Input Rate

The input rate is the per-token price tag on everything you feed into an AI model. It's the meter running every time the model reads. This isn't just your latest prompt; it's the entire working context: the system prompt, all the tools it knows about, every file it has processed, and the full transcript of your conversation up to that point. For an AI agent, this context stack gets resent on every single turn.

This is where the real money goes. Our data from 157,670 Claude Code API responses between April and July 2026 showed that visible output accounted for a mere 12.1% of the bill. The other 87.9%? That was input. Specifically, cache reads made up 48% and cache writes 39%. Uncached input was a rounding error. The input rate is the base multiplier for these dominant costs. A cache read, for instance, is priced at 0.1 times the base input rate, while a cache write can be 1.25 or 2 times the base rate, depending on the cache TTL. Even though cache reads are cheap per token, their sheer volume because they are resent on every turn makes them the biggest line item on your invoice. It sets the stage for all of it.

The input rate is not the output rate. This is the biggest AI billing misconception. Most people obsess over how much the model writes, thinking that's the primary driver of cost. They shorten answers, cut verbosity, starving the model of its purpose. Our data proves this is a fool's errand. Optimizing for shorter answers targets only 12% of your bill. The output rate, while higher per token than the input rate, applies to a much smaller volume of tokens.

It's also not just the tokens you explicitly type. It isn't a simple character count of your prompt. The model processes a lot more than what you see. For example, Anthropic's extended thinking documentation makes it clear that you are billed for full thinking tokens as output, even if the API returns a condensed summary. While those are technically output tokens, the input rate is what defines the baseline cost for the vast majority of the data the model processes internally and externally. Don't confuse the cost of generating text with the cost of processing context. One is a rounding error, the other is the bill.

Anthropic's pricing, 26 July 2026, shows Claude Opus 5 and Opus 4.8 both have an input rate of $5 per million tokens. Their output rate is $25 per million. Now, consider Claude Fable 5. Its input rate is $10 per million tokens, double that of Opus. Its output rate is $50 per million, also double.

Output rates alone suggest Fable 5 is just twice as expensive across the board. But that's not how it works. Our two-month workload, re-priced against Fable 5, came out about 1.9 times the Opus 5 cost. Almost exactly its input multiplier. Why? Because the input rate applies to the 87% of your bill that is cache traffic.

Imagine a typical agent turn. You send a small prompt, maybe 100 tokens. But the agent resends 50,000 tokens of context. If that context is cached, it's a cache read at 0.1x the input rate. For Opus 5, that's $5/M * 0.1 = $0.50/M. For Fable 5, it's $10/M * 0.1 = $1.00/M. That 50,000 tokens of context costs $0.025 for Opus 5 and $0.05 for Fable 5. Now, if your session idles past five minutes, that context needs a fresh cache write. For Opus 5, that's $5/M * 1.25 = $6.25/M. For Fable 5, it's $10/M * 1.25 = $12.50/M. The 50,000 tokens now cost $0.3125 for Opus 5 and $0.625 for Fable 5. The output, even if it's 2,000 tokens at $25/M for Opus 5, is only $0.05. The input rate, and its multipliers, are what drive the invoice. High input rate means expensive model, period.

Focus optimization squarely on the input rate. This means ruthless context discipline. 1. **Cut what gets resent:** Fewer files in context, narrower reads, compact the transcript aggressively, and end a session instead of letting it idle past Anthropic's default five-minute cache TTL. A coffee break is a billable event if your session expires. 2. **Match the model to the input rate:** Route routine agent turns to models with a cheaper input tier. For example, Sonnet 5, with its introductory $2 per million input tokens through August 31, 2026, is a massive win for everyday tasks compared to Opus 5's $5. Reserve the expensive tiers like Opus or Fable for turns where their advanced capabilities genuinely pay off, not for every casual interaction.

**When not to:** 1. **Do not obsess over output length:** Asking for shorter answers is a waste of time and degrades the quality of the model's work. You are optimizing 12% of the bill at the expense of utility. It's a bad trade. 2. **Do not let sessions idle:** Leaving an agent session open and unused for more than five minutes means the next turn pays for a cache write at 1.25 times the input rate instead of a cheap cache read at 0.1 times. That's a 12.5x cost increase. 3. **Do not use high input rate models indiscriminately:** Deploying Claude Fable 5 ($10/M input) for tasks that Sonnet 5 ($2/M input) can handle is like using a sledgehammer to crack a nut, then paying for the sledgehammer rental by the hour. It's overkill, expensive. Understand the input rate of your chosen model and align it with the value of the task.

The input rate is the silent killer of your AI budget, dictating the cost of every piece of context your agent re-reads.