
Ultrafast is a paid speed tier, not a new model. Today it works with GPT-6 Astra: in the API for everyone (with low limits), and in Codex and ChatGPT Work only on the Pro 500 plan and some Enterprise and Edu plans. In the API it costs $60 per million input tokens and $300 per million output tokens. In Codex it consumes your plan limit 8 times faster. GPT-6.1 Sol in Ultrafast version is expected in the coming days. The mode does not support EU data processing, which rules it out for some companies.
We cover the rest of the conference announcements in our DevDay 2026 recap.
What Ultrafast is
Beyond Batch mode, OpenAI now has four processing tiers in the API: cheaper and slower Flex (half price), Standard, Fast (called Priority until July 30, 2026, double price), and the new Ultrafast. They differ in generation speed and price, the model stays the same. OpenAI describes Ultrafast as a “premium tier for tasks where speed matters most.”
The mode debuted on August 13, 2026 as a limited preview for GPT-5.6 Sol. At that time OpenAI talked about speeds up to 750 tokens per second and a 14x speedup on Cerebras hardware. At DevDay the mode launched more broadly, but with GPT-6 Astra, a larger model, and at lower speed.
How fast it really is
Sources give different numbers, so we’re listing them side by side:
- DevDay summary: up to 8 times faster (300 tokens/s) in Codex and up to 6 times in API,
- API documentation: “up to 8 times faster than Standard tier”,
- Dan Shipper from Every, who saw demos before launch: 250 tokens/s.
From OpenAI’s numbers, standard Astra in Codex generates about 37.5 tokens per second (300 divided by 8). OpenAI notes it’s comparing only generation speed, not total task completion time. In the API they recommend using a WebSocket connection, because without a persistent connection network latency eats into the gains, especially with many tool calls.
We didn’t test Ultrafast mode ourselves: it’s not available on OpenRouter, and we didn’t use our direct OpenAI API account for this test. For comparison: in our test of GPT-6.1 Sol in standard mode through OpenRouter, it output about 58 tokens per second including network latency.
Which models and where
| Model | Ultrafast in API | Ultrafast in Codex and ChatGPT Work |
|---|---|---|
| GPT-6 Astra | yes, for everyone, low limits | Pro 500, selected Enterprise and Edu plans |
| GPT-6.1 Sol | announced | announced, up to 8x faster |
| GPT-5.6 Sol | limited preview for selected companies | no |
| GPT-6 Sol, GPT-6 Luna | no, only Fast | no, only Fast |
In Codex, other personal plans (Plus, Pro 100, Pro 200) don’t have access to Ultrafast even after purchasing credits. In Enterprise the mode is disabled by default and enabled by the workspace owner. We covered the plan details with Ultrafast in our article about ChatGPT Pro 500.
How much Ultrafast costs in the API
Rates from OpenAI API pricing as of September 29, 2026, per million tokens, for requests up to 272,000 input tokens.
| GPT-6 Astra | Input | Cached | Output |
|---|---|---|---|
| Flex | $5 | $0.50 | $25 |
| Standard | $10 | $1 | $50 |
| Fast | $20 | $2 | $100 |
| Ultrafast | $60 | $6 | $300 |
Above 272,000 input tokens, Ultrafast costs $120 for input and $450 for output. Default limits for Astra in Ultrafast are 500,000 tokens per minute at account tiers 1-3, one million at tier 4, and 5 million at tier 5. You enable the mode with the parameter service_tier="ultrafast".
Important note: Ultrafast only works with US or global processing. It does not support regional EU endpoints, and in ChatGPT it’s not available for workspaces that require processing outside the US.
How much it costs in Codex and ChatGPT Work
In subscriptions you don’t pay per token, you consume your limit. Codex documentation gives multipliers relative to standard mode:
| Mode | Plan limit consumption | Purchased credits consumption |
|---|---|---|
| Fast | 2.5x | 2x |
| Ultrafast (GPT-6 Astra) | 8x | 6x |
In Pro 500, Ultrafast first consumes your plan limit, then credits after it’s exhausted. With only Ultrafast, the 25x Plus limit is enough for roughly what three Plus limits would give you in standard mode.
When Ultrafast is worth it
We calculated an example agent task in the API: 50,000 new input tokens, 150,000 cached tokens, and 20,000 output tokens on GPT-6 Astra. This is our example, not a measurement.
- Standard: $0.50 + $0.15 + $1.00 = $1.65. At 37.5 tokens/s, generation alone takes about 9 minutes.
- Ultrafast: $9.90. At 300 tokens/s, generation takes about 67 seconds.
You save about 8 minutes for $8.25. If your hourly rate is $100, 8 minutes of waiting is worth $13.33, so on a single task the premium doesn’t pay for itself. It pays off where waiting blocks more people or money: production outages, live customer service, “change and check” loops with many short steps. In the August Ultrafast preview, OpenAI gave similar examples: incident response, transaction analysis, and customer service. This is our opinion.
If you care mainly about cost, GPT-6.1 Sol in standard mode costs $2 and $10 per million tokens, 30 times less than Astra in Ultrafast, and according to OpenAI approaches Astra performance on some tasks. When Sol gets its own Ultrafast, it might be the most sensible choice for fast agents. OpenAI hasn’t announced pricing for that version yet.
01OpenAI DevDay 2026: Dots, GPT-6.1 Sol, Pro 500 and MoreOver 20 announcements at OpenAI DevDay 2026: always-on Dots agents, GPT-6.1 Sol at one-fifth the cost of Astra, Ultrafast mode, ChatGPT Space and Pro 500.Read →
02GPT-6.1 Sol: Pricing, Availability & Non-English Writing TestGPT-6.1 Sol from DevDay 2026: API pricing in USD, differences from GPT-6 Sol and Astra, availability in ChatGPT plans, and our non-English writing test results.Read →
03OpenAI Dots: What They Are and Where They WorkOpenAI Dots are always-on AI agents in ChatGPT. Each Dot has its own cloud computer, connects to 4000+ apps, and works even when you're not chatting.Read →
04ChatGPT Space and Pages: What They Are and How to Get StartedChatGPT Space replaces Library with a team workspace for Pages, files, and presentations. Available in Pro, Business, and Enterprise plans.Read →Frequently asked questions
How many tokens per second does OpenAI Ultrafast deliver?
According to OpenAI, up to 300 tokens per second in Codex with GPT-6 Astra, up to 8 times faster than standard. In the API up to 6 times faster.
How much does Ultrafast cost in the API?
For GPT-6 Astra, $60 per million input tokens and $300 per million output tokens, six times more than standard tier.
Which models support Ultrafast?
Today GPT-6 Astra. GPT-6.1 Sol will get it in the coming days, and GPT-5.6 Sol is in limited preview for selected companies.
Is Ultrafast available on the Plus plan?
No. Among personal plans only Pro 500 has it. Purchasing credits on Plus, Pro 100, or Pro 200 doesn’t unlock it.
Sources and verification date
OpenAI
More on this from QUANTSLANT
ai tools·ChatGPT Space and Pages: What They Are and How to Get StartedChatGPT Space replaces Library with a team workspace for Pages, files, and presentations. Available in Pro, Business, and Enterprise plans.