Giving 40 Marketers a Claude Sonnet 5.5 Budget

01 · What forty marketers needed before the first token moved

The ticket landed like a fire alarm. A non-engineering department, forty strong, wanted model access by Monday. Not a sandbox, not a workshop, not a slide deck about artificial intelligence: real access to real models, for real work, with a real budget line. If you lead a team that runs heavy, sustained workloads, you already know the pattern. The demand does not arrive politely, and it does not wait for a platform roadmap.

The first job is not technical. It is translation. Non-engineering colleagues do not think in tokens, latency budgets or context windows; they think in deliverables. A campaign brief. A research digest. A hundred product descriptions rewritten to a house style. Your task is to convert those deliverables into a metered plan that survives contact with a finance review, and to do it before anyone has burned through a month of budget in an afternoon.

Start with three questions in the kickoff meeting. What work should the models take over? What must still be reviewed by a human? And who pays when the workload spikes? Write the answers down where everyone can see them. Most failed rollouts fail here, not in the plumbing: a department gets unrestricted access, usage multiplies, nobody can explain the invoice, and the tool is quietly taken away.

02 · Two models, one policy: a default tier and a heavy tier

Resist the buffet. Offering a long model menu to a new department produces thrash: people switch models mid-task to compare, costs multiply, and outputs become impossible to review consistently. Pick two, name them in the policy, and make everything else an exception with a named approver.

For the default tier, choose something fast and economical for the highest-volume work: drafting, summarising, tagging, first-pass rewriting, meeting-note cleanup. Claude Haiku 4.5 is the kind of model that fits that slot in a heavy workload, where the value comes from thousands of small calls rather than a few heroic ones. For the heavy tier, keep one stronger model for long documents, nuanced positioning, and anything that will be shown to a customer. Claude Sonnet 5.5 works well there, especially for tasks that mix a large pile of source material with careful synthesis.

• › Default tier: bulk and repetitive work, automatic, no approval needed.
• › Heavy tier: long-context and customer-facing work, allowed but logged.
• › Exceptions: anything else, written justification, team lead sign-off.

Document the swap rule too: when a task in the default tier produces three unusable drafts in a row, the writer escalates to the heavy tier rather than retrying blindly. Retries are the most common hidden cost in a non-engineering rollout.

03 · Token budgets, cache writes and cache reads, explained without jargon

Every call is metered on real tokens, and this is the part your department must understand before launch. A long prompt with a ten-page brief attached costs more than a short one. A week of re-reading the same brand guidelines in every request costs more than a setup that reuses them. Spell this out in plain language: input, output, and the two cache line items.

Cache writes happen when a stable block of content, such as house style rules or a standing brief, is stored for reuse. Cache reads happen when later calls pull that stored content instead of paying to send and process it again. The two are priced separately, so the rollout meeting should cover one habit: put anything repeated on every call into a stable, reusable block, and keep the changing part of the request short. This single practice is often the difference between a predictable monthly spend and a budget that doubles when a campaign goes live.

This is also where a relay earns its keep. Quanta is a pay-as-you-go API relay built for power users and teams with heavy, sustained workloads: one balance covers many models, and it works with the tools engineers and power users already run. When a department shares one balance instead of juggling separate accounts, budgeting becomes a single conversation. If you want the department's traffic on one metered account, open a Quanta account once and hand the team their keys.

04 · Keys, quotas and the approval chain that keeps spend predictable

Give each squad its own key, named for the squad and the purpose, never a shared department-wide credential. When a key is named, a bill becomes attributable, and attribution is the only honest budgeting tool you have. Rotate keys when people change roles, and retire keys for experiments that end.

Layer quotas on top. A soft limit triggers a warning to the squad lead; a hard limit stops automated work and requires a decision. Set the hard limit against a real number from finance, not a guess. If nobody can agree on a monthly figure, agree on a weekly one and review it every Friday for the first month.

Define the approval chain in advance: who can raise a limit, who approves a new model, who signs off on a new automated pipeline. Put the referral benefit somewhere sensible as well. People who refer others earn 19% of their top-ups, which matters when a rollout spreads from one department to three and someone has to fund the growth.

05 · Week-one rituals that turn curiosity into steady throughput

Adoption is a habits problem. Run a week-one rhythm: a kickoff that shows two live workflows end to end, a shared prompt library seeded with the department's own templates, and a short office hour where people bring half-finished work instead of abstract questions. Review output quality on real deliverables, not toy examples.

Then measure. Track cost per deliverable, not cost per user, and review it alongside the quality signal your department already trusts: revisions requested, drafts discarded, time to first usable version. Pair every automated workflow with a named human owner, and write down what the model is allowed to do with customer data, unpublished copy and anything legal will care about later.

Finally, plan the second month before the first ends. Expect usage to double when the novelty wears off and the workloads arrive. Expect the cache discipline to slip. Expect someone to ask for a third model and a fourth. If your policy names the default tier, the heavy tier, the exception path and the budget owner, those requests become small decisions instead of emergencies, and the rollout becomes infrastructure instead of an experiment.