OpenAI's fastest tier reached Europe on 8 October — on every model except the flagship

On 8 October 2026 OpenAI turned on its Ultrafast service tier for GPT-6.1 Sol across the API, Codex and ChatGPT Work. The pitch, from OpenAI’s own developer account: “near-Astra intelligence at up to 8x faster speeds,” priced at $12 per million input tokens and $60 per million output — six times Sol’s standard $2/$10, and, as OpenAI points out, just 1.2x the cost of standard Astra.

That is a good speed story and it is not the story.

The story is one sentence further down the announcement: Ultrafast is “available in all supported regions, including support for data residency in the US and EU.” Alongside it, OpenAI added EU data residency to GPT-6.1 Sol Fast and GPT-6 Luna Fast.

Set that next to what OpenAI’s documentation still says about the flagship. GPT-6 Astra’s Ultrafast supports US data residency and global processing only. It does not support EU or other non-US regional processing endpoints. Astra’s Fast tier has never been available under EU residency either — a gap this desk documented on 6 September, three days after Astra shipped.

So the sentence a European buyer should read out of 8 October is not “Ultrafast is 8x faster.” It is: OpenAI’s speed tiers are now purchasable in the EU on every model except the one it sells as its best.

The thesis: in Europe, latency is now a model choice, not a price choice

Everywhere else in OpenAI’s lineup, speed is something you buy with money. Pass service_tier: "fast" and pay 2x. Pass service_tier: "ultrafast" and pay 6x. The model is unchanged; you are buying a place in the queue and the silicon underneath it.

Under EU data residency that stops being true. There, the escalation ladder has exactly one model on it:

Standard rates per million tokens from OpenAI’s pricing documentation, for requests under 272,000 input tokens; above that threshold input doubles and output bills at 1.5x for the whole request. The two Ultrafast rates and Astra’s Fast rate are published figures; Sol and Luna Fast are the documented 2x multiplier applied to their standard rates, so confirm them in the dashboard before budgeting. EU regional processing endpoints carry a 10% uplift on models released on or after 5 March 2026, so an EU-resident Sol Ultrafast request actually meters at about $13.20/$66.

Read down the right-hand column and the decision writes itself. A European team that needs sub-second interactive generation cannot have it on Astra. It can have it on Sol. The constraint is not budget — a US team with the same budget has both — it is jurisdiction. And because Sol scores one point below Astra on Artificial Analysis’ index while costing a fifth as much, the jurisdictional constraint happens to push European buyers toward the model the arithmetic already favoured.

That is an unusual shape. Regional feature gaps normally cost you something. This one costs a point of benchmark score and saves 80% of the rate.

What the 6x actually buys, and what it costs in allowance

Ultrafast’s multiplier is uniform — 6x on input, 6x on output, on both models — which makes it one of the few prices in this market you can reason about without a spreadsheet. OpenAI describes up to 8x faster generation than Sol Standard, and roughly 300 output tokens per second in Codex.

The throughput ceilings differ sharply between the two models, and in Sol’s favour. On the Build usage tier, Sol Ultrafast runs to 1 million tokens per minute against Astra Ultrafast’s 500,000; on Grow, 40 million against 5 million. Sol is not merely cheaper to run fast, it is allowed to run fast at eight times the volume.

The cost that does not appear on the rate card is allowance. In Codex and ChatGPT Work, Ultrafast is gated to Pro 500 ($500/month), eligible usage-based Enterprise plans and credit-based Edu plans, with Enterprise administrators enabling it per user and the tier off by default in Enterprise workspaces. Reporting puts the drain at 8x against included Codex limits and 6x against purchased credits.

Put that next to the subscription arithmetic from DevDay: Pro 500 sells 25x the Plus allowance for $500, and every rung of OpenAI’s ladder now costs the same $20 per 1x of Plus usage. Run a Pro 500 seat entirely in Ultrafast and its effective headroom falls to roughly 3x of Plus — less than a Pro 100 seat at a fifth the price. The 25x figure and the 8x figure are drawn from the same budget in opposite directions, and nothing in the product tells you the exchange rate.

The honest framing is the one this desk reached in August, when Ultrafast arrived as a Cerebras-powered preview with no price, no SLA and no region list: the speed tier is a latency surcharge the subscription absorbs on your behalf. Two of those three unknowns are now resolved. The price is 6x. The region list exists and excludes the EU on Astra. The SLA is still missing — OpenAI has published no latency commitment for Ultrafast on any model, which means the 8x is a performance claim, not a contractual one.

Where this leaves the flagship

OpenAI shipped Astra on 3 September at $10/$50, exactly 2.5x GPT-5.6 Sol on every line item, with the argument that you should price the task rather than the token. Three things have happened since, all in the same direction.

On 22 September, GPT-6 Sol and Luna halved while Astra held, taking the flagship premium from 2.5x the mid-tier to 5x. On 29 September, GPT-6.1 Sol arrived with cached input at $0.10 and a score one point off Astra’s. And on 8 October the speed tier that would have justified the premium on interactive work became five times cheaper on Sol and, in Europe, available only there.

None of that makes Astra a bad model. It leads on long documents and output-token efficiency, and computer-use work is genuinely what it is for. It does mean the set of jobs where Astra is the correct purchase has narrowed three times in five weeks, and the remaining set is specific: US-resident, genuinely agentic desktop work, where the capability gap is worth 5x the tokens and 5x the Ultrafast rate on top.

What to do this week

If you run latency-sensitive traffic from an EU-resident project, this is the week the option appeared. Set service_tier: "ultrafast" on gpt-6.1-sol in a staging project, measure p95 against your standard path, and price it at $13.20/$66 with the regional uplift rather than the headline $12/$60. There is no equivalent path on Astra and no announced date for one.

If you hold a Pro 500 seat for Ultrafast, measure what fraction of your Codex work actually runs on it. At an 8x drain, a seat used entirely in Ultrafast has less effective headroom than Pro 100. The tier is worth paying for where a human is blocked and waiting; route background and batch work to standard explicitly rather than by default.

If you are mid-way through an Astra business case, re-run it with GPT-6.1 Sol Ultrafast as the comparator rather than Astra Standard. At $12/$60 against $10/$50 you are comparing one point of benchmark score to an 8x throughput difference at a 20% premium — a trade that was not available to evaluate a week ago.

If you are comparing this against a non-OpenAI EU option, the comparison is no longer vendor-level. Our guide to which AI APIs you can actually buy with EU data residency sets out the full model-by-tier matrix, including the two routes that offer a published SLA and the one cloud where the residency premium is 20% rather than 10%.

If you are buying subscriptions rather than tokens, nothing here changes Plus or Pro 100. Our ChatGPT review carries the current plan table and the ladder arithmetic, the pricing tracker carries the dated change log, and the API cost calculator applies the long-context surcharge that still catches people above 272,000 tokens.

Questions or corrections? Email Pick Right. Want the full list? See all news.