A command that reruns a fixed coding benchmark on your own machine, comparing ordinary Claude Code runs with Claude Code guided by Heimdall.

Use it to inspect the benchmark suite and measurement plan, or run the tasks and compare tokens, elapsed time, passing tests, and cost.

It lets you verify published measurements by actually running the tasks instead of relying on hardcoded claims, while a dry run avoids API calls and agent runs.

bench — Command by randomittin

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods

npx agentmods add commands/randomittin/heimdall/bench

Clone the repo

git clone --depth 1 https://github.com/randomittin/heimdall

Per session 101 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.

When invoked 994 The whole file, excluding the scripts and references it only reads on demand.

Security scan A 0 findings. Scan, not verified.

Origin original No closer match found in the catalogue.

Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

Model	Per session	Once invoked
Fable 5	$0.00101	$0.00994
Opus 5	$0.00051	$0.00497
Sonnet 5	$0.00020	$0.00199
Haiku 4.5	$0.00010	$0.00099

Measured yesterday against content hash 6ec5ee615455, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

bench scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

commands/bench.md · 89 lines

How it starts

The opening of the file, as written. The whole thing — 89 lines — stays where its author put it; the contents beside it link to each section on GitHub.

/hmd:bench — reproduce the public benchmark table

Runs bin/heimdall-bench — the documented entry door over the measurement harness (bin/benchmark). It does NOT invent numbers: it produces them by running a fixed suite of representative coding tasks under two arms (raw Claude Code vs Claude Code driven through Heimdall) and emitting the exact table format the published evals/ table uses. The number in any Heimdall claim is one a stranger can reproduce here.

When to use

Someone challenges a benchmark number — point them at heimdall bench so they reproduce it on their own machine instead of trusting the README.
Before a launch, to regenerate the table on a pinned model (the flagship).
To see, with zero API spend, exactly what a real run measures.

Instructions

Always start dry (the default). Validates the suite and prints the capture plan with NO API calls and NO agents spawned:
```
heimdall-bench
heimdall-bench --dry          # explicit, identical
```
This lists every task, its category, and its verify[] steps, then prints the exact arm commands a live run would issue and what each metric is parsed from. A cold stranger can run this without burning a token.
Run live only when the user opts in. A --live run invokes the claude CLI for every task × arm and spends real API tokens. Pin the model for a reproducible, publishable table:
```
heimdall-bench --live --model <model-id>
```
It writes evals/benchmark/results.jsonl (one machine-readable line per task × arm) and evals/benchmark/results.md (the published-format markdown table). Numbers come from the CLI's own usage accounting and from actually running each task's verify[] — nothing is hand-tuned.
Reprint the last live table without re-running:
```
heimdall-bench --table
```
Honesty principle. Publish everything measured, including tasks where Heimdall loses on tokens, wall time, or cost. A pricier arm that ships 7/7 beats a cheaper arm that ships 3/7; hiding the trade-off would make the whole table untrustworthy.

Read the full file on GitHub · 89 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

yesterday First seen · 89 lines · 0 tokens per session scan A 6ec5ee615455

Subscribe to this mod's changes

bench is a command published in the GitHub repository randomittin/heimdall (5 stars, last pushed 11d ago), licensed MIT. It adds 101 tokens to every session and 994 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.

Code review Planning Testing ai-coding anthropic claude-code claude-plugin cli developer-tools

bench

Getting it into your agent

What it costs to keep this loaded

Grade A, and why

Nothing flagged

How it starts

/hmd:bench — reproduce the public benchmark table

When to use

Instructions

What this file has done since we first saw it

Other commands, from other repositories

autocode

autocode-new

dispatcher

standup

dev-planner

iterate