11ai-benchmarks-faq

Benchmarks

Answer questions about the 11ai-benchmarks plugin: which reporting skill fits a job, how the analyzers discover and classify harness usage, how pricing, timing, and cost states are derived, what each report section and warning means, and how the skills are maintained. Routes every question to the plugin's own contracts, references, scripts, and tests and answers with a citation. Use when a user asks how the benchmark skills behave, why a report shows a value, which benchmarks skill to run, or what a report section means.

terminal
$ npx skills add rj11io/11ai --skill 11ai-benchmarks-faq

Try it — say this to your agent

Use 11ai-benchmarks-faq for this task.

The full playbook · v0/plugins/11ai-benchmarks/skills/11ai-benchmarks-faq/SKILL.md

11ai Benchmarks FAQ

Answer questions about this plugin by routing them to the plugin's own files, not from memory. This skill owns no behavior: the sibling skills' contracts, references, scripts, and tests are the truth, and the routing table below says which file answers which question. The only content this skill owns is the comparison table and the glossary.

Answer contract

  • Read the routed source before answering. Never answer a behavior question from memory.
  • Cite the file, and the section when one is named, in every answer.
  • If the sources do not cover the question, say so and name the skill that owns the behavior. Do not guess.
  • Escalate tiers when the question demands it: a question about what a skill promises stops at its contract; a question about what exactly happens in an edge case goes to the tests.
  • If a routed sibling file is missing because only this skill was installed, read the same skill's page at https://ai.rj11.io/skills/ followed by the skill name, or the rj11io/11ai repository on GitHub, instead.

Ground-truth ladder

Tier Meaning Sources
contract What a skill promises Sibling SKILL.md files
reference Coverage tables and catalogs Files under sibling references directories
behavior What actually happens, asserted Sibling scripts and tests

Covered skills

  • 11ai-benchmarks-machine — machine-wide usage, cost, and timing reports.
  • 11ai-benchmarks-project — reports for one repository and its attributed threads.
  • 11ai-benchmarks-single-thread — a report for one exact thread and its sub-agents.
  • 11ai-benchmarks-pricing-update — maintains the bundled pricing catalogs; writes no report.

Which skill do I want?

Skill Scope Input Output
11ai-benchmarks-single-thread One thread plus its spawned sub-agent threads A thread selector, or the active thread Markdown and HTML report package under the thread folder
11ai-benchmarks-project One repository and every thread attributed to it The repository root folder Markdown and HTML report package under the thread folder
11ai-benchmarks-machine Every readable usage store on the machine Nothing Markdown and HTML report package on the Desktop
11ai-benchmarks-pricing-update The bundled pricing catalogs of the three reporting skills Official provider pricing pages Updated, synchronized catalog copies

Routing table

Every row: a question, the file that answers it, the text to find in that file, and its tier on the ground-truth ladder.

Choosing and running

Question Source Anchor Tier
What does the machine-wide skill promise end to end? ../11ai-benchmarks-machine/SKILL.md "## Contract" contract
What does the project skill promise end to end? ../11ai-benchmarks-project/SKILL.md "## Contract" contract
What does the single-thread skill promise end to end? ../11ai-benchmarks-single-thread/SKILL.md "## Contract" contract
What does the pricing-update skill actually do? ../11ai-benchmarks-pricing-update/SKILL.md "## Workflow" contract
How do I run the machine-wide analyzer and which flags exist? ../11ai-benchmarks-machine/SKILL.md "## Contract" contract
How do I select one exact thread to analyze? ../11ai-benchmarks-single-thread/SKILL.md "## Workflow" contract
How do I include exported usage from an unsupported harness? ../11ai-benchmarks-machine/SKILL.md "--include" contract
What checks must pass before a run counts as complete? ../11ai-benchmarks-machine/SKILL.md "## Completion checks" contract

Harness coverage and classification

Question Source Anchor Tier
Which coding harnesses are covered and where do their usage stores live? ../11ai-benchmarks-machine/references/harnesses.md "# Native harness stores" reference
How are user-facing surface and underlying runtime told apart? ../11ai-benchmarks-machine/references/harnesses.md "## Surface and billing semantics" reference
Which usage shapes count as authoritative, and which never become costs? ../11ai-benchmarks-machine/SKILL.md "## Supported usage shapes" contract
How does declared workspace metadata attribute usage to a project? ../11ai-benchmarks-machine/references/harnesses.md "## Supplemental workspace attribution" reference

Cowork sessions and sub-agents

Question Source Anchor Tier
How are Cowork root and sub-agent transcripts grouped into sessions? ../11ai-benchmarks-machine/references/harnesses.md "## Claude Cowork" reference
What happens to detected remote Cowork sessions without readable usage? ../11ai-benchmarks-machine/scripts/analyze-llm-cost-global.mjs "remote Cowork" behavior
Where do measured Cowork sessions and sub-agent runs show up in a report? ../11ai-benchmarks-machine/scripts/analyze-llm-cost-global.mjs "Cowork coverage" behavior
Is Claude usage ever counted twice across Desktop, Cowork, and CLI surfaces? ../11ai-benchmarks-machine/references/harnesses.md "Claude Desktop metadata" reference

Pricing and cost states

Question Source Anchor Tier
Where do prices come from, and can I override them locally? ../11ai-benchmarks-machine/SKILL.md "## Pricing" contract
Which rate applies when usage predates the known price history? ../11ai-benchmarks-pricing-update/tests/pricing-history.test.mjs "uses the earliest available rate when usage predates known history" behavior
Which rate applies when a thread crosses a price-change boundary? ../11ai-benchmarks-pricing-update/tests/pricing-history.test.mjs "uses the main price at the finish date when an aggregated thread crosses a boundary" behavior
Which rate applies when usage has no date at all? ../11ai-benchmarks-pricing-update/tests/pricing-history.test.mjs "uses the latest rate active at report time when usage is undated" behavior
Why is a model listed as requiring a pricing update, and what does that exclude? ../11ai-benchmarks-machine/scripts/analyze-llm-cost-global.mjs "Models requiring a pricing update" behavior
Which official sources may feed the catalog, and how are they reviewed? ../11ai-benchmarks-pricing-update/references/provider-sources.md "# Official provider pricing sources" reference
What schema and rate periods does the pricing catalog use? ../11ai-benchmarks-pricing-update/references/provider-sources.md "## Catalog schema" reference
Is a specific model priced in the bundled catalog right now? ../11ai-benchmarks-pricing-update/references/pricing.json - reference
How are the three bundled catalog copies kept identical? ../11ai-benchmarks-pricing-update/scripts/sync-pricing-catalog.mjs - behavior
Who is allowed to change the bundled catalogs? ../11ai-benchmarks-pricing-update/SKILL.md "## Catalog ownership" contract
Why does the report not match my invoice or subscription bill? ../11ai-benchmarks-machine/SKILL.md "API-equivalent estimate" contract
How is response latency measured and which harnesses does it cover? ../11ai-benchmarks-machine/SKILL.md "Response latency" contract

Timing

Question Source Anchor Tier
What is the difference between wall time and estimated active time? ../11ai-benchmarks-machine/SKILL.md "## Timing" contract
How is timing measured for one thread and its sub-agents? ../11ai-benchmarks-single-thread/SKILL.md "## Timing semantics" contract

Reports and outputs

Question Source Anchor Tier
Where are report packages written by default? ../11ai-benchmarks-machine/SKILL.md "11ai-benchmarks-machine-reports" contract
Where does the project report land, and how is its folder named? ../11ai-benchmarks-project/SKILL.md "11ai-benchmarks-project-reports" contract
Where does the single-thread report land, and how is its folder named? ../11ai-benchmarks-single-thread/SKILL.md "11ai-benchmarks-single-thread-reports" contract
Are unreadable or malformed records silently dropped from the totals? ../11ai-benchmarks-machine/scripts/analyze-llm-cost-global.mjs "malformed" behavior
What must every report contain, and how does the HTML behave? ../11ai-benchmarks-machine/SKILL.md "## Report requirements" contract
How is usage attributed to monthly, quarterly, and yearly periods? ../11ai-benchmarks-machine/SKILL.md "## Period attribution" contract
What does the Pricing coverage section of a report tell me? ../11ai-benchmarks-machine/scripts/analyze-llm-cost-global.mjs "## Pricing coverage" behavior
What does the Anomalies and limitations section flag? ../11ai-benchmarks-machine/scripts/analyze-llm-cost-global.mjs "## Anomalies and limitations" behavior
What does the Historical pricing selection detail show? ../11ai-benchmarks-machine/scripts/analyze-llm-cost-global.mjs "Historical pricing selection" behavior
How is cost broken down by model and reasoning effort? ../11ai-benchmarks-machine/scripts/analyze-llm-cost-global.mjs "Cost by model by effort" behavior
How does a report classify each observed harness? ../11ai-benchmarks-machine/scripts/harness-support.mjs "classifyThread" behavior

Maintenance and regression

Question Source Anchor Tier
What must happen when one sibling skill changes behavior? ../../README.md "## Cross-skill maintenance" reference
What is the report regression gate before a behavior change ships? ../../README.md "regression gate" reference

Glossary

Definitions only; the numbers, paths, and defaults live in the routed sources above.

Term Meaning
Harness The combination of surface, runtime, store, and billing mode that identifies how a thread ran and was recorded
Surface The user-facing product a thread ran in, such as claude-code, claude-desktop-code, codex-cli, or chatgpt-work
Runtime The underlying engine that executed the thread and wrote its usage ledger, such as claude or codex
Store The usage store the numbers were read from, such as claude-projects, codex-rollouts, or opencode-db
Usage source The kind of evidence behind the numbers: a native transcript, a local database, or an imported export
Confidence How trustworthy the counters are: harness-reported tokens versus values derived from a record's shape
Billing mode Whether usage was billed per token through an API key or covered by a subscription
Output TPS Blended output tokens per second: output tokens divided by estimated active seconds, over threads with measured timing; includes tool time and excludes time-to-first-token
Active time / response Blended turnaround: estimated active time divided by counted usage responses; a work-rate proxy, not provider latency
Response latency Per-response duration measured by a per-harness method (record-timestamps, turn-events, request-events, or row-durations); includes network, queue, decode, and in-turn tool time, never time-to-first-token
Thread One conversation or session as the unit of attribution, cost, and timing
Known cost Cost computed from measured tokens matched to a bundled catalog rate
Unpriced model A measured model with no catalog match; its cost stays unavailable instead of zero
Temporal fallback Applying a documented substitute rate when usage falls outside known price history
Measured Cowork session A Cowork session whose local transcripts were readable and counted
Sub-agent run A child thread spawned by a root thread and reported as part of its tree
Active time Estimated time the model was actually working, as opposed to wall-clock elapsed time
Attribution time The timestamp that decides which price period and report period a thread belongs to

Want the source?

This page renders the skill's markdown straight from the repository.

Open on GitHub
More Benchmarks skills