The Token Bill Is Mostly Reprocessed Context, and Caching Is the Largest Lever

Across enterprise Claude Code deployments, Anthropic reports the average cost is around $13 per developer per active day and $150 to $250 per developer per month, with 90% of users staying under $30 per active day. That is the number to budget against, and most of it is not the model being expensive. It is context being sent again and again: tool schemas the agent reads before it reads your request, files it re-opens because the task was vague, and tool results that flow through the window twice.

The bill is driven by reprocessed context, not clever prompting

Anthropic's code execution with MCP names the two cost drivers directly. Tool definitions overload the window: an agent connected to thousands of tools can "process hundreds of thousands of tokens before reading a request." Intermediate tool results compound it: a long transcript passed through context twice adds "an additional 50,000 tokens." Loading those same tool definitions on demand as code instead of upfront took one tool-heavy scenario "from 150,000 tokens to 2,000 tokens, a time and cost saving of 98.7%." Cloudflare reports the same pattern under the name Code Mode. Treat the 98.7% as the ceiling of one scenario, not an average: the saving scales with how many tools and how large the results are.

The other multiplier is re-reading. Anthropic's own guidance notes that "vague requests like 'improve this codebase' trigger broad scanning," and that a single "go to definition" replaces a grep followed by reading several candidate files. A scoped task in a verifiable environment is also the cheaper task, because the agent stops hunting.

A cache read costs about a tenth of fresh input

Prompt caching is the largest single lever because the read discount is steep. On Claude, a cache hit costs 10% of the standard input price; the write costs 1.25x for the 5-minute window and 2x for the 1-hour window, so caching "pays off after just one cache read for the 5-minute duration." The pattern is not vendor-specific: OpenAI prices cached input at 0.1x standard on the GPT-5 line, and Gemini context caching runs near 0.1x of input but adds a separate per-hour storage fee the others do not charge. The catch is that the discount only applies to a stable prefix. Any byte change near the front (a timestamp in the system prompt, a reordered tool list) invalidates everything after it, and no vendor publishes what fraction of a real coding session's input actually hits cache, which is what determines realized savings. The code-execution and tool design that keeps the prefix stable is the same design that keeps the cache warm.

Match the model tier to the task, not the org to the flagship

Per-token prices as of June 2026 cluster by tier. On Claude: Opus 4.8 at $5 / $25 per million input/output, Sonnet 4.6 at $3 / $15, Haiku 4.5 at $1 / $5. Output runs four to eight times input across every vendor, so verbose agents are expensive agents. Anthropic's own routing advice is to reserve the flagship: "Sonnet handles most coding tasks well and costs less than Opus. Reserve Opus for complex architectural decisions. For simple subagent tasks, specify model: haiku." Analyst and vendor claims that tiered routing cuts cost 40% to 80% (one, another) are marketing figures with no neutral measurement on merged-PR workloads; treat the direction as sound and the magnitude as unproven. One hidden cost rides under flat per-token rates: Anthropic notes the newer tokenizer can use up to 35% more tokens for the same fixed text, so a model that looks same-priced can cost more per task.

Multi-agent topologies multiply the bill and pay off only at high task value

Fanning work across subagents is the fastest way to multiply spend. Anthropic measured that multi-agent systems use about 15x more tokens than chat, and reports separately that agent teams use roughly 7x more tokens than standard sessions because each teammate holds its own context window. The multiplier can pay: in one eval a lead-plus-subagent system "outperformed single-agent Claude Opus 4 by 90.2%," and "token usage by itself explains 80% of the variance" in score. But Anthropic gates it to "tasks where the value of the task is high enough" and notes coding parallelizes less cleanly than research. Cognition argues against the pattern entirely on coordination and cost grounds. If you run teams, the containment levers are small teams, cheaper models for the teammates, and shutting them down when their work is done.

Batch and off-peak cut offline work in half, but not the live loop

All three major vendors discount asynchronous work by 50%: Anthropic, OpenAI, and Gemini each halve input and output for batch processing, and Anthropic's stacks with prompt caching. The discount does not reach an interactive agent: batch is asynchronous, so it fits evals, triage passes, and offline migrations, not the turn-by-turn session a developer is watching.

Cost per merged change is still the metric nobody publishes

Companies publish PR volume, not PR cost. Stripe merges 1,300+ minion PRs a week and Uber's uReview runs across ~65,000 weekly diffs at "an order of magnitude less" than third-party tools, but neither attaches a dollar figure to a merged change. The only worked cost-per-PR numbers in circulation, such as Vantage's $31.58 per PR, are explicitly hypothetical. Gartner sizes the market at $9.8 to $11.0 billion annualized as of April 2026 and predicts that "by 2028, AI coding costs will overtake the average developer's salary" as pricing shifts from seats toward consumption. That is a planning assumption, not a measurement. This is the same gap measuring outcomes names: no company publishes unit economics, so instrument cost per change yourself before debating it. See hiring vs tokens for what the salary-overtake forecast does and does not imply.

Instrumenting cost per change across more than one vendor CLI runs into uneven reporting. One CLI ends each turn with a structured result carrying the turn's dollar cost and token usage; another's result omits cost entirely, leaving only a token-count estimate against the price sheet. A tracker that sums turn metadata across a mixed fleet holds real dollars for some sessions and estimates or blanks for others. Label each figure actual or estimated on the dashboard rather than averaging the two into one number, and put the normalization in the same per-vendor adapter that already translates each CLI's event shapes.