One Balance, 100 Engineers, Real Token Metering

01 · What Broke First When the Team Passed 100 Engineers

At ten engineers, AI usage is mostly tribal knowledge. A few people write strong prompts, everyone copies them, and the bill is small enough that nobody needs a spreadsheet. At one hundred engineers, the same habits stop scaling. Requests arrive from product, infrastructure, support and research at the same time, prompts drift apart, and the same question gets answered three different ways with three different settings. The first thing to break is rarely the model. It is coordination.

Treat the jump as an operating change rather than a tooling change. Decide who owns the shared prompt library, who approves new workflows, and who reviews generated code before it lands. Write those roles down before the headcount doubles again, because retrospective rules are far harder to enforce than rules set during rollout. Make one person accountable for prompt standards and one for spend, and make both names visible to the whole team.

The second thing to break is budget visibility. When everyone spends from their own account, you cannot tell whether a spike comes from a new feature, a stuck retry loop, or someone running an entire corpus through a long context window. Centralising usage is the moment when metering discipline becomes part of the job description instead of an afterthought.

02 · Route Work by Task Shape, Not by Personal Preference

Teams at this size often let every engineer pick a favourite model and keep it forever. That works until the workload mixes short classification bursts with deep reasoning chains and large retrieval passes. A better rule is to route by task shape. Fast, repetitive jobs that run thousands of times a day belong on a light, inexpensive option such as Claude Haiku 4.5, where latency and volume matter more than depth. Migrations, design reviews and security-sensitive reasoning belong on something like Claude Opus 5.5, where one careful answer is worth the extra tokens.

Write the routing rule into the same document as your prompt standards, and give each workflow a name: triage, refactor, docs, tests, evaluation. Named workflows make it possible to compare usage week over week and to change a default without surveying the entire company. They also give new hires a map of what the team actually does with models, which is more useful than any onboarding slide.

Keep a small escape hatch. Some tasks genuinely need a different model, and blocking exceptions entirely pushes usage into personal accounts you cannot see. Allow exceptions with a short written justification, then review them monthly to decide whether a pattern deserves its own workflow and its own routing rule.

03 · Cache Writes and Cache Reads Rewrite the Budget Math

Once usage is centralised, the bill stops being one number and becomes a shape. Calls are metered on real tokens, and caching changes that shape quickly, because cache writes and cache reads are priced separately from fresh input. A long system prompt or a large retrieval context that you regenerate from scratch on every request will cost more than the same content served from cache, and the difference compounds at team scale.

The practical habit is to keep stable content at the front of your context and variable content at the end. Shared tool descriptions, policy text, repository maps and standing instructions should not be rebuilt per call. Teams that make this a rule in their prompt libraries cut token consumption without touching model quality, and they get more predictable response times as a side effect.

Watch the two cache counters separately. A healthy pattern shows a steady share of cache reads alongside stable overall spend, which means your context structure is holding. A rising share of cache writes usually signals a rotating header, a timestamp or a per-user preamble that silently invalidates the cache on every call. Fixing one such header can matter more than any prompt rewrite that quarter.

This is exactly where a pay-as-you-go relay earns its place. Quanta meters every call on real tokens with cache writes and cache reads broken out, so a team lead can tell the difference between a caching regression and a genuine increase in workload instead of guessing from one blended total.

04 · Guardrails and Rollout Waves Before the Next Hundred

Scaling usage is as much about what the model is allowed to touch as about what it can do. Decide early which repositories, customer records and secrets are out of scope for prompts, and put that answer in the onboarding checklist rather than in a chat thread that scrolls away. Assume that anything pasted into a prompt will be reviewed by someone else later, and coach accordingly.

Review generated code the way you review code from a new hire: automated checks first, human sign-off for anything touching permissions, billing or data retention. At a hundred people, consistency beats individual cleverness. One team lead approving prompt library changes is far cheaper than three incident reviews, and it keeps the shared defaults trustworthy.

Roll out in waves instead of all at once. Start with one workflow, one team and one measurement window, then add teams once the numbers look sensible. Every wave should answer three questions: which workflow ran, how many tokens it consumed, and whether the cache behaved as expected. Those answers give you a baseline before the next group arrives and makes later comparisons meaningful.

Keep the model layer flexible too. A single balance covering many model families lets you move a workflow between options as latency, quality and cost shift, without renegotiating access for every team each time someone wants to experiment.

05 · The Weekly Numbers Worth Reviewing in Standup

A short weekly review keeps scaling honest. Track four things: tokens per active engineer, the split between cache writes and cache reads, usage by named workflow, and the number of exceptions to your routing rule. None of this needs a dashboard at the start; a shared sheet is enough until the patterns are clear and someone asks for automation.

The most useful signal is usually the outlier rather than the average. One workflow consuming several times its share often means a retry loop, an oversized context, or a prompt that quietly grew a documentation dump. Fixing a single workflow like that can pay for the entire review habit and restore trust in the numbers.

The second signal is adoption quality. Count how many engineers use the shared prompt library versus their own private copies. Private copies are not a failure, but they are a sign that the shared version is too rigid or too slow to change. Treat that as a product problem for your internal tooling, and ask the copycats what they were trying to fix.

Finally, keep the commercial side tidy as well. Quanta is built for heavy, sustained workloads and works with Claude Code, Codex, Gemini CLI and any OpenAI SDK, so your existing tools keep working as the team grows; its referral programme also credits you with 19% of the top-ups made by teams you bring along. If your next step is moving from scattered accounts to one metered balance, open an account and migrate a single workflow first, then measure it for a week before you move the rest.