# AI Cost Benchmark 2026: What a Completed Task Costs

> What does an AI task cost? Machine cost per completed task for 38 current models and 12 business scenarios, from list prices, as of 18 August 2026.

- Canonical URL: https://www.convios.com/en/insights/ki-kosten-benchmark-2026
- Language: English
- As of: 18 August 2026
- Publisher: Convios GmbH
- License: CC BY 4.0
- Scope: 38 current models, 12 business scenarios, 2 scenarios with documented solve rate

38 current models, 12 business scenarios, vendor list prices. The tables show the machine share of a completed task in euros and, where the calculator knows a solve rate, the cost per solved task.

## The range per scenario

Between the cheapest and the most expensive current model lies, for every one of the twelve scenarios, more than a factor of one hundred. The median shows where the field stands; the factor shows how much model choice matters.

| Scenario | Models | cheapest model | Median | most expensive model | Factor |
| --- | --- | --- | --- | --- | --- |
| Answering a support ticket | 38 | Ministral 3 3B (0.001 €) | 0.008 € | gpt-5.4-pro (0.23 €) | 462× |
| Reading an email and drafting the reply | 38 | Ministral 3 3B (0.000 €) | 0.005 € | gpt-5.4-pro (0.14 €) | 479× |
| Writing or revising a text | 38 | Ministral 3 3B (0.001 €) | 0.015 € | gpt-5.4-pro (0.51 €) | 636× |
| Sales email and proposal draft | 38 | Ministral 3 3B (0.001 €) | 0.023 € | gpt-5.4-pro (0.75 €) | 576× |
| Classification and routing | 38 | Ministral 3 3B (0.000 €) | 0.003 € | gpt-5.4-pro (0.06 €) | 318× |
| Extracting data from documents | 38 | Ministral 3 3B (0.002 €) | 0.034 € | gpt-5.4-pro (0.84 €) | 352× |
| Knowledge-base question | 38 | Ministral 3 3B (0.001 €) | 0.017 € | gpt-5.4-pro (0.47 €) | 394× |
| Research agent over documents | 38 | Ministral 3 3B (0.016 €) | 0.282 € | gpt-5.4-pro (13.10 €) | 814× |
| Writing or changing code | 38 | Ministral 3 3B (0.005 €) | 0.087 € | gpt-5.4-pro (3.27 €) | 605× |
| Working through a long document | 25 | gpt-5.4-nano (0.097 €) | 0.858 € | gpt-5.4-pro (28.47 €) | 295× |
| Meeting notes | 38 | Ministral 3 3B (0.002 €) | 0.029 € | gpt-5.4-pro (0.75 €) | 375× |
| Translation | 38 | Ministral 3 3B (0.001 €) | 0.013 € | gpt-5.4-pro (0.53 €) | 1,059× |

## The cheapest models per scenario

For each scenario the five cheapest models by machine cost, plus the cheapest model in each class. The input sizes are listed below each table; they are example sizes from the calculator presets, not measurements.

### Answering a support ticket

Input 3,700 tokens, output 400 tokens, one call per task, 2 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0005 € |
| Ministral 3 8B | Edge class | 0.0008 € |
| Qwen3.5 9B (Together) | Compact class | 0.0009 € |
| Mistral Small 4 | Compact class | 0.0010 € |
| Ministral 3 14B | Edge class | 0.0011 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0032 €); Qwen3.7-Plus (Together) (Workhorse class, 0.0022 €); Qwen3.5 9B (Together) (Compact class, 0.0009 €)

### Reading an email and drafting the reply

Input 2,000 tokens, output 300 tokens, one call per task, one minute of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0003 € |
| Ministral 3 8B | Edge class | 0.0004 € |
| Qwen3.5 9B (Together) | Compact class | 0.0005 € |
| Mistral Small 4 | Compact class | 0.0006 € |
| Ministral 3 14B | Edge class | 0.0006 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0019 €); Qwen3.7-Plus (Together) (Workhorse class, 0.0013 €); Qwen3.5 9B (Together) (Compact class, 0.0005 €)

### Writing or revising a text

Input 4,000 tokens, output 1,200 tokens, one call per task, 5 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0008 € |
| Ministral 3 8B | Edge class | 0.0012 € |
| Qwen3.5 9B (Together) | Compact class | 0.0015 € |
| Ministral 3 14B | Edge class | 0.0016 € |
| Mistral Small 4 | Compact class | 0.0020 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0058 €); Qwen3.7-Plus (Together) (Workhorse class, 0.0044 €); Qwen3.5 9B (Together) (Compact class, 0.0015 €)

### Sales email and proposal draft

Input 6,000 tokens, output 1,200 tokens, one call per task, 20 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0013 € |
| Ministral 3 8B | Edge class | 0.0019 € |
| Ministral 3 14B | Edge class | 0.0026 € |
| Qwen3.5 9B (Together) | Compact class | 0.0026 € |
| Mistral Small 4 | Compact class | 0.0030 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0088 €); Qwen3.7-Plus (Together) (Workhorse class, 0.0067 €); Qwen3.5 9B (Together) (Compact class, 0.0026 €)

### Classification and routing

Input 1,500 tokens, output 30 tokens, one call per task, one minute of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0002 € |
| Mistral Small 4 | Compact class | 0.0003 € |
| Ministral 3 8B | Edge class | 0.0003 € |
| Qwen3.5 9B (Together) | Compact class | 0.0003 € |
| gpt-5.4-nano | Compact class | 0.0004 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0010 €); Qwen3.7-Plus (Together) (Workhorse class, 0.0007 €); Mistral Small 4 (Compact class, 0.0003 €)

### Extracting data from documents

Input 15,000 tokens, output 600 tokens, one call per task, 5 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0024 € |
| Ministral 3 8B | Edge class | 0.0036 € |
| Mistral Small 4 | Compact class | 0.0041 € |
| Qwen3.5 9B (Together) | Compact class | 0.0042 € |
| Ministral 3 14B | Edge class | 0.0049 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0131 €); Qwen3.7-Plus (Together) (Workhorse class, 0.0087 €); Mistral Small 4 (Compact class, 0.0041 €)

### Knowledge-base question

Input 8,000 tokens, output 400 tokens, one call per task, 3 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0012 € |
| Ministral 3 8B | Edge class | 0.0018 € |
| Mistral Small 4 | Compact class | 0.0020 € |
| Qwen3.5 9B (Together) | Compact class | 0.0023 € |
| Ministral 3 14B | Edge class | 0.0024 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0065 €); Qwen3.7-Plus (Together) (Workhorse class, 0.0048 €); Mistral Small 4 (Compact class, 0.0020 €)

### Research agent over documents

Caution case, see cost per solved task below.

Input 25,000 tokens, output 1,200 tokens, 8 calls per task, 5 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0161 € |
| Ministral 3 8B | Edge class | 0.0242 € |
| Mistral Small 4 | Compact class | 0.0315 € |
| Ministral 3 14B | Edge class | 0.0322 € |
| gpt-5.4-nano | Compact class | 0.0479 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0968 €); DeepSeek V4 Flash (Workhorse class, 0.0968 €); Mistral Small 4 (Compact class, 0.0315 €)

### Writing or changing code

Input 15,000 tokens, output 1,500 tokens, 3 calls per task, 10 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0054 € |
| Ministral 3 8B | Edge class | 0.0081 € |
| Ministral 3 14B | Edge class | 0.0108 € |
| Mistral Small 4 | Compact class | 0.0113 € |
| Qwen3.5 9B (Together) | Compact class | 0.0136 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0341 €); DeepSeek V4 Flash (Workhorse class, 0.0293 €); Mistral Small 4 (Compact class, 0.0113 €)

### Working through a long document

Input 300,000 tokens, output 3,000 tokens, one call per task, 15 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| gpt-5.4-nano | Compact class | 0.0965 € |
| Gemini 3.1 Flash-Lite | Compact class | 0.1236 € |
| Gemini 3.5 Flash-Lite | Compact class | 0.1516 € |
| gpt-5.6-luna | Compact class | 0.1898 € |
| DeepSeek V4 Flash | Workhorse class | 0.2114 € |

Cheapest model per class: DeepSeek V4 Pro (Frontier class, 0.6342 €); DeepSeek V4 Flash (Workhorse class, 0.2114 €); gpt-5.4-nano (Compact class, 0.0965 €)

### Meeting notes

Input 15,000 tokens, output 800 tokens, one call per task, 3 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0020 € |
| Ministral 3 8B | Edge class | 0.0031 € |
| Mistral Small 4 | Compact class | 0.0035 € |
| Qwen3.5 9B (Together) | Compact class | 0.0036 € |
| Ministral 3 14B | Edge class | 0.0041 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0113 €); Qwen3.7-Plus (Together) (Workhorse class, 0.0075 €); Mistral Small 4 (Compact class, 0.0035 €)

### Translation

Input 2,000 tokens, output 2,000 tokens, one call per task, 2 minutes of rework

| Model | Class | Machine per task |
| --- | --- | --- |
| Ministral 3 3B | Edge class | 0.0005 € |
| Ministral 3 8B | Edge class | 0.0008 € |
| Ministral 3 14B | Edge class | 0.0010 € |
| Qwen3.5 9B (Together) | Compact class | 0.0011 € |
| Mistral Small 4 | Compact class | 0.0019 € |

Cheapest model per class: Mistral Large 3 (Frontier class, 0.0052 €); Qwen3.7-Plus (Together) (Workhorse class, 0.0041 €); Qwen3.5 9B (Together) (Compact class, 0.0011 €)

## Cost per solved task where a solve rate is documented

Where a continuously maintained benchmark series provides a solve rate per model, the calculator prices in attempts, error cost and review time. For the other scenarios no such series exists, and no number is shown on purpose.

### Answering a support ticket

Benchmark series: tau2 · 2 models with a value

| Model | Solve rate | Attempts | Machine per task | per solved task |
| --- | --- | --- | --- | --- |
| gpt-5.2 | 73 % | 1.6 | 0.025 € | 1.07 € |
| gpt-5.1 | 59 % | 2.0 | 0.022 € | 1.34 € |

### Research agent over documents

Benchmark series: browsecomp · 15 models with a value

| Model | Solve rate | Attempts | Machine per task | per solved task |
| --- | --- | --- | --- | --- |
| gpt-5.6-sol | 92 % | 1.3 | 1.207 € | 11.01 € |
| Claude Opus 5 | 91 % | 1.3 | 2.749 € | 13.95 € |
| gpt-5.6-terra | 88 % | 1.4 | 0.506 € | 15.01 € |
| Claude Sonnet 5 | 85 % | 1.4 | 0.867 € | 18.17 € |
| Claude Mythos 5 | 88 % | 1.4 | 4.189 € | 18.19 € |
| gpt-5.6-luna | 83 % | 1.4 | 0.053 € | 18.75 € |
| DeepSeek V4 Pro | 83 % | 1.4 | 0.259 € | 18.86 € |
| gpt-5.5 | 84 % | 1.4 | 1.279 € | 18.88 € |

## Method

1. Prices are vendor list prices in USD per 1 million tokens, without discounts, checked on 18 August 2026 on the vendors' pricing pages. Converted at the ECB reference rate of 1.1576 USD per euro.
2. Machine cost per task: input, output and cache shares following the logic of the AI cost calculator, one attempt, no batch processing, no router, no US residency. Models whose context window cannot hold the input are omitted from that table.
3. Cost per solved task: machine cost divided by the solve rate from the named benchmark series, corrected for dependent retries (factor 1.2), plus error and review cost according to the scenario profile. Where no series exists, nothing is interpolated.
4. Only models with an active, bookable endpoint and a published list price are included.
5. Your own sizes, cache shares, batch processing and error costs can be set in the calculator.

### Changes

- 18 August 2026: first edition with 38 current models. Updated quarterly with the calculator's price status.

### How to cite

Convios GmbH (2026): AI Cost Benchmark 2026, as of 18 August 2026, https://www.convios.com/en/insights/ki-kosten-benchmark-2026. Free to use with attribution (CC BY 4.0).

- [Open the AI cost calculator with your own sizes](https://olivergausmann.com/en/insights/ai-cost-calculator)
- [The article behind the method: cutting AI cost without losing quality](https://olivergausmann.com/en/insights/ki-kosten-senken-modellwahl)
- HTML: https://www.convios.com/en/insights/ki-kosten-benchmark-2026