LLM cost optimisation, with code quality checked at each step

A coding agent bill combines several kinds of cost. Each of Claude Code, OpenAI Codex, OpenCode, Cursor and GitHub Copilot can be bought as seats or subscriptions, as metered requests to a model or gateway, or as a combination. We break your spend down from the invoices, not from list prices. We also look at the usage limits that come with subscriptions, because being throttled, paying for overage or waiting on blocked work has a price as well.

• Seats and subscriptions
• Metered requests to models and gateways, tokens included
• Capacity you have reserved or committed to
• Runs that were retried or abandoned

A smaller token bill is worth little if more changes are rejected. We add seats, subscriptions and metered requests together with rework, reverts and review findings, and compare what each accepted change costs.

Freed capacity turns into cash only when the invoice changes. With seats or reserved capacity, that happens when a contract or an allocation changes. You learn which savings will reach an invoice, and from which month.

Seats and metered usage reward different habits. With a seat, one session a day costs as much as twenty. With metered billing, the cost rises with the size of the context, the number of retries and the runs in parallel. Prompt caching, where both the provider and the tool offer it, makes repeated context cheaper, for example instruction files or large source files sent with every request.

Before we change a plan or a default, we look at how your teams actually use their tools. Putting heavy users on metered billing, or occasional users on seats, raises the bill instead of lowering it.

Your platform team works on this with us, and we change only one model choice at a time.

• A catalogue of approved models, with someone named as its owner
• Default models and routing rules for each type of task, based on results from your own work
• Attribution by team at the gateway, matched against the providers' invoices
• Budgets and caps tested under parallel and streamed requests, since a limit that holds for one call can fail under load
• A written process for teams that need a higher limit
• Fallbacks for when a provider is slow or down, rehearsed in drills

OpenCode keeps the provider and the model in versioned configuration. Your contracts still decide the bill, so we check the configured endpoints against your invoices.

When a workflow adds a second model for review or runs agents in parallel, the same repository context is sent several times, and every retry sends it once more. We count these requests per accepted change and add the review time the workflow costs or saves your engineers. How we attribute the cost depends on how it is billed: by provider or gateway invoices for a tool such as Kilo, by comparing seats with actual metering, and by whatever your plan meters for cloud agent work, such as Devin sessions. Paying more per change can still make a workflow worth it.

Setting up Kilo · Handing tasks to Devin · Using pstack in Cursor

• Spend for each team, tool and task
• The cost of an accepted change before and after every change to the setup
• Acceptance and rework next to cost, so a saving is reported only where quality held