LLM cost management is the practice of pricing, attributing and capping what your company spends on language models, and LLM cost optimization is lowering that spend per useful output without losing quality. This guide gives you four layers with one concrete practice each, a 10-lever optimization table, model choices by task, and a 30-day rollout plan.
All model rates are ATP list rates from the model pages, as of October 2026. Vendor pricing facts link to the vendor's own pages.
Layer 1: token economics
The practice: price the three request shapes that make up most of your traffic, per 1,000 requests, before you argue about the monthly total.
A support reply with 1,284 input and 412 output tokens on claude-sonnet-4-6 ($3 / $15 per 1M) costs 1,284 × 3 ÷ 1,000,000 + 412 × 15 ÷ 1,000,000 = $0.010032, or $10.03 per 1,000 requests. At 300,000 replies a month that is $3,009.60. Once you have that number per shape, every lever below has a dollar value. The full method, with five models side by side, is in LLM token cost: how to calculate it.
Layer 2: credits as the settlement unit
The practice: fund one monthly budget in credits and push it down the hierarchy, so finance reconciles one unit while engineering keeps model-level detail.
On ATP Token, 1 credit = USD 0.01. A $3,000 month is 300,000 credits, which is three Scale top-ups of USD 1,000 each. Credits flow organization → workspace → project, and each level shows Available, Received, Allocated and Consumed. Pay-as-you-go credits do not expire, and top-ups are non-refundable, so fund for the month you are planning rather than the year.
Layer 3: ownership through projects and keys
The practice: one project per service per environment, each with its own key.
Six services in staging and production means 12 projects and 12 keys. A key belongs to exactly one project and inherits its allowed models and credit balance, so the Usage page's breakdown by key reads directly as spend by owner. When a contractor leaves, revoking their project's key stops it immediately without breaking the other 11 services. The pattern is explained in one project, one key.
Layer 4: caps and a weekly review
The practice: give every project a fixed allocation, and look at usage once a week.
A project can only spend what it was allocated. A staging project with 5,000 credits has a $50 ceiling: when its balance runs out, calls return 402, and a project that ends up past its allocation is flagged In debt until topped up. The staging job that ran all weekend costs about $50.
For the review, ATP request logs keep 7 days of per-request rows (model, status, input and output tokens, request ID), so a weekly look catches anomalies while the detail still exists. Month-end reconciliation uses billing events, the 90-day ledger that is the billing source of truth.
10 levers for LLM cost optimization
Effects are either quoted from the vendor's pricing page or worked out with the support-reply example above (1,284 in / 412 out on claude-sonnet-4-6, $0.010032).
Levers 1 to 4 and 9 apply to any provider. Levers 5 and 6 are vendor pricing programs with their own rules, so price them per vendor. Levers 7, 8 and 10 are controls you configure in the ATP console.
Choosing models by task
Lever 1 usually saves the most, and it needs an evaluation first. Run 50 to 100 real prompts from each workload through two or three candidates and compare quality before moving traffic.
Token counts also differ between models. Anthropic notes the Claude 4.7+ tokenizer produces approximately 30% more tokens for the same text, so compare cost on your own prompts rather than on rate cards.
Agents: measure cost per completed task
A coding or research agent sends many requests per job, and each step can resend the full context. Anthropic reports that Claude Code averages around $13 per developer per active day and $150 to $250 per developer per month, and that agent teams use approximately 7× more tokens than standard sessions in plan mode (Claude Code costs).
Cost per request stays flat while cost per task climbs, so track both. Put agents in their own project with their own allocation, so a looping agent hits its own cap instead of draining production. For coding agents specifically, see the coding agents cost control checklist.
30-day rollout plan
- Week 1: list every key and vendor account in use, and the service behind each. Price the top three request shapes per 1,000 requests.
- Week 2: create the organization, workspaces and one project per service per environment. Move the three highest-spend services to project keys; the change is a base URL and key swap for OpenAI, Anthropic and Gemini SDKs.
- Week 3: allocate monthly credits per project, set allowed models, and move agents into separate projects. Apply levers 1 to 4 to the most expensive workload.
- Week 4: close the month from billing events, compare Consumed against Allocated per project, and set next month's allocations from that.
Metrics scoreboard
Setting this up in ATP Token
- Create a Team organization, then workspaces and projects in Resources.
- For each project, choose its allowed models and allocate credits. In a team organization, keys spend org-allocated credits, not a personal wallet.
- Create one
atp-key per project. The secret is shown once; the API keys page lists every key across workspaces and lets you revoke any of them. - Read the Usage page weekly by model and by key, and use request logs to drill into a spike within 7 days.
