Model APIs now sell the same model at several speeds. A request can be processed slower and for less, at the standard rate, or faster for a premium, and the parameter that chooses is one field in the request body. OpenAI calls the options service tiers (Flex, Standard, Fast, Ultrafast)1; Anthropic has fast mode for Claude Opus 5.5 and a Priority Tier it no longer sells23; Google's Gemini API has Flex and Priority, both in preview45.
The short answer to "which tier should we use" is: standard by default, a slow tier for anything nobody is waiting on, and a fast tier only on the few user-facing paths where output speed is measurably the bottleneck. The rest of this guide gives the prices, the promises and the failure behaviour you need to make that call.
The tiers side by side
| Vendor and tier | How to ask for it | Price against standard | What it promises | When capacity runs out |
|---|---|---|---|---|
| OpenAI Flex | service_tier: "flex" | Batch rates (half) | Slower responses, occasional unavailability | 429 Resource Unavailable, not charged |
| OpenAI Fast | service_tier: "fast" or "priority" | Twice Standard | Up to 2.5x faster, more consistent latency | Ramping too fast downgrades to Standard, billed at Standard |
| OpenAI Ultrafast | service_tier: "ultrafast" | Six times Standard on GPT-6.1 Sol | Fastest tier; no figure in the guide | Separate rate limits; behaviour not documented |
| Anthropic fast mode | speed: "fast" | Twice standard on Opus 5.5 | Up to 2.5x output tokens per second | 429 or 529 error |
| Anthropic Priority Tier | Existing commitments only | Committed capacity | Prioritized requests, 99.5% uptime target | Overflow falls back to standard |
| Gemini Flex | service_tier: "flex" | Half | Minutes, 1 to 15 minute target | 503 or 429, no fallback |
| Gemini Priority | service_tier: "priority" | 75 to 100 percent more | Second response times, non-sheddable | Downgraded to Standard, billed at Standard |
Every row comes from the vendor's own page6782354. Two of the speed claims are "up to", which is a ceiling, not a typical value. None of the three vendors publishes a latency number for its fastest tier that you can plan against.
OpenAI: Flex, Fast and Ultrafast
OpenAI's pricing page lists five ways to process a request: Standard, Batch, Flex, Fast and Ultrafast, and notes that Priority processing was renamed Fast mode on 30 July 20269. The changelog entry says requests tagged priority automatically use Fast mode10.
| GPT-6.1 Sol, per million tokens, up to 272K input | Input | Cached input | Output |
|---|---|---|---|
| Flex (same as Batch) | $1.00 | $0.05 | $5.00 |
| Standard | $2.00 | $0.10 | $10.00 |
| Fast | $4.00 | $0.20 | $20.00 |
| Ultrafast | $12.00 | $0.60 | $60.00 |
Fast is a pay-as-you-go tier for user-facing applications; OpenAI describes "up to 2.5× faster speeds and more consistent latency"7. It can be set per request or as a project default. Standard and Fast share one rate limit for a given model. The catch is the ramp limit: if traffic grows too quickly, some Fast requests are downgraded to standard speed, billed at standard rates, and report service_tier: "default". OpenAI's rule of thumb is that once traffic reaches 1M input tokens per minute, you increase it by no more than 50 percent every 15 minutes, and it advises against running large ETL or batch jobs in Fast mode7. GPT-6 Astra's Fast mode has no latency SLA7.
Ultrafast is OpenAI's fastest tier. The guide lists GPT-6.1 Sol and GPT-6 Astra, available to all API users, with its own rate limits (for GPT-6.1 Sol, 1M tokens per minute on the Build tier by default) and a strong recommendation to use WebSockets, because "without a persistent connection, network overhead can reduce the latency gains"8. The changelog dates Ultrafast for GPT-6.1 Sol in the Responses API to 8 October 202610. When OpenAI announced Ultrafast in limited preview on 13 August, for GPT-5.6 Sol, it said "up to 14x faster than Standard processing"10; the current guide gives no speed figure and does not describe what happens when Ultrafast capacity is short8. GPT-6 Luna has no Ultrafast price9.
Flex bills at Batch rates for Responses or Chat Completions requests "in exchange for slower response times and occasional resource unavailability", is in beta with limited model availability, and returns 429 Resource Unavailable when there is no capacity, without charging6. OpenAI suggests raising the SDK timeout to 15 minutes and either backing off or retrying with service_tier: "auto" when finishing matters more than price6.
Since 21 October 2026, OpenAI also lets you choose which service tiers each project may use in Platform settings10. That is the control a finance lead should ask about first.
Anthropic: fast mode, and the end of Priority Tier
Anthropic's fast mode runs "the same model with a faster inference configuration", with up to 2.5x higher output tokens per second; the gain is in output speed, not time to first token, and is most visible when streaming2. You opt in with speed: "fast" on a Messages request. On 15 October 2026 fast mode for Claude Opus 5.5 left research preview: no access request, no beta header, and a default fast mode rate limit in the Start, Build and Scale usage tiers. It is also available on Claude in Microsoft Foundry for Azure-hosted deployments11, and on Claude Models on Google Cloud (Anthropic managed)2.
Fast mode for Opus 5.5 costs $8 input and $40 output per million tokens across the full context window, twice the standard $4 and $20; caching and data residency multipliers stack on top, and it is not available with the Batch API or with a Priority Tier commitment2. When the fast mode limit or capacity is exhausted the API returns a 429 or 529, and the response's usage.speed field says which speed served a successful request. Falling back to standard speed is your code's job, and Anthropic warns that it misses the prompt cache, because requests at different speeds do not share cached prefixes2.
Priority Tier is a different thing: committed capacity that Anthropic prioritizes over other traffic, with a 99.5 percent uptime target and overflow falling back to standard3. The page now opens with a warning that commitments "are no longer available for purchase"; existing ones run to their contract end, and it does not support Fable 5.1, Opus 5.5, Sonnet 5.5 or Haiku 5.53. If a Priority Tier commitment is part of your reliability plan, it ends with the contract or with a move to a current model, whichever comes first.
Google: Gemini Flex and Priority
Google's Gemini API offers both, in preview, on the Interactions API45. Priority costs 75 to 100 percent more than Standard and is "designed for second response times for interactive, user-facing AI tools". It is non-sheddable, its default limits are 0.3 times the standard rate limit, and when limits are exceeded overflow requests are downgraded to Standard and billed at Standard rather than failing; the x-gemini-service-tier response header shows it4. Flex is half price with a latency target of 1 to 15 minutes, can be preempted during standard-traffic spikes, returns 503 or 429 with no server-side fallback, and Google recommends client timeouts of 10 minutes or more5. Both list the current Gemini models, including Gemini 4 Argon and Gemini 3.8 Flash.
What the premium costs per request
The arithmetic is simple and it is the point: a fast tier doubles or sextuples the cost of every request it touches. Applying it to the whole workload by setting a project default is the most expensive way to buy speed, and OpenAI's ramp limit means a sudden switch may not even deliver it7.
When paying for speed is worth it
Paying for speed works when three things are true together: a person is waiting, the answer is long enough that output rate dominates what they feel, and faster output changes something you measure. A streamed code edit, a long drafted reply or a voice agent's turn can meet all three.
It usually loses in these cases:
- Short answers. A 50-token classification is dominated by time to first token and network, and Anthropic is explicit that fast mode targets output speed, not time to first token2.
- Agent loops. If most of a turn is spent in tools, retrieval or your own code, faster generation shortens the smallest part.
- Background work. Evaluations, enrichment and nightly jobs belong on Flex or Batch, where the same tokens cost half65.
- Capacity insurance. OpenAI Fast and Gemini Priority downgrade under load, and Anthropic fast mode errors. None of them is a guarantee of speed when demand spikes, which is often when you wanted it.
The slow tiers lose when the work has a deadline: a Flex request can wait minutes or fail with no fallback, so anything a user or a downstream SLA depends on needs a retry path to standard.
The CEO question: which paths justify it
For a leader, the decision is not per vendor but per path. Ask the team for a list of user-facing paths with, for each: the model, the median and 95th percentile time to first token and total response time, the output length, and the business metric that latency is believed to move. Then ask three questions.
- Where is time actually spent? If time to first token or tools dominate, a fast tier buys little.
- What would the premium cost at current volume? Use the multipliers above: twice standard for OpenAI Fast and Anthropic fast mode, six times for GPT-6.1 Sol Ultrafast, up to twice for Gemini Priority.
- Who can turn it on? A project default or one config change can move all traffic to a premium tier. OpenAI's per-project tier permissions exist for this10; make sure someone owns them.
A sound policy fits on one line: standard everywhere, Flex or Batch for background work, a fast tier on named paths with a measured reason and an owner, reviewed monthly against the invoice.
Watching tiers and their cost with INFRO
INFRO is a control plane that sits between your applications and the provider accounts you already hold. Your providers bill you directly at whatever tier served each request; INFRO does not change provider prices or capacity.
Measure before you pay. Each request INFRO records carries its latency and time to first token, and usage analytics return p50, p95 and p99 latency by model, project and key, so you can see which paths are slow before buying speed. Traffic changes raise a finding when a scope's p95 latency or time to first token shifts.
Fall back without code. A route with failover moves a request to another target on your own connections when one fails, which covers the "premium tier is out of capacity" case for errors the gateway sees.
Parameters pass through. On /v1/chat/completions, top-level parameters INFRO does not model are forwarded to the provider unchanged (chat completions reference), so service_tier reaches an OpenAI connection as you sent it.
What INFRO does not do today: its per-request cost estimate uses the standard list or negotiated rate and does not apply a tier premium, so a Fast or Ultrafast request is estimated at its standard price. Provider cost reconciliation compares INFRO's estimates with your provider's own cost report, which is where tier charges show up; treat the gap as the cost of the tiers you chose.