A tool page should answer one question: what can I actually do with this? Each entry lists the documented capabilities it carries, the models behind it, and what it refuses to do.
These records describe what providers officially document. benchr has not run these capabilities itself, so nothing here is a test result. Each record shows the vendor page it was read from and the date.
Ledger updated: September 3, 2026
Claude Code
Coding agent
AnthropicDocs checked September 1, 2026
An agentic coding tool that reads a codebase, edits files, runs commands and works with git - in a terminal, an IDE, a desktop app and the browser.
What you can actually do with it
Build features and trace bugs across many files, then run the tests and fix what fails.
Stage changes, write commit messages, create branches and open pull requests directly.
Load project rules from CLAUDE.md, run hooks around edits and commits, and package repeated procedures as skills.
Connect to your own systems through MCP, and run unattended on a schedule or from a GitHub event.
What it will not do
Work without a Claude subscription or Console account on most surfaces.
Auto-update when installed through Homebrew or WinGet - only the native install does.
Constrain itself: permissions and hooks are configuration, not defaults you can assume.
Status
Work through an entire repository, not a pasted fileDocumented
Hand off a task and walk away from the machineDocumented
Teach the model a procedure once and reuse itDocumented
One screenshot in, a working page outPartly documented
Give a model your own tools and data through MCPDocumented
The surface that carries the server-side tools: browser control, desktop control, sandboxed code execution, web search with citations, batching and prompt caching.
What you can actually do with it
Run Anthropic-executed tools inside a turn, so search and code execution do not need a client-side loop.
Cut repeat-context cost to a tenth with prompt caching, and bulk work to half price with the Batches API.
Read PDFs up to 32 MB and 600 pages, and images up to 8000x8000 px, by file reference rather than base64.
Run Agent Skills inside the code execution container by skill_id.
What it will not do
Offer every tool on every platform - browser use, code execution and web search each have their own platform gaps across Bedrock, AWS, Google Cloud and Microsoft Foundry.
Give code execution or Skills zero-data-retention eligibility.
Allow Skills on the API to reach the network or install packages at runtime.
Status
Drive a real browser, not a scraperDocumented
Operate a desktop applicationDocumented
Let the model write and run real code mid-answerDocumented
Read a 600-page PDF including its chartsDocumented
Answer from the live web with citations attachedDocumented
Stop paying twice for the same fileDocumented
Run the whole job overnight at half priceDocumented
Teach the model a procedure once and reuse itDocumented
One screenshot in, a working page outPartly documented
Fit a whole codebase in one prompt, and find its limitPartly documented
benchr