Cut your model bill by 89% without changing what you ship

The number on the price list is not your bill. Your bill is that number times what you send, times how often you send it, times a set of multipliers most people never read. This guide works through all four on one workload and ends at 11% of where it started. Every figure below comes from the vendors' own documentation, checked on August 23, 2026.

The workload

Say you answer support questions. 200,000 requests a month. Each request carries 12,000 input tokens, of which 8,000 are a fixed system prompt plus product documentation and 4,000 are the actual question and history. Each answer is about 800 output tokens.

That is 2,400 million input tokens and 160 million output tokens a month. On Claude Opus 5 at $5 and $25 per million, the bill is $12,000 for input and $4,000 for output. Call it $16,000.

Lever one: cache the part that never changes

Those 8,000 fixed tokens go out 200,000 times a month, identical every time. Prompt caching charges you to store them once and then charges a tenth of the input price to read them.

Anthropic publishes the multipliers plainly: a 5-minute cache write costs 1.25x the base input price, a 1-hour write costs 2x, and a cache read costs 0.1x. The documentation states the break-even outright: caching pays off after one cache read on the 5-minute duration, or after two reads on the 1-hour duration. At 200,000 requests a month you are far past both.

The fixed 1,600 million tokens now cost $0.50 per million instead of $5. Input drops from $12,000 to $800 plus $4,000 for the variable part. Bill: $8,800.

• 5-minute cache write: 1.25x base input price.
• 1-hour cache write: 2x base input price.
• Cache read: 0.1x base input price.
• Pays off after one read (5 minutes) or two reads (1 hour).

Lever two: batch anything that can wait

Both Anthropic and OpenAI cut token prices in half for asynchronous batch processing, input and output alike. OpenAI states a 24-hour turnaround and a separate pool of significantly higher rate limits. Anthropic's batch rates are listed per model: Opus 5 at $2.50 and $12.50, Sonnet 5 at $1 and $5.

Support answers that a person reads within the hour are not batchable. Overnight classification, backfills, evaluation runs, summarizing yesterday's tickets and enriching a catalog all are. In most systems more of the volume can wait than the team assumes, because nobody ever asked which parts actually need to be synchronous.

Anthropic also states that caching multipliers stack with the batch discount. Half of $8,800 is $4,400.

Lever three: stop buying the top model for every task

Claude Sonnet 5 costs $2 and $10 per million against Opus 5's $5 and $25. Haiku 4.5 costs $1 and $5. Classification, extraction, routing, tagging and short answers rarely need the top tier, and the honest way to find out is to run the same evaluation set against the cheaper model rather than to assume either way.

Move this workload to Sonnet 5, keeping caching and batch: the cached portion costs $0.10 per million, the variable input $1, the output $5. That is $160 plus $800 plus $800. Bill: $1,760, from $16,000.

Lever four: the multipliers nobody asked for

Fast mode on Claude Opus 5 and Opus 4.8 costs $10 and $50 per million, double the standard rate, and it stacks on top of caching and data residency multipliers. It is worth it when a person is waiting and not worth it when a queue is.

Pinning inference to the United States through the inference_geo parameter multiplies every token category by 1.1 on Claude 4.6 and later. OpenAI charges a 10% uplift for regional processing on models released on or after March 5, 2026. If a contract requires it, pay it. If nobody asked, do not.

OpenAI's price list splits every flagship model into short and long context: gpt-5.6-sol is $4 and $20 short, $8 and $30 long. Both are promotional rates, stated as available at least through November 21, 2026. The page does not say where the boundary sits, so find that threshold in the docs before you design a long-context workload around the headline price. Anthropic goes the other way and bills the full 1M token window at standard rates on Claude 4.6 and later, stating that a 900,000-token request costs the same per token as a 9,000-token one.

One caveat that is not a lever

Claude 4.7 and later, and Claude Mythos Preview, use a newer tokenizer that produces roughly 30% more tokens for the same text than the previous one. Anthropic says the exact increase depends on the content.

That means a per-token price comparison between Opus 5 and, say, Sonnet 4.6 understates the newer model's cost on identical text. It does not tell you anything about comparisons across vendors, because that would need measurement on your own text rather than arithmetic on a price list. If your budget is tight enough that 30% matters, count tokens on a real sample before you commit.