Price AI work by the cost of a completed task, not by the price of a token. Models use very different numbers of tokens for the same work. Reasoning settings, retries, tool calls, and human review all change the real cost. A model with a higher token price can be cheaper per task, because it uses fewer tokens and fails less. So the unit to budget, compare, and negotiate on is the cost per successfully completed task at your own quality bar.

Why Token Price Misleads

Four effects separate the token price from the cost of the work.

  1. Reasoning tokens are billed. OpenAI's reasoning guide says reasoning tokens are "billed as output tokens" even though you do not see them. Anthropic and Google bill thinking tokens the same way.
  2. Token counts vary by model and version. Artificial Analysis reported that Claude Opus 4.6 used 58 million output tokens to run its benchmark index, about twice Opus 4.5. The token price was identical, so the cost of the same work doubled.
  3. Settings change token use several times over. Anthropic's effort documentation says the effort parameter affects all tokens in the response, including tool calls. Researchers have measured reasoning models using 1,953% more tokens than conventional models to answer "2+3" (Chen et al.).
  4. Tools and retries add tokens. Every tool definition is sent as input on every request. Anthropic's pricing page lists about 4,500 extra input tokens per request for its computer-use tools. A failed attempt costs the same as a successful one.

The Evidence: Cost per Task Reorders the Models

The ARC Prize publishes the retail cost per task for each model and setting on its ARC-AGI-2 benchmark. The data shows how far cost per task and price per token can diverge.

Score vs cost per task on ARC-AGI-2, selected models and effort settingsClaude Opus 5.5Claude Sonnet 5.5Other modelsScore (%)60%70%80%90%100%$0.01$0.1$1$10Retail cost per task (US$, log scale)Low: Retail cost per task (US$, log scale) $0.24, Score (%) 70%Medium: Retail cost per task (US$, log scale) $0.34, Score (%) 88%High: Retail cost per task (US$, log scale) $0.41, Score (%) 93%Retail cost per task (US$, log scale) $0.67, Score (%) 93%Max: Retail cost per task (US$, log scale) $1.85, Score (%) 92%Sonnet High: Retail cost per task (US$, log scale) $0.28, Score (%) 67%Sonnet XHigh: Retail cost per task (US$, log scale) $0.55, Score (%) 84%Haiku 5.5: Retail cost per task (US$, log scale) $0.09, Score (%) 61%Opus 5: Retail cost per task (US$, log scale) $1.45, Score (%) 88%GPT-6.1 Sol: Retail cost per task (US$, log scale) $0.14, Score (%) 92%Gemini 4 Argon: Retail cost per task (US$, log scale) $0.33, Score (%) 91%LowMediumHighMaxSonnet HighSonnet XHighHaiku 5.5Opus 5GPT-6.1 SolSingle runs at retail prices. ARC-AGI-2 is a puzzle benchmark,not a business workload. Use it to see the method.
CHARTScore vs cost per task on ARC-AGI-2, selected models and effort settings. Source: ARC Prize Foundation, ARC-AGI-2 leaderboard, data as of October 28, 2026; chart by OpenNash.

Using the list prices on Anthropic's pricing page:

  • Pricier per token, cheaper per task. Opus 5.5 costs twice as much per output token as Sonnet 5.5. Yet Opus 5.5 at high effort scored 93.3% at $0.41 per task. Sonnet 5.5 at its highest setting scored 83.9% at $0.55.
  • A 40x token gap becomes a 3x task gap. Haiku 5.5 is 40 times cheaper per output token than Opus 5.5. Per task, it was about 2.8 times cheaper.
  • Settings matter as much as the model. The same Opus 5.5 ranged from $0.24 to $1.85 per task, a 7.7x spread. The most expensive setting scored lower than a cheaper one.
  • A new model generation beat a price cut. From Opus 5 to Opus 5.5, cost per task fell 72% while the list price fell 20%.

Princeton's Holistic Agent Leaderboard found the same pattern across agent benchmarks. Higher reasoning effort reduced accuracy in the majority of its runs.

Vendors now measure their own progress this way. Anthropic's S-1 filing says Claude Opus 5.5 "reduced customer costs per task by approximately 40%" compared to Opus 5, from a lower price and better efficiency combined.

The Formula: Cost per Completed Task

Use this formula for each workflow:

Cost per completed task = (token cost of all attempts + tool fees + infrastructure) รท success rate + human review cost + rework cost

Input Where to get it
Token prices, cache and batch discounts Vendor pricing pages
Tokens per attempt, including reasoning and tools Your API usage logs
Attempts per task (retries, fallbacks, agent loops) Your traces
Success rate at your quality bar Your own eval set, graded on your acceptance rules
Review minutes and share of tasks reviewed Your QA sampling and time studies
Loaded labor rate Your HR data, or the BLS employer cost average

A worked example shows why the last two lines matter. Take the Opus 5.5 result above: $0.408 per attempt at 93.3% success, or $0.437 per completed task. Now add 2 minutes of human review on 20% of tasks, at the BLS private-industry employer cost of $46.89 per hour. That adds $0.31. Review is about 40% of the total.

Worked example: what one completed task costsModel tokens, adjusted for failed attempts ($0.44)Human review, 2 minutes on 20% of tasks ($0.31)Model tokens, adjusted for failed attempts ($0.44): 0 (58%)58%Human review, 2 minutes on 20% of tasks ($0.31): 0 (42%)42%Model cost $0.408 per attempt at 93.3% success. Review at $46.89per hour, the BLS private-industry employer cost (June 2026).
CHARTWorked example: what one completed task costs. Source: ARC Prize leaderboard; US Bureau of Labor Statistics, Employer Costs for Employee Compensation; chart by OpenNash.

In many workflows, the cheapest way to cut cost is to cut review time with better checks, not to switch to a cheaper model. Our post on approval gates and exception routing covers how to send fewer cases to people.

Discounts Change the Answer

Two discounts apply to most production workloads, and both are in the vendors' own pricing pages.

Discount Anthropic OpenAI Google
Batch processing (results within hours) 50% off 50% off 50% off
Cached input 2.5% to 10% of input price, by model 5% to 10% of input price, by model 5% to 10% of input price

A workflow that reuses a long system prompt or document set pays a small fraction of the listed input price. A workflow that can wait for results pays half. See what prompt caching is and when it saves money.

Compare With Outcome Pricing

Some vendors now sell completed outcomes instead of tokens.

Vendor Unit Price
Intercom Fin Per outcome $0.99
Zendesk AI agents Per automated resolution $1.50 committed, $2.00 pay as you go
Salesforce Agentforce Per conversation $2

Outcome prices are easy to budget. Compare them with your own cost per completed task, and read each vendor's definition of an outcome. Intercom, for example, counts a conversation as resolved when the customer confirms the answer or leaves without asking for more help. Audit a sample against your own quality bar.

Re-Price Every Quarter

The cost of a fixed level of capability falls fast. Epoch AI measured about 47% per quarter from 2023 to 2026. A workflow priced a year ago is probably overpaying.

Prices do not only fall, though. MIT researchers found that the price of running frontier models is rising 3x to 18x per year, because models are larger and reason longer (Gundlach and coauthors). Introductory list prices also end. Google's Gemini pricing page lists scheduled increases for some models. Anthropic's effort documentation tells users to re-run effort tests on their evals with each new model.

Routing captures the fall. Ramp's engineering team reports that a router that prices in failure risk "cut our AI spend by more than 25%" while reducing errors. Research systems such as FrugalGPT and RouteLLM show similar savings from sending easy requests to cheaper models.

A Quarterly Cost Review

  1. Define the task and the quality bar before you price anything.
  2. Run an effort sweep. Test each candidate model at each setting on your eval set.
  3. Calculate cost per completed task with the formula above, including review.
  4. Tier the models. Assign each task the cheapest setting that passes its bar.
  5. Turn on caching and batch wherever the workflow allows.
  6. Compare with outcome pricing where a vendor sells the same result.
  7. Repeat every quarter and on every major model release. See know your token mix for how to track usage by workflow.

Know Your Cost per Task With OpenNash

OpenNash helps your team measure and cut the cost of AI work. Our forward-deployed engineers:

  1. Build eval sets from your own work, with clear pass and fail rules.
  2. Measure cost per completed task for each workflow, including review time.
  3. Run effort and model sweeps and route each task to the cheapest setting that passes.
  4. Turn on caching, batch, and budgets, with cost reports per team and per workflow.
  5. Repeat the review every quarter, so falling prices reach your costs.

Book a call with an OpenNash forward-deployed engineer or email us at hello@opennash.com. In the first session, we review your AI spend and your top use cases. Then we show you which ones deliver the highest business value and the fastest ROI.