You can buy a faster answer now. Check who is waiting first

You put a model behind something a customer touches. A chat box on the pricing page, a phone line, a form that writes back. It works, and it is slow. Not broken-slow, just slow enough that people type a second message before the first answer lands, or close the tab. So you swap in the small cheap model, watch the answers get worse, and file it under trade-offs.

That trade-off used to be close to a law of physics. It is now a line on a price list. OpenAI and Anthropic both sell their flagship models at two speeds, and depending on which model you are on, the fast one costs two to four times the standard one [1][2]. They also both sell a slower tier at half [1][2]. Which means the interesting question is no longer which model is fast enough. It is who is waiting, and what a few seconds of their attention is worth to you. This guide is for someone running a business of one to about ten people who already has a model in front of a customer, or inside a loop they sit and watch. It is not for a team with a latency budget and an inference engineer. It is not for someone whose only use of AI is typing into a chat window, where none of this reaches your bill.

Speed became a priced tier, not a property of the model

Read the current price lists side by side and the pattern is clear, though the size of the premium is not the one usually quoted.

On OpenAI’s list, GPT-5.6 Sol costs $2 per million input tokens and $10 per million output tokens on standard processing, and $8 and $40 in Fast mode [1]. That is four times, not double. GPT-6 Astra is $5 and $25 standard against $20 and $100 fast, also four times [1]. GPT-5.3-Codex is the exception, at $1.75 and $14 standard against $3.50 and $28 fast, which is double [1]. If you are budgeting, look up the multiple on the model you actually call. OpenAI also renamed this tier partway through the year: “Priority processing was renamed Fast mode on July 30, 2026” [1].

Anthropic’s premium is the smaller one and it is uniform. Claude Opus 5 is $5 per million input tokens and $25 per million output on standard pricing, and $10 and $50 in Fast mode, exactly double [2][3]. Anthropic is unusually direct about what you are buying: Fast mode delivers “up to 2.5x higher output tokens per second” from the “same model weights and behavior (not a different model)”, with only the inference configuration changed [3].

The same menu runs downward. Batch processing is half price at both vendors [1][2]. OpenAI’s flex tier prices tokens “at Batch API rates” while still answering synchronously [6]. So for one set of model weights you can now pay anywhere from half to four times, and the only variable is how long you are prepared to wait.

The most-quoted example of the ceiling being raised is OpenAI’s Ultrafast preview, announced on 13 August 2026, which runs GPT-5.6 Sol “up to 14× faster than Standard processing” and up to 750 output tokens per second on Cerebras hardware [4]. That is one announcement and it is still a limited preview [4]. The durable fact is the price list underneath it, which will outlast any particular model name on it.

Two clocks run during a response, and your user feels one at a time

There are two numbers that get called latency and they have different cures.

The first is time to first token: the pause between the request going out and anything at all appearing. The second is output tokens per second: how fast the answer arrives once it starts. A user staring at a spinner is complaining about the first. A user watching a paragraph crawl out over 20 seconds is complaining about the second.

This distinction decides which purchase is correct, and the vendors say so themselves. Anthropic states that with Fast mode, “speed benefits are focused on output tokens per second (OTPS), not time to first token (TTFT)” [3]. If your problem is the silence before the first word, the premium tier is the wrong thing to buy. That problem lives in your prompt length, your retrieval step, and your caching, not in generation speed.

Diagnosing it costs nothing. Turn on streaming, run 10 real requests, and note two times for each: when the first word appeared, and when the last one did. If most of the wait is before the first word, go to caching. If most of it is after, generation speed is genuinely your constraint.

Half of your calls have nobody waiting, and that half is cheaper

Before buying speed, sort your model calls into two piles: a person is on the hot path, or a person is not.

The second pile is usually larger than operators expect. Nightly summaries of yesterday’s enquiries. Tagging a backlog. Enriching a spreadsheet. Drafting content you will review tomorrow morning. Running your own evaluations. None of these has anyone sitting there, and all of them can go to the tier that costs half.

Anthropic’s Message Batches API gives 50 percent off standard API prices, with most batches completing within 1 hour and results available when all messages have completed or after 24 hours, whichever comes first [7]. A batch that has not finished within 24 hours expires [7]. A single batch is limited to 100,000 requests or 256 MB, whichever is reached first, and results stay downloadable for 29 days [7]. Batch discounts stack with prompt caching discounts [7].

OpenAI’s flex processing is the middle option: batch-level pricing, but you still call the API and get an answer back, accepting slower response times and the occasional 429 Resource Unavailable when capacity is short, for which “you will not be charged” [6]. The default timeout is 10 minutes with an official OpenAI SDK, and OpenAI’s own sample code raises it to 15 [6]. The trade is that you have to write the retry logic, with exponential backoff or a fallback to standard processing [6].

Do this sorting first, because it changes the size of the problem. Once the not-waiting pile is moved to half price, whatever is left in the waiting pile is a much smaller bill, and paying a premium on a small bill is a much easier decision.

The free levers move the clock more than the paid one does

OpenAI publishes a latency guide built on seven principles, and two of its numbers should change what you do first [5].

“Cutting 50% of your output tokens may cut ~50% of your latency” [5]. That is close to a one-to-one return, and it is free. Capping the output token limit, asking for the answer in three sentences, and using a structured output format with minimal syntax all pay off immediately.

Against that: “Cutting 50% of your prompt may only result in a 1–5% latency improvement” [5]. So the long system prompt you have been meaning to trim is almost certainly not your speed problem. Trim it for cost if you like, but not for the clock.

The third free lever is streaming, which OpenAI calls the single most effective approach to making users wait less, cutting the waiting time to a second or less [5]. It does not make anything faster. It moves the perceived wait to near zero, which is what you were actually buying. The remaining principles are worth reading in order: make fewer requests by combining sequential steps, parallelise the steps that do not depend on each other, and do not default to a model at all where a lookup, a pre-computed value or ordinary code would do [5].

Then there is caching, which is the correct fix for the first-token clock. Anthropic says of prompt caching that “you will generally see improved time-to-first-token for long documents”, with cache hits billed at 0.1x the base input price, 5-minute cache writes at 1.25x and 1-hour writes at 2x [8]. The catch is the minimum cacheable prompt length, which differs by model: 512 tokens on Claude Opus 5, 1,024 on Claude Sonnet 5, and 4,096 on Claude Haiku 4.5 [8]. Shorter prompts are processed without caching and no error is returned, so a cache you think is running may not be [8]. OpenAI’s equivalent prices cached input at 10 percent of the base input rate, with cache writes at 125 percent [1]. If you send the same instructions and the same reference document on every call, this is the cheapest second you will ever buy back.

Price the routing decision before you price the speed decision

The gap between two speeds of one model is small next to the gap between two models.

On OpenAI’s current list, GPT-5.6 Luna costs $0.10 per million input tokens and $0.60 per million output tokens [1]. GPT-5.6 Sol in Fast mode costs $8 and $40 [1]. That is 80 times the input price and about 67 times the output price. Any conversation about whether to pay a speed premium is a rounding error next to the question of which model should have been handling that call in the first place.

So price the volume on the smaller model first. If a small model handles the task acceptably, it is both cheaper and, as OpenAI’s own guide notes, generally faster, because smaller models usually run faster [5]. The fast tier of a flagship is what you buy when a small model genuinely cannot do the job and a person genuinely cannot wait.

Run it with the defaults, which use the standard and Fast mode output prices for GPT-5.6 Sol [1], and the premium comes to about $189 a month. For most small operations that is the whole finding: the fast tier is affordable and beside the point, or the wait was never a generation-speed problem to begin with. It only becomes a real number at volumes where you would already have someone watching the bill.

What you are buying today is a preview with edges

Neither end of this menu is a switch you flip.

Anthropic’s Fast mode is a research preview that requires access, granted through an account manager or a waitlist [3]. It runs only on Claude Opus 5 and Claude Opus 4.8, and only on the Claude API and Claude Managed Agents, not on Amazon Bedrock, the Claude Platform on AWS, Google Cloud or Microsoft Foundry [3]. You enable it by setting speed: "fast" with the fast-mode-2026-02-01 beta header [3]. It is not available with the Batch API or with Priority Tier, it has dedicated rate limits separate from standard Opus limits, and requests at different speeds do not share cached prefixes [3]. That last one matters, because it means you cannot cheaply A/B the two speeds on a cached workload without paying to rebuild the cache each time.

OpenAI’s Ultrafast is available “in a limited preview today to a select group of customers”, and OpenAI says it will expand access as capacity grows [4]. It publishes no price [4].

Flex has the opposite edge. It is available and cheap, but the 429s are real and you own the retry [6].

What still goes wrong

Every speed figure here is a ceiling. “Up to 14× faster than Standard processing” and 750 output tokens per second are OpenAI’s best-case numbers for Ultrafast [4], and “up to 2.5x higher output tokens per second” is Anthropic’s for Fast mode [3]. Your own prompts, with your own input and output lengths, will land somewhere below those. Measure before you commit, and measure again after the preview opens up, because a curated preview cohort and general availability are different load conditions.

The prices are less stable than they look. OpenAI’s GPT-5.6 Sol pricing is promotional at least through 21 November 2026 [1], and the tier itself was renamed from Priority processing on 30 July 2026 [1]. Fast mode at Anthropic is a research preview [3] and Ultrafast has no published price at all [4]. Anything you build a margin on today should be re-priced when those states change, and a workflow that only pencils out at promotional rates is a workflow with a date on it.

Then there is the honest limit of the whole topic. In a business of your size, the model is often not the slow part. The slow part is the polling interval on your automation tool, a retrieval step that reads more than it needs, a webhook that queues, or a human approval sitting in someone’s inbox overnight. Buying a faster tier fixes none of that, and it is the one purchase in this guide that gives you a receipt without giving you a result. Time the whole path before you pay to speed up one segment of it.