Updated · Sources & corrections
In this guide
Sonnet and Opus are model families, not permanent performance specifications. A useful comparison records the exact model versions, access route and evaluation date. The Claude 4 release linked below documents Sonnet 4 and Opus 4; it should not be read as a claim that these are the newest available models.
First decide what you are comparing
Comparing two API model IDs isolates a different question from comparing two complete coding-agent workflows. An agent also contributes prompts, tools, context retrieval and permissions. Keep those conditions fixed before attributing an outcome to the underlying model.
Use current prices for the exact models
Provider, region, cached input, output length and batch processing can change billed cost. Record the actual input/output token usage and the rate applicable to that request. A subscription price is not an API token price. The original vendor release provides historical positioning; your provider’s current billing page is authoritative for a purchase.
Measure cost per successful task
- Choose tasks with independently checkable acceptance criteria.
- Keep the starting repository, prompt and available tools fixed.
- Repeat each task, including failed runs and retries.
- Record elapsed time, token usage, human corrections and whether tests passed.
- Compare cost and time only for results that meet the same acceptance criteria.
A cheaper token rate can still lead to a more expensive task if it requires several retries. A more expensive model can also provide no practical benefit on a simple task. Neither outcome can be decided from the family name alone.
Match the evaluation to your workload
For routine code edits, include small changes with tests. For architectural work, include requirements that cross files and capture incorrect assumptions. For writing, use a review rubric rather than a single preference score. Do not use a coding score to claim a universal reasoning or writing winner.
This page does not publish an independent Sonnet-versus-Opus benchmark or a measured throughput ratio. Earlier blanket speed and success-rate claims have been removed. The purpose is to help you evaluate a versioned choice without turning vendor positioning into a universal result.
Related: Gemini 2.5 Pro and Claude Opus 4 and comparing the terminal agents.
Sources and verification scope
This article distinguishes documented features and vendor-reported results from independent testing. No original head-to-head benchmark dataset is published for this page; hardware-memory figures retained in the guide are estimates, not measured guarantees.
Editorial update, 21 October 2026: source references and verification scope added; identified command, feature or wording errors corrected. Read our sourcing and correction policy.