# Two Tasks Out of Twenty Eat 43% of Your AI Budget: How Reasoning Effort Optimization and Retry Cut Costs by 40%

> 2 of 20 agent tasks drive 43% of AI spend. See how reasoning effort optimization and low-to-high retry cut production AI agent costs by 40% without losing quality.

- Published: 2026-10-20
- Author: Master Mind
- Canonical: https://aimaster.fi/en/artikkelit/kaksi-tehtavaa-kahdestakymmenesta-nielee-43-prosenttia-tekoalybudjetista-paattelysyvyyden-optimointi

In a 20-problem multi-step agent benchmark (WideSearch), the two most expensive tasks accounted for 43% of total spend while the cheapest half took only 10% (Anthropic, 2026). Reasoning effort optimization — adjusting how deeply a language model thinks and calls tools on each request — is the lever that stops a handful of runaway runs from eating your AI budget. When verifiable tasks run at low effort first and only failed checks retry at high effort, total cost drops by about 40% at the same pass rate (Anthropic, 2026).

Most leadership teams still compare language models by price per million tokens. That spreadsheet looks tidy in a slide deck, right up until the first production invoice lands on the CFO's desk. In production, you do not pay for tokens in isolation — you pay for completed business tasks.

## Why Do Two Tasks Out of Twenty Consume 43% of Your AI Agent Budget?

In multi-step AI agent workflows, costs follow a steep power law rather than an even average. In Anthropic's 2026 WideSearch measurements across 20 agent problems, the top 10% of tasks (2 of 20) drove 43% of the entire bill, while the cheapest 10 tasks combined accounted for just 10%. Averages hide the exact point where your margin leaks.

Why does a single task suddenly cost ten or twenty times more than a normal run? An agent loop re-sends the entire growing conversation — system prompt, tool definitions, and every prior step — on every turn. A 40-turn task sends its first turn 40 times, so unoptimized cost grows roughly with the square of the turn count (Anthropic, 2026).

In a growth company's daily operations, this tail spend appears when an order-validation or invoice-matching agent hits an ambiguous PDF attachment or a conflicting ERP row. Instead of escalating cleanly, a model left on high default reasoning keeps generating long internal thinking blocks and repeated tool searches. The token meter spins for minutes on a single ticket.

## What Is Reasoning Effort Optimization and How Does the Effort Parameter Work?

Reasoning effort optimization means setting a language model's work depth per task using the `output_config.effort` parameter (`low`, `medium`, `high`, `xhigh`, or `max`) instead of running every request at maximum capacity (Anthropic, 2026). The effort setting governs every output token the model produces: internal thinking blocks, tool calls, function arguments, and final text responses.

Most language models default to `high` effort (while Claude Opus 5.5 and Claude Haiku 5.5 default to `medium`), spending as many tokens as needed for thorough answers. At lower effort levels, the model combines operations into fewer tool calls, skips unnecessary preambles, and avoids overthinking routine lookups (Anthropic, 2026).

## What Happens When You Compare Models by Cost Per Completed Task Instead of Token Price?

Comparing models by unit token price misleads buyers because a stronger or better-calibrated model finishes the job in fewer turns. On Anthropic's SWE-bench Pro subset (2026), Claude Opus 5.5 carries twice the per-token price of Claude Sonnet 5.5, yet at `low` effort it solved 87.4% of tasks for $0.12 per solved task — matching Sonnet 5.5 at its `high` default (87.6% solved at $0.11). Stepping up to the frontier model Claude Fable 5.1 at default `high` effort solved 92.3% at $1.19 per task, whereas Claude Opus 5.5 at `medium` solved 92.8% for $0.22 — less than a fifth of the cost (Anthropic, 2026).

| Strategy / Configuration | Typical Use Case in a Growth Company | Measured Cost & Quality Impact (Anthropic, 2026) |
| --- | --- | --- |
| Everything on Frontier Model at High Effort | Unoptimized default setup where every ticket runs at full depth | Baseline cost ($1.19 / solved task on SWE-bench Pro subset with Fable 5.1) |
| Right-Sized Model at Medium Effort (Opus 5.5) | Multi-step ERP, CRM, and procurement agent workflows | 92.8% solved at $0.22 / task — ~80% lower cost per solved task than Fable 5.1 default |
| Low Effort First + Retry Failures at High Effort | Tasks with verifiable outputs (JSON schema, SQL/code check, order rules) | ~40% lower total cost than running everything at high effort, at the same pass rate |
| Prompt Audit on Legacy System Prompts | Removing old "verify twice" and 6-step scripts after model upgrades | Lower cost per ticket and higher accuracy by eliminating redundant tool calls |

## How Does Low Effort Plus High-Effort Retry Cut Agent Costs by 40%?

Whenever an AI agent's output can be verified automatically — against a schema, a calculation rule, or a test suite — running every request at `low` effort first and retrying only failed attempts at `high` effort cuts total cost by about 40% at the same pass rate compared with running everything at `high` (Anthropic, 2026). Most routine business inputs pass on the first lightweight attempt.

Where can a growth company use this pattern immediately? Any process with a deterministic check qualifies:

- Purchase invoice and order matching: A lightweight check verifies whether line items sum to the invoice total and match the purchase order in the ERP. Only mismatches trigger a high-effort run.
- Customer support and claims triage: 80–90% of incoming tickets match standard product categories at `low` effort. Complex multi-part claims escalate to `high` effort only when confidence or rule checks fail.
- Database queries and report generation: If a generated query fails validation or returns an empty set, the orchestrator re-runs that single step with deeper reasoning.

This requires reliable automated checks. That is why we pair effort routing with [AI agent evaluations and QA in production](https://aimaster.fi/en/artikkelit/75-prosentin-tarkkuus-putoaa-42-prosenttiin-tekoalyagenttien-evaluointi-ja-laadunvarmistus), so your platform knows within milliseconds whether a low-effort output passed or needs escalation.

## How Do You Change Reasoning Effort Without Breaking Your Prompt Cache?

Changing the top-level `effort` parameter mid-conversation invalidates the prompt cache, forcing the model to re-write the entire context prefix at 1.25x input price instead of reading it at 0.025x to 0.1x (Anthropic, 2026). On a 100,000-token agent session, a single broken cache turn costs $0.50 instead of $0.02 on Claude Opus 5.5 (25 times more) and $1.25 instead of $0.025 on Claude Fable 5.1 (50 times more).

In production traffic, well-built agent loops read a median 84% of their input tokens from the cache, and the top 10% read 94% or more (Anthropic, 2026). If you naively flip top-level effort settings between turns in the same conversation, you wipe out the savings gained from [prompt caching in AI agent cost optimization](https://aimaster.fi/en/artikkelit/prompt-caching-leikkaa-ai-agentin-syotekuluista-90-prosenttia-tuotannossa).

To keep the cache intact, follow two architectural rules. First, run the high-effort retry as a clean separate call on the verifier boundary, or use per-message effort controls (`output_config.effort` inside a mid-conversation system message on supported models like Claude Opus 5.5, Sonnet 5.5, and Fable 5.1), which preserve the cached prefix (Anthropic, 2026). Second, audit your prompts whenever you upgrade models: removing legacy instructions like "verify twice" or mandatory 6-step scripts stops newer models from making four tool calls on a simple "where is my package" query.

## How Do We Build Cost-Controlled AI Agents on the AIMASTER Path?

AIMASTER is a Finnish AI partner that puts AI into production for growth companies — not as slide decks, but as working systems. We serve Finnish companies with €1M–€1B in revenue, and we build reasoning effort optimization directly into our three-stage productized path:

- **[Master Plan](https://aimaster.fi/tuotteet/master-plan)** is an AI strategy sprint that maps where AI generates the most value for your company — measured in euros. We calculate the target cost per completed task for your key processes and identify where tail spend risks hide before writing a line of code.
- **[Master OS](https://aimaster.fi/tuotteet/master-layer)** is the company's AI Operating System: a platform that safely connects the company's existing systems (CRM, ERP, documents), data, and agents for AI use. In Master OS, we configure session budgets, prompt caching, and the automated verifiers that route tasks between `low` and `high` effort without breaking cache prefixes.
- **[Master Mind](https://aimaster.fi/tuotteet/master-mind)** is a suite of AI agents that runs on top of Master OS and handles business processes autonomously. Custom AI solutions are delivered with an agile sprint model: one sprint is 3 development days. In our ongoing partnership, we also run prompt audits during model upgrades so old instructions never inflate your bill.

When we built the public [Aini AI assistant for Jaajo Linnonmaa](https://aimaster.fi/artikkelit/aini-tekoalyapulainen-jaajo-linnonmaalle) and automated marketing workflows for digital marketing agency Tagomo, the goal was never to burn maximum reasoning tokens on every routine message. Fast, bounded execution on routine turns and deep reasoning only where the task demands it is what separates a production system from an expensive prototype.

Want to see which processes in your company can run 40% leaner with automated effort routing and where AI delivers the highest euro return? Start with a [Free Master Mind Analysis](https://aimaster.fi/analyysi).

## Frequently asked questions

### What is reasoning effort optimization in AI agents?

Reasoning effort optimization controls how many tokens a language model spends on internal thinking, tool calls, and explanations before returning an answer. By adjusting the effort parameter (low, medium, high, xhigh, or max) per task, your company stops paying maximum reasoning costs on routine checks.

### Why do a few agent runs consume most of the AI budget?

In Anthropic's 2026 WideSearch benchmark of 20 multi-step agent problems, the two most expensive tasks accounted for 43% of total spend while the cheapest half took only 10%. A single complex task triggers dozens of tool calls and re-sends the growing conversation history on every turn.

### How much does the low-to-high effort retry pattern save?

Running verifiable tasks at low effort first and automatically retrying only failed outputs at high effort cuts total cost by about 40% at the same pass rate compared with running every task at high effort (Anthropic, 2026).

### Does changing the effort level break prompt caching?

Changing the top-level effort setting mid-session invalidates the prompt cache, which can make a 100,000-token prefix 12.5 to 50 times more expensive on that turn (Anthropic, 2026). Keep top-level effort constant within a cached session or use per-message effort controls on supported models.

### How does AIMASTER implement reasoning effort optimization?

In the Master Plan sprint, we identify which business processes have verifiable outputs and calculate their target cost per completed task in euros. In Master OS and Master Mind, we build the automated verifier and effort routing in 3-day sprints so your agents escalate reasoning depth only when needed.