LLM API Pricing: 6 Costs the Rate Card Does Not Show

TL;DR: The per-million-token number on a model page is the start of the calculation, not the answer. Six mechanisms sit between it and your invoice, and each one is capable of moving the bill further than picking a different vendor would. This page names all six, links to the measurement behind each, and indexes the current price page for every model we track.

Last updated 2026-08-31. Multipliers quoted below were read from each vendor’s own caching and pricing docs on that date.

What are the six costs the rate card does not show?

1. Why does caching sometimes make the bill worse?

On current flagship models both halves of the cache are repriced, and only one of them is a discount. Anthropic states the multipliers outright:

5-minute cache write tokens are 1.25 times the base input tokens price · 1-hour cache write tokens are 2 times the base input tokens price · Cache read tokens are 0.1 times the base input tokens price

OpenAI landed on the same shape for GPT-5.6 and later, where “cache writes cost 1.25× the standard, uncached input-token rate” and reads cost 0.1x that rate. Earlier OpenAI models carry no cache-write charge at all. So caching is a bet on hit rate: write once and read once and you have paid 1.35x the ordinary input cost instead of 2x, but write and never read and you have simply paid 25% extra.

The two vendor implementations differ enough to change your architecture: Anthropic vs OpenAI prompt caching has the cost math and three cache-miss fixes. For what a real bill looks like when the hit rate is high, LLM API cache hit math works the case where the pricing page implies $50 and the invoice says $4. One detail that breaks naive cache reasoning: DeepSeek matches on complete units, not byte-by-byte prefixes, so a one-character change at the wrong place can cost you the whole hit.

2. Why does the same workload cost more at 3pm than at 3am?

DeepSeek moved the entire V4 family to peak and off-peak billing at 16:00 UTC on 2026-08-16. The same workload run at two different times of day produces two different bills, and any cost model built before that date is wrong by a multiplier rather than a percentage.

DeepSeek API price increase has the windows and the multiplier per tier. Six ways to pay less as prices rise is the mitigation side.

3. Why does one extra token double the whole request?

This is the one that surprises people most. Grok 4.6 costs $2 per million input and $6 per million output until your prompt reaches 200K tokens. Past that line, every token in the request bills at double, not only the tokens above the threshold. A 201K-token prompt does not cost slightly more than a 199K one; it costs twice as much.

The 200K cliff and when 4.5 wins has the arithmetic and the crossover.

4. Why does your cost estimator undercount reasoning models?

If your cost estimator computes prompt_tokens × input + completion_tokens × output, check whether the model’s reasoning tokens are inside completion_tokens. On Grok 4.6 they are not. On one measured coding workload that formula undercounted the billed output by 77% and the call as a whole by about 71%.

Opus 5 vs Grok 4.6 cost is the head-to-head where this surfaced: 3.6x the bill for 1.75x the code, and the accounting only reconciles once reasoning tokens are counted properly.

5. Do aggregators charge on tokens or on money?

OpenRouter does not add a per-token markup on top of provider rates. It charges a 5.5% platform fee on credit-card top-ups, dropping to 5% on crypto, plus a fixed per-transaction component that makes small top-ups proportionally worse. That is a different shape of cost from a markup, and it means the fix is topping up in larger chunks rather than switching models.

Every fee OpenRouter charges, itemized has the full breakdown.

6. When is a subscription worse than a metered key?

A flat monthly plan converts a variable cost into a fixed one, which is worth something. It also converts a spend problem into an availability problem: when the window drains, you stop working rather than paying more.

Grok Build at $30 a month versus roughly $1 per million on the API is the worked comparison. Codex weekly limit and a drop-in API that caps spend is the same trade seen from the other side, where a metered key with a hard cap gives you both predictability and a way to keep working.

Where does the current rate for each model live?

Rates move monthly, so this index points at the page we keep current for each model rather than restating numbers that will age.

For a single flat table across vendors rather than a page each, AI API pricing comparison has twenty-three models side by side with cache and batch math. Read the date line inside it before quoting a number: the table body reports a late-May 2026 verification, and several rates have moved since, DeepSeek’s most of all.

Video pricing works differently enough to need its own map. The Seedance API guide has the per-second, per-resolution structure.

How do you pay less without changing models?

Four routes, in rough order of how much they return per hour spent.

1. Route by task. Hybrid routing in Claude Code cuts cost about 80% by sending the cheap 90% of turns to a cheap model. Reduce AI API costs by 60% is the general version.
2. Use the free tiers that are actually usable. Free LLM API tiers ranked for coding separates the ones with real caps from the ones that die on your first agent loop. Per model: DeepSeek V4 Flash, GLM 5.2, Kimi K2.7 Code.
3. Take the discounts that exist. Discounted LLM APIs lists 13 models currently below list price.
4. Cap the downside. A hard spend cap converts a runaway loop from an invoice into an error, and errors are cheaper. If you are handling those errors, LLM API error codes covers which billing failures arrive as 402 and which arrive as 429.

References

• OpenAI API pricing
• Claude pricing
• Gemini API pricing
• DeepSeek API pricing
• xAI models and pricing
• OpenRouter models