Skip to content

Model Recommendations

Complexity score (0-100)

ctok computes a complexity score from the input before picking a model:

SignalMax weight
Task type base (bug-fixarchitecture)12-60
Input volume (log-scaled)35
File breadth (number of attached files)20
Content heterogeneity (≥3 kinds)6
Keyword signals (security, distributed, migrate, …)±12 each
Large expected output8

Bands: simple < 30 ≤ normal < 65 ≤ deep

Model selection

BandRecommended modelWhy
Simplehaiku-4-5Narrow edits, cosmetic, docs - Haiku is fast and cheap.
Normalsonnet-4-6The default workhorse - strong reasoning, good price.
Deepopus-4-7Architecture, hairy bugs, cross-cutting refactors. Worth the premium.

When input exceeds 150k tokens, the recommender nudges toward Sonnet or Opus regardless of complexity band (long-context tasks benefit from stronger models).

Effort levels

Effort starts from the complexity band and is adjusted by task type:

Task typeAdjustment
architecturexhigh
debugginghigh
feature+1 level from band
reviewlow or medium
documentationlow
bug-fixno override
refactorno override
generalno override

Levels: lowmediumhighxhigh

Pricing reference

ModelInputOutput
haiku-4-5$0.80/M$4.00/M
sonnet-4-6$3.00/M$15.00/M
opus-4-7$15.00/M$75.00/M

Anthropic list rates as of May 2026. ctok shows estimates only - verify at anthropic.com/pricing.