# AI coding model costs: an illustrative token-price basket

Canonical: https://www.pointfive.co/blog/the-pointfive-coding-task-index

A transparent token-price example, with separate input and output rates and clear limits on what it says about real coding-task costs.

By: Dave Anderson

Published: 2026-07-06

[All articles](https://www.pointfive.co/blog) 

Dave Anderson  CMO, PointFive  July 6, 2026  3 min read

A fixed token basket can illustrate how model prices differ. It cannot establish what different models cost to complete the same coding task: they may produce different token counts, retries, and outcomes.

Correction, September 17, 2026: this article previously described an illustrative basket as a task benchmark and included an unsupported future Sonnet price change. The comparison below replaces those claims and the derived monthly engineer-cost charts.

## An illustrative token basket

Assume 200,000 uncached input tokens and 30,000 output tokens in total, spread across requests that meet the model's limits and standard pricing conditions. The calculation is 0.2 \* input rate + 0.03 \* output rate, with rates in USD per million tokens.

| Model | Input / million | Output / million | Basket cost |
| --- | --- | --- | --- |
| Claude Haiku 4.5 | $1 | $5 | $0.35 |
| Claude Sonnet 5 | $2 | $10 | $0.70 |
| Claude Opus 5 | $5 | $25 | $1.75 |
| Claude Opus 5.5 | $4 | $20 | $1.40 |

Source: [Anthropic standard API pricing](https://platform.claude.com/docs/en/about-claude/pricing) , checked September 23, 2026. This example excludes caching, batch discounts, tools, taxes, regional premiums, and negotiated terms. It is not an exhaustive model comparison.

For Sonnet 5, the input portion is $0.40 and output is $0.30. The result is $0.70 for this basket, not a measured price for a completed coding task.

## What the basket leaves out

Models use different tokenizers and may need different amounts of reasoning, tool use, and revision. Holding token counts constant does not hold useful work constant. The basket also says nothing about accuracy or whether a resulting change passes review.

A task benchmark would need published tasks, model versions, settings, repeated runs, token accounting, retries, and success criteria. Monthly engineer costs additionally depend on workload and utilization; multiplying this basket by an assumed task count would be a scenario, not observed field data.

## Use real task outcomes for decisions

Collect billed usage for representative work, record completion and quality, and compare total cost per successful outcome. TokenShift helps teams inspect supported coding-agent usage and evaluate optimization alongside governance and task outcomes.

---
Source: the public page above. Product screenshots and illustrative interfaces are examples, not live customer data.