What this measures
This is a capability audit, not a task benchmark. Each row is a tool behavior, such as a bounded read window or a wait mode on a background process. Each row records whether a harness ships the behavior in its default configuration.
The rubric has 61 rows: 24 for the shell and 37 for the read. Every harness was scored against the same rows and the same weights. A full cell means the behavior ships, a half cell means partial or behind a flag, and an empty cell means the behavior was not found.
This page does not run GitHub issues, a repository, or a test suite.
The frontier
The x-axis ranks the 61 capabilities from the most modeled savings to the least. The y-axis adds them up as you move right, so each step is one capability's weight. It is a log scale on the gap to 100%, so the top of the field is readable. The dashed line at the top is 100%, or every capability shipped.
The leading line is the frontier: the most savings any harness captures at each step. Midas holds it and ends at 98.2%. Command Code is next at 95.9%. Hermes follows at 77.8%.
The distance between a line and the ceiling is the savings that harness leaves unclaimed. The log scale spreads the top of the field. The first capabilities carry most of the weight, so one missing expensive behavior costs more than several missing cheap ones. The gap between Midas and Command Code is the lead.
Token efficiency frontier
61 shell and read capabilities, ranked by modeled savings
Capability audit · Nov 04, 2026 · 11 harnesses · 659,000 token ceiling
Share of the 659,000 ceiling. Log scale on the gap to 100%.
Audit date: 4 November 2026
The weights estimate the tokens each behavior saves per million tokens of tool traffic.
The field
This ranked list uses the same scores. The bar is the share of the 659,000 ceiling each harness captures. The evidence table later splits the score into the shell and read tools.
- 01Midas98.2
- 02Command Code95.9
- 03Hermes77.8
- 04Claude Code62.6
- 05OpenClaw55.4
- 06Grok Build52.3
- 07Cline45.6
- 08Kilo Code44.9
- 09OpenCode34.9
- 10Codex27.6
- 11pi22.1
For task outcomes rather than tool-surface grades, see the Terminal-Bench 2.1 page.
Where the gap comes from
The matrix shows all 61 rows as cells. Midas is nearly solid. Codex is strong on the shell and almost empty on the read side, which is the large empty block in its row. That gap pulls its frontier line down.
The evidence
Every harness is pinned to a version and a commit. Open a version to read the evidence behind each cell.
| Harness | Version | All | Full | ||
|---|---|---|---|---|---|
| Midas | 0.1.85 | 97.2 | 99.0 | 98.2 | 58/61 |
| Command Code | 1.88.2 | 94.2 | 97.5 | 95.9 | 54/61 |
| Hermes | dce2ecb | 83.0 | 73.3 | 77.8 | 40/61 |
| Claude Code | 2.1.311 | 68.6 | 57.4 | 62.6 | 29/61 |
| OpenClaw | 0a6c013 | 63.0 | 48.9 | 55.4 | 24/61 |
| Grok Build | bc7f02e | 66.9 | 39.6 | 52.3 | 22/61 |
| Cline | 48d6385 | 35.8 | 54.0 | 45.6 | 16/61 |
| Kilo Code | 5e02825 | 40.8 | 48.3 | 44.9 | 23/61 |
| OpenCode | 10765ff | 25.0 | 43.4 | 34.9 | 15/61 |
| Codex | 88f7765 | 53.7 | 4.9 | 27.6 | 10/61 |
| pi | 853a80d | 11.0 | 31.7 | 22.1 | 10/61 |
How this was measured
The audit covers 61 capabilities across 11 harnesses, with a modeled ceiling of 659,000 tokens and zero model calls. This page grades whether a harness ships a tool behavior. It does not run a repository or a test suite.
Most public comparisons measure model quality. The harness is where the token bill is decided, so this audit grades the tool surface directly, with evidence a reader can check.
The rubric covers 61 capabilities: 24 for the shell and 37 for the read. Every harness was scored against the same rows and the same rules.
Each row carries a weight, the modeled tokens it saves per 1M tokens of tool traffic. The weights sum to 659,000, the ceiling. A harness that ships every row captures all of it.
Open harnesses were read at the pinned commits in the table. Two ship compiled bundles, so those columns come from the bundle and the published documentation.
Every positive cell in the raw sheets carries a file and line, a bundle offset, or a probe transcript, so you can check the score.
Row definitions follow Command Code's Token Efficiency Frontier benchmark, published in August 2026.
Run the audit
The rubric and all eleven score sheets are published with this page. The verifier recomputes every total from them and compares the result to the table above.
curl -fsSLO https://trymidas.io/benchmarks/verify.mjs node verify.mjs
It reads rubric.json and the sheet behind each version, then prints the same shell, read, and all scores. A mismatch exits non-zero.
Node 20 or later. The script itself is verify.mjs.
Use Midas
This audit is not a product trial. To use Midas on your own repository, install it with one command, then start it in your project.
Audit date: 4 November 2026. The Midas row is version 0.1.85.