Which model should review your code
Octopus lets you choose the model that reviews your pull requests, for a whole organisation or for a single repository. This guide covers what the choice changes and what each model has cost per review. It ends with how to switch.
What changes when you change the model
- Cost per review. Octopus Cloud bills AI usage at 2x the provider's list price. Among the models below, list prices run from $1 to $10 per 1M input tokens.
- How long a review takes. Most reviews finish in 30 to 90 seconds, and a pull request with hundreds of files can take a few minutes. Higher reasoning effort means a slower review.
- How much reasoning goes into it. Some models think before they answer. The Reasoning Effort setting controls how hard those models reason. Models without extended thinking ignore it.
What each model cost per review
These are the median amounts charged per review on Octopus Cloud over the 30 days up to 14 November 2026. They cover reviews on platform keys and include the 2x platform rate. The last column is the range that holds the middle half of reviews.
| Model | Median per review | Middle half of reviews |
|---|---|---|
| Claude Haiku 4.5 | $0.09 | $0.07 to $0.11 |
| GPT-5.3 Codex | $0.16 | $0.15 to $0.19 |
| Qwen3.8-Max | $0.17 | $0.14 to $0.22 |
| Claude Sonnet 5.5 | $0.42 | $0.23 to $0.64 |
| Claude Opus 5.5 (Cloud default) | $0.76 | $0.57 to $0.98 |
| Claude Opus 5 | $1.03 | $0.73 to $1.41 |
| Claude Fable 5.2 | $1.52 | $1.19 to $2.13 |
| GPT-6 Astra | $1.67 | $1.54 to $2.33 |
| Claude Fable 5.1 | $2.38 | $1.98 to $3.10 |
This is cost data, not quality data: we have not measured which model finds more problems, so nothing here ranks the models on review quality. Each model reviewed the pull requests its own users sent it, so the rows are not a test on identical pull requests. Your figures will depend on the size of your changes.
Which model to pick
- If you have no reason to change, keep the Cloud default, Claude Opus 5.5.
- If a repository opens many pull requests and cost matters most, give it a cheaper model. Claude Haiku 4.5, GPT-5.3 Codex and Qwen3.8-Max each had a median below $0.20 per review. Claude Sonnet 5.5 came in at $0.42.
- If you want the deepest review, for example on security-sensitive code, pick a Fable model. Fable is the tier Anthropic places above Opus. Fable 5.2 has the lower list price of the two: $8 input and $40 output per 1M tokens, against $10 and $50 for Fable 5.1.
- If a pull request is very large, Claude Fable 5.2 and Fable 5.1 can review the complete change instead of a shortened diff, as long as the change fits the size, cost and time limits for a single review.
- If you already pay a provider, add your own API key. Octopus sends requests through your key and deducts no credits for them, and the provider bills you. Other platform operations, such as embeddings, still use credits.
How to switch
Open Settings and choose Models, under AI & Reviews.
- For the whole organisation: in Organization Defaults, pick a Review & Chat Model. Reasoning Effort is in the same card.
- For one repository: in Repository Models, find the repository and pick its Review & Chat Model. This overrides the organisation default.
- For your own key: add it under Provider Keys, next to Models in Settings.
A repository without its own model uses the organisation default. Provider list prices for every model are on the pricing page.