Updated
We test the tools businesses pay for, from hosting to password managers, and publish every number.
Insight · 9 minute read
The DAYJOB Tax AI benchmark found that Claude Sonnet 5.5 passed only 10.8% of attempts at completing professional tax assignments. The October 2026 results show how AI agents struggle with conflicting records, missed tax issues and complex work that requires professional judgement.
By Vidyadhar.
Report · 29 Oct 2026
GitHub Copilot Fusion uses multiple AI models to handle coding tasks without requiring developers to select each model manually. Independent testing found that Fusion Max scored 60 on the Coding Agent Index, only two points behind Claude Code with Opus 5.5 Max, while using approximately half the tokens and costing about half as much per task.
Report · 28 Oct 2026
Couchbase semantic caching reduced LLM calls by 40% to 60% in company demonstrations announced in October 2026. The approach can avoid repeated model requests, but actual savings depend on workload repetition, cache accuracy and operating costs.
Report · 28 Oct 2026
DuckDB 2.0 introduces Agent Mode, which reduced query-output tokens by approximately 59% in DuckDB's own AI agent experiment. However, the reduction did not produce a meaningful improvement in overall agent costs or runtime.
Every number on this site is one we measured, with the method written next to it.
Read the full methodNewsletter
New reports, test results and price changes, about once a month. No spam, and every email has an unsubscribe link.
