Guides / llm fast mode pricing
LLM fast mode pricing, and what it excludes
Fast mode is a premium latency tier that bills the same tokens at a higher rate. OpenAI charges double on most models and 2.5 times on GPT-5.5, and Anthropic charges double on two Opus models. Culpa, a local-first LLM cost, margin, and forecast ledger, prices each call from a versioned price book, so a tier change shows up as a cost rather than a setting.
Why this happens
Fast mode buys latency with money and the exchange rate isn't uniform. OpenAI renamed priority processing to Fast mode on 2026-07-30 and accepts either service_tier value, so old code and new documentation describe the same billing. Its multiplier is 2x on the GPT-5.6 family and on GPT-5.4, and 2.5x on GPT-5.5, which is the one worth knowing because it's the model many teams are still on. Anthropic's is a research preview covering Claude Opus 5 and Opus 4.8 only, at exactly double, and the restrictions around it matter more than the rate. It runs on the first-party Claude API only rather than on partner cloud platforms. It's unavailable with the Batch API, so the two biggest price modifiers in the catalogue can't be combined. And on Opus 4.6 a fast request runs at standard speed and bills at standard rates, so asking for it and silently not getting it counts as supported behaviour.
What this usually looks like
- Latency improved after a config change and nobody priced what it cost.
- A service_tier value is set in code and no one remembers which model it applies to.
- You tried to batch a fast-mode workload and the request was refused.
- Your GPT-5.5 spend is higher per token than your GPT-5.6 spend, on the same tier.
- A fast-mode flag is set on a model that ignores it, so you pay standard and think you're paying more.
Free, no card, no account
Run the free Cost Leak Scan
It shows your most expensive conversation before you install anything.
Mistakes that cost the most
| Mistake | Why it hurts | Do instead |
|---|---|---|
| Assuming the fast multiplier is the same on every model. | OpenAI charges 2x on most and 2.5x on GPT-5.5, so a per-model check is the only reliable one. | Read the fast-mode table per model rather than applying one multiplier across a family. |
| Planning to batch a fast-mode workload. | Anthropic states fast mode isn't available with the Batch API, so the 50% discount and the premium tier are mutually exclusive. | Decide which one the workload actually needs, because you can't have both. |
| Setting a fast flag globally. | Some models error, some silently run standard, and some bill double. All three look like working code. | Set it per call site, and check the reported cost rather than the config. |
| Buying latency you can't measure the value of. | Doubling the rate is easy to justify in a meeting and hard to justify against a number nobody produced. | Price the premium on last month's real volume before enabling it, then compare against what the latency earned. |
Run this check tonight
- Grep for service_tier and speed flags, and list which models each one reaches.
- Price last month's volume at the standard rate and at the fast rate, and look at the gap.
- Check whether any fast-flagged model silently ignores the flag.
- Confirm nothing in your batch pipeline also requests fast mode.
The same volume on three tiers
100M input and 10M output tokens a month on GPT-5.5, priced from OpenAI's own pricing page read 2026-08-03. GPT-5.5 is chosen deliberately because it carries the highest published fast multiplier. Standard is $5.00 and $30.00 per million, fast mode is $12.50 and $75.00, and batch is half of standard. Token volume is modelled and every rate is published.
The spread between the cheapest and dearest way to run the same tokens is five times, and none of that comes from a discount you negotiate. It's three settings. Knowing which one each workload runs on is worth more than most optimisation work.
Every number, with its confidence and source
| Figure | What it means | Confidence | Source |
|---|---|---|---|
| $2,000.00 per month | modelled GPT-5.5 volume on fast mode, against $800.00 standard and $400.00 batched | calculated | 100M input and 10M output tokens at OpenAI's published GPT-5.5 rates read 2026-08-03: fast mode $12.50 and $75.00 per million gives $2,000.00, standard $5.00 and $30.00 gives $800.00, batch at half of standard gives $400.00. Token volume is modelled and all three rate pairs are published. |
| 2.5x | OpenAI's published fast-mode multiplier on GPT-5.5, against 2x on every other model listed | provider-reported | developers.openai.com/api/docs/pricing, read 2026-08-03. GPT-5.5 standard is $5.00 and $30.00 per million and its fast-mode row is $12.50 and $75.00, which is 2.5x on both. The GPT-5.6 family and GPT-5.4 are 2x on both. Reported because a single multiplier applied across a family would be wrong on the model many teams are still running. |
What a generic answer can’t know
A provider invoice shows what you were charged and not which tier produced it, so a bill that doubled because someone set a flag looks identical to a bill that doubled because traffic grew. Culpa prices each call from a versioned price book and records what it was actually billed at, which is what turns a tier into a line you can see. That matters most for the silent cases in this tier's own documentation: a model that errors on a fast request tells you immediately, and a model that runs at standard speed and bills at standard rates tells you nothing at all. Beyond that, Culpa sets the premium against the revenue the customer who triggered it pays you, so buying latency becomes a margin decision rather than a config change.
Questions founders ask next
What does fast mode cost?
On OpenAI, double the standard rate on the GPT-5.6 family and GPT-5.4, and 2.5 times on GPT-5.5. On Anthropic, double on Claude Opus 5 and Opus 4.8, which are the only two models it covers, at $10.00 and $50.00 per million against a standard $5.00 and $25.00. Read from both pricing pages on 2026-08-03.
Was fast mode called something else?
Yes. OpenAI's pricing page states priority processing was renamed Fast mode on 2026-07-30, and that requests can use either service_tier value. So older code and newer documentation describe the same billing, which is worth knowing when auditing what a flag does.
Can I combine fast mode with batching?
No. Anthropic states plainly that fast mode isn't available with the Batch API. Those are the two largest modifiers in the catalogue, one doubling the rate and the other halving it, and they can't be applied to the same request.
What happens if I request fast mode on a model that doesn't support it?
It depends on the model, and both behaviours exist. Anthropic states Opus 4.7 returns an error for a fast request, while Opus 4.6 runs at standard speed and bills at standard rates. The second is the one to watch, because the flag is set, the code works, and nothing you asked for happened.
On your infrastructure
Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.
How Culpa works
Find the culprit. Not just the total.
Your dashboard shows what you spent. It stops short of who spent it. Culpa shows the conversation, the user and the feature behind it.
Your prompts stay local.
Culpa runs on your own infrastructure. What you send to a model reaches us at no point.
Every dollar has a name.
Follow any charge to the conversation, the user, the feature and the customer behind it.
See the bill before it lands.
Cost your next feature before you ship it. You get the likely bill and the worst case, at best, median, p90 and p99.
Three steps to your first answer.
Change one base URL.
Or drop in the Python or TypeScript library.
Find your most expensive conversation.
In the first session, not the first week.
Cost your next feature before you ship it.
Why the bill went up
Example dashboardCalls traced
418,209
across 3 projects
Spend this week
$378.41
+ $182 vs last week
Failed calls
312
74% retried, and you paid for all of them
+ $182 this week traced to one culprit
Spend over 14 days
Most expensive users
Next week forecast
Graded against reality. Accuracy shown as results land.
Keep reading
Sources: OpenAI API pricing, Anthropic pricing. Last reviewed 2026-08-03, rates effective 2026-07-02. Plain text version.