Top 5 Ways to Stop Wasting Claude Code Tokens (2026 Guide)

Everyone is asking how to get access to the newest frontier model. Wrong question. The model is not the moat; the process you build around it is. Your usage limit is a budget, and routing decides whether it buys you four days of work or one. Here are the five rules we run at Summone, tested live on real jobs, with the receipts.

~70%
cheaper than running everything on the frontier model, in our live architect-and-workers test
Summone field test, July 2026
40%
of the price: Sonnet now matches last month's flagship Opus intelligence
Artificial Analysis, July 2026
Worse
is what maximum thinking effort scored on a mid-tier model against the hardest benchmark problems
FrontierCode benchmark, 2026
2
worker errors caught by the architect's verify pass on one real job: drafts, not facts, until checked
Summone field test, July 2026

The problem: one expensive model, default settings, everything

Most companies we walk into run every task, from renaming a variable to designing a system, on the most expensive model at default settings. That is the whole failure mode. Frontier models cost several times more per token than mid-tier ones, agentic sessions multiply token use, and the meter runs whether the task needed the intelligence or not. The five rules below fix it in order of leverage.

1. Extract the method into a skill (you keep the process, not the intelligence)

The biggest waste is not the price per token; it is using an expensive model to supply discipline a cheap model could follow if someone wrote it down. So write it down. We distilled the working discipline of a frontier model into a reusable Claude skill: five gates, in order, for any hard task.

  1. Scope. Define done, and how you will check it, before touching anything.
  2. Evidence. Open the real file or data. Never design from memory.
  3. Attack. Try to break your own answer before committing to it.
  4. Verify."It ran" is not verification. Check the actual output.
  5. Report. Lead with the answer. Separate verified from assumed.

Load that skill into Sonnet or Opus and the output gets noticeably sharper. You cannot transfer the intelligence between models. You can transfer the method, and the method is most of what you were paying for on routine work. (New to skills? Our guides to the top Claude skills for UI/UX engineers and for DevOps engineers cover how they work and how to install them safely.)

2. Route every task: the right model, not the best model

A routing table is one decision made once, instead of the wrong decision made fifty times a day. Ours, as of July 2026:

The routing table

Frontier model (Fable 5): planning, orchestration and review of high-stakes work. Never the workhorse.
Opus, high effort: the architect. Design decisions and hard debugging only.
Sonnet, medium effort:the daily workhorse. It now matches last month's Opus at 40% of the price.
Haiku: mechanical scans, admin, classification. Twice the speed, a fifth of the cost.

The Artificial Analysis price-to-intelligence data makes this blunt: Sonnet sits alone in the most attractive quadrant, and every older Opus is now dominated, meaning the same price buys less intelligence. If your default model is anything other than the mid-tier workhorse, you are paying flagship prices for commodity work.

3. Escalate the model, not the effort dial

The counterintuitive finding worth stealing: on the hardest problems in the FrontierCode benchmark, cranking a mid-tier model to maximum thinking effort made it worse, not better. Measured overthinking, not vibes. Extra effort only converted to accuracy on the frontier model. Effort is also a second cost dial: higher settings multiply both token burn and latency.

So the rule is simple. When a task is genuinely hard, escalate the model. Do not max the effort dial on a cheaper model and hope. The tell that you have it backwards: if you keep being tempted to push the mid-tier model's effort higher, that is the signal the task belongs a tier up.

4. Frontier as architect, cheap models as workers

We tested the whole setup live on a real job. The frontier model scoped the work and dispatched two subagents: a Sonnet worker and a Haiku scout. Then it verified their output. Both workers did good work. Both also made small errors that the verification pass caught. The run came out roughly 70% cheaper than doing everything on the frontier model, with no loss in the shipped result.

The lesson

Worker output is a draft, not a fact, until the orchestrator checks it. The savings come from the cheap models doing the volume; the quality comes from the expensive model doing the checking. Skip the verify pass and you have not saved money, you have deferred the cost to production.

5. Treat your usage limit as a budget

If you are on a subscription rather than API billing, all of this still applies, and arguably more. A weekly usage limit is a budget whether or not a card is attached. Routing decides whether it buys you four days of work or one. Spend the frontier allowance on frontier-worthy work: planning, review, the decisions that are expensive to get wrong. Let the workhorse models eat the volume.

A routing setup like this typically cuts spend by well over half with no quality loss on routine work. The models will keep changing; the prices in the table above will be stale in a quarter. The process is yours to keep.

The short version

Write the discipline down as a skill so cheap models run a frontier process. Route every task to the cheapest model that can do it. When a task is hard, escalate the model, not the effort dial. Let the frontier model plan and verify while workers do the volume. And spend your usage limit like it is money, because it is. Most companies run everything on one expensive model at default settings. Don't be most companies.

Sources

  1. Artificial Analysis (July 2026). Anthropic model comparison: intelligence vs price. artificialanalysis.ai/providers/anthropic
  2. FrontierCode benchmark (2026). Accuracy vs cost per task across effort settings, hardest-50 diamond subset.
  3. Anthropic. Claude Code skills documentation. code.claude.com/docs/en/skills
  4. Summone field test (July 2026): live architect-and-workers run, frontier orchestrator with Sonnet worker and Haiku scout; cost and error figures observed on a real client job.

Saving Claude Code Tokens: Frequently Asked Questions

Almost always because every task, from renaming a file to designing an architecture, runs on the most expensive model at default settings. Frontier models cost several times more per token than mid-tier ones, and long agentic sessions multiply that. Routing routine work to a cheaper model and reserving the frontier model for planning and review is the single biggest lever, and it typically cuts spend by well over half with no quality loss on routine work.

Sonnet on medium effort. As of mid-2026 it matches last month's flagship Opus intelligence at roughly 40% of the price (Artificial Analysis data), which puts it alone in the price-to-intelligence sweet spot. Save Opus for genuine architecture decisions and hard debugging, and Haiku for mechanical scans, admin and classification, where it is about twice the speed at a fifth of the cost.

Not reliably, and sometimes the opposite. On the FrontierCode benchmark's hardest problems, pushing a mid-tier model to maximum effort scored WORSE than a lower setting: measured overthinking, not vibes. Extra effort only converted to accuracy on the frontier model. When a task is genuinely hard, escalate the model, don't max the effort dial and hope.

Use the /model command in a session to switch between available models, or set a default in your settings. Subagents can be dispatched on cheaper models than the main session, which is how the architect-and-workers pattern is wired up in practice.

A working discipline written down as a reusable Claude skill, so cheaper models run the process a frontier model follows naturally: Scope (define done and how you'll check it before touching anything), Evidence (open the real file or data, never design from memory), Attack (try to break your own answer before committing), Verify ("it ran" is not verification, check the actual output), and Report (lead with the answer, separate verified from assumed). Loading it into Sonnet or Opus makes the output noticeably sharper without frontier prices.

Yes, arguably more. A weekly usage limit is a budget whether or not a card is attached. Routing decides whether that limit buys you four days of work or one. Running everything on the most expensive model at default settings is how a Monday allowance dies on Tuesday.

Want this routing setup running in your team?

We build exactly this for SMEs: model routing, reusable skills and verified agentic workflows, through hands-on Claude Code training and AI consulting. Book a free 45-minute call and we'll map where your token budget is leaking.

Book a free call →