OpenAI Fast mode: when paying for 2.5x speed is worth it

OpenAI Fast mode is the old priority processing tier under a new name, and its headline promise, up to 2.5x faster than Standard, comes with a price multiplier that is not the same on every model. On GPT-5.6 Sol you pay exactly twice the Standard rate; on GPT-5.5 you pay 2.5 times. This post gives you the per-model premium, a way to price each second you save, and the cases where Fast mode is money wasted.

• Renamed: priority processing became Fast mode on 30 July 2026. service_tier: "fast" and service_tier: "priority" both work.
• Speed: "up to 2.5x faster" than Standard. The guide states this figure explicitly for gpt-5.6-sol.
• Price on Sol: twice Standard, $8 / $40 per 1M input / output tokens for short-context requests.
• Next tier: Ultrafast, up to 14x faster for GPT-5.6 Sol, in limited preview since 13 August.

What changed when priority processing became OpenAI Fast mode

The rename is backward compatible. The 30 July changelog entry says requests tagged priority automatically use Fast mode, and on the same date OpenAI raised Fast mode's speed for gpt-5.6-sol to the "up to 2.5x" figure. On 5 August it extended Fast mode to long-context prompts over 272K tokens on Sol, Terra and Luna.

One detail will confuse your dashboards. For GPT-5.6 and earlier models the response's service_tier field returns priority even when you sent fast, and the usage dashboard groups those requests as priority too. If your logging filters on the string fast, it will count nothing.

How do I enable OpenAI Fast mode?

You enable Fast mode per request with service_tier: "fast" on the Responses or Chat Completions API, or for a whole project under Settings, General, Project Service Tier. With the project setting, requests that do not specify a tier default to Fast, and the Fast mode guide says traffic transitions "gradually" rather than at once.

client.responses.create(
    model="gpt-5.6-sol",
    input="...",
    service_tier="fast",
)

OpenAI Fast mode pricing: the premium is not always 2x

OpenAI Fast mode pricing sits in its own table on the pricing page, headed "Fast mode", separate from Standard, Batch and Flex. Short-context rates per 1M tokens, input / output, as of 21 September 2026:

Model	Standard	Fast mode	Multiplier
gpt-6-astra	$10.00 / $50.00	$20.00 / $100.00	2x / 2x
gpt-5.6-sol	$4.00 / $20.00	$8.00 / $40.00	2x / 2x
gpt-5.6-terra	$2.00 / $12.00	$4.00 / $24.00	2x / 2x
gpt-5.6-luna	$0.20 / $1.20	$0.40 / $2.40	2x / 2x
gpt-5.5	$5.00 / $30.00	$12.50 / $75.00	2.5x / 2.5x
gpt-5.4	$2.50 / $15.00	$5.00 / $30.00	2x / 2x
gpt-5	$1.25 / $10.00	$2.50 / $20.00	2x / 2x
gpt-5-mini	$0.25 / $2.00	$0.45 / $3.60	1.8x / 1.8x
o3	$2.00 / $8.00	$3.50 / $14.00	1.75x / 1.75x
gpt-4.1	$2.00 / $8.00	$3.50 / $14.00	1.75x / 1.75x

The GPT-5.6 family and GPT-6 Astra are a clean 2x. GPT-5.5 is 2.5x. o3 and gpt-4.1 are cheaper to speed up, at 1.75x, and gpt-5-mini sits at 1.8x. Some models you may still run, including gpt-5-nano, gpt-5-pro and o3-pro, do not appear in the Fast table at all. Cached input discounts still apply in Fast mode, which makes caching the first thing to fix before you pay for speed; our write-up on prompt caching covers how.

The cost per saved second, worked through

Speed has a price per second only once you fix a latency baseline, and OpenAI does not publish one. The numbers below are assumptions, labelled as such. Replace them with your own p50 latency and token counts.

• Assumption: a GPT-5.6 Sol request with 2,000 input and 600 output tokens.
• Assumption: 8 seconds end to end on Standard.
• Standard cost: 2,000 x $4 / 1M + 600 x $20 / 1M = $0.008 + $0.012 = $0.020.
• Fast cost: 2,000 x $8 / 1M + 600 x $40 / 1M = $0.016 + $0.024 = $0.040.
• Premium: $0.020 per request.

At the full 2.5x, 8 seconds becomes 3.2, saving 4.8 seconds: $0.020 / 4.8 = about $0.0042 per saved second. "Up to" is a ceiling, though. If your traffic sees 1.5x (our assumption, not OpenAI's), 8 seconds becomes 5.3, you save 2.7 seconds, and each one costs about $0.0075, nearly twice as much.

Scale it to decide. At 100,000 requests a month the premium is $2,000. At the full speed-up that buys 480,000 seconds, about 133 hours of users not waiting, or roughly $15 per hour of waiting removed. That is the number to put in front of a product owner, next to what a slow answer costs in abandoned sessions.

Is OpenAI Fast mode worth the cost?

OpenAI Fast mode is worth the cost where a person watches the response arrive and traffic is steady. The guide calls it "ideal for high-value, user-facing applications with regular traffic where latency is paramount". A support chat, an in-app copilot or a checkout assistant fits. It is the wrong choice in five cases:

1. Batch and ETL jobs. The guide says to avoid running them in Fast mode. Nobody waits, so the Batch tier at half the Standard rate is the better fit.
2. Spiky traffic. If traffic ramps too fast, some requests are downgraded to Standard speed and billed at Standard rates, returning service_tier: "default". The rule of thumb: above 1M input tokens per minute, grow by no more than 50% every 15 minutes.
3. Fine-tuned models and embeddings. Fast mode does not support them.
4. GPT-6 Astra with EU data residency. Fast mode is unavailable for that combination, and Astra's Fast mode carries no latency SLA.
5. When a cheaper lever exists. The reasoning guide points to gpt-5.6-luna for "the lowest cost and latency". Dropping reasoning effort or model tier may save more seconds than Fast mode, for less money.

The strongest case against our framing is that latency is not linear in value: shaving 5 seconds off a 30-second agent run may matter far less than shaving 2 off a 3-second chat reply. That is true, and it argues for applying Fast mode per route, not per project. In the AI builds we run, only the handful of routes a user watches are candidates. Fast mode and Standard share one rate limit per model, so the split costs you no extra quota.

Ultrafast mode: 14x, limited preview, no public price

OpenAI announced Ultrafast mode on 13 August as a new API service tier for GPT-5.6 Sol that "runs up to 14x faster than Standard processing", available in limited preview to select customers. No price is published on the pricing page, so there is no cost per saved second to compute yet. Plan with Fast mode and treat Ultrafast as a future option, not a line in this quarter's budget.

If you are choosing between Sol, Terra and Luna before you choose a tier, our comparison of GPT-6 Astra vs GPT-5.6 covers the model decision, and our AI engineering team can instrument the latency baseline this calculation needs. The next step is the same either way: log p50 and p95 latency per route for a week on Standard, then price Fast mode against those numbers rather than against "up to".