A clear win on getting the code right.
We gave Claude Code five feature tasks that depend on team rules written outside the code, with and without Mios. With Mios, the first answer was fully correct in about half the runs (5 of 11), against none of 41 without it. It took almost half as many prompts to reach correct code, for a total cost close to Claude Code alone.
Runs of November 12 and 13, 2026 · Mios build of November 12, commit 251d8392 · Claude Sonnet 5.5
How we measured it.
- 01Five real feature tasks
Five features on the repository of a SaaS product in production. Each one touches several files and plugs into existing code: route, controller, service, test and docs. The agent gets the ticket and nothing else.
- 02Hidden tests
The agent never sees the tests that grade it; they are copied in only for grading. One check per task for the feature itself, one per team rule, 24 checks in all. No model takes part in the grading.
- 03Rules written outside the code
Every task depends on rules the team wrote in chat, shared documents and messaging. 11 of them appear nowhere in the repository, so an agent follows them only if something brings them in.
- 04Five setups, one model
Claude Code alone, with Obsidian on the repository’s docs, with Graphify on the repository, with both, and with Mios. Every setup runs Claude Sonnet 5.5 on the same clone, with the same ticket and the same time limit.
- 05Two runs per cell
Two runs per task and setup. On the one task where the two runs disagreed, a third run for Mios and for the best other setup.
- 06The same review for everyone
When the first answer fails a test, the same session gets a review comment on the failing parts, word for word the same for every setup. From the fifth prompt on, it also states the missing rule and shows the failing tests, so every setup ends with correct code. Up to 8 prompts.
- 07Bars written first
What counts as a win and what counts as noise was written down before the first run. Two things changed along the way, both listed in the limits below: the prompt cap went from 4 to 8, and the review loop is now counted per task instead of per run.
One prompt, the first answer.
What each setup hands back after the ticket alone. Time and cost are medians per task. Cost is what Claude Code spends, plus the Mios server for Mios.
| Setup | Fully correct | Checks passed | Off-repo rules followed | Time | Cost |
|---|---|---|---|---|---|
| Claude Code alonethe ticket and the repository | 0 of 10 runs | 23%5.5 of 24 | 0 of 11 | 90 s | $0.266 |
| + Obsidianthe repository’s docs as notes | 0 of 11 runs | 40%9.5 of 24 | 0.5 of 11 | 110 s | $0.279 |
| + Graphifya graph of the repository | 0 of 10 runs | 33%8 of 24 | 0 of 11 | 97 s | $0.300 |
| + Obsidian and Graphifyboth of the above | 0 of 10 runs | 33%8 of 24 | 0 of 11 | 111 s | $0.299 |
| + Miosthe team’s memory over MCP | 5 of 11 runs | 84%20.2 of 24 | 9.3 of 11 | 108 s | $0.395 |
Off-repo rules followed: the average over the runs, out of the 11 rules found nowhere in the repository. Against Claude Code alone, the first answer with Mios costs 48% more per task and takes 19% longer. By our bars the time gap is noise; the cost gap is not.
Until the code is right.
Every setup ended with all tests passing. Here is what it took for the five tasks, each done once, which is what a team pays. Under each figure, the gap with Claude Code alone.
| Setup | Prompts | Time | Claude Code tokens | Claude Code cost | Mios server cost | Total cost |
|---|---|---|---|---|---|---|
| Claude Code alonethe ticket and the repository | 16.5reference | 954 sreference | reference | $2.241reference | no server | $2.241reference |
| + Obsidianthe repository’s docs as notes | 17.5+6% | 1,029 s+8% | +10% | $2.352+5% | no server | $2.352+5% |
| + Graphifya graph of the repository | 17+3% | 1,064 s+12% | +9% | $2.330+4% | no server | $2.330+4% |
| + Obsidian and Graphifyboth of the above | 17+3% | 1,055 s+11% | +17% | $2.423+8% | no server | $2.423+8% |
| + Miosthe team’s memory over MCP | 9−45% | 675 s−29% | −18% | $2.034−9% | $0.436 | $2.471+10% |
Claude Code tokens and cost are read from each session. The Mios server cost is the model calls the Mios server made, measured; the other setups call no server. Per task, the median tells the same story: 1.5 prompts for Mios against 3 for Claude Code alone, 132 s against 184 s, and $0.404 against $0.403 in total cost.
Where it helps, what it costs.
Where Mios helps
- Correct code. On the first answer, Mios passed 84% of the checks (20.2 of 24). The best other setup, with Obsidian, passed 40%. Mios followed 9.3 of the 11 off-repo rules on average; no other setup followed more than 0.5.
- Fewer rounds. 9 prompts for the five tasks, against 16.5 for Claude Code alone: 45% fewer.
- Less time. 675 s until correct code, against 954 s for Claude Code alone: 29% less.
What it costs
- The first answer costs more: $0.395 per task against $0.266 for Claude Code alone, 48% more at the median.
- Until correct code, Claude Code itself spends 9% less with Mios. Add the Mios server and the total is 10% higher: $2.471 against $2.241 for the five tasks.
- 98% of the server cost is the edit guard, a check that runs on Claude Sonnet before the agent writes. That is where the money goes, and the part we need to make cheaper.
A clear win on getting the code right. With Mios, about half the first answers (5 of 11) were fully correct, against none for every other setup. Reaching correct code took 45% fewer prompts and 29% less time. Cost is on par, not lower yet: to correct code, Claude Code itself spends 9% less with Mios, and the Mios server brings the total to 10% more than Claude Code alone, within the 25% noise we set before the runs.
What this bench does not show.
- 01Short tasks
A first answer took 90 to 110 s at the median, and correct code 2 to 4 minutes. Long tasks were not measured.
- 02Tuned on the same sources
Most of the gap comes from tasks whose source conversations were used to tune Mios: three tasks carry 13.5 of the 14.7 points between Mios and Claude Code alone. The bench shows that Mios serves well the facts it ingested well. It does not show that ingestion works on conversations it has never seen.
- 03Few runs
One to three runs per cell. A cost or time gap under 25% is noise and settles nothing.
- 04What Mios misses
Mios misses one rule in every run of one task, and its result varies from run to run on another. We have not yet found out whether it fails to find the rule, or finds it and the agent does not follow it.
- 0511 rules, not 13
Two of the 13 rules written outside the repository do not separate the setups: one passes in every run, Claude Code alone included, and the other can be reached through data already in the repository. We count the 11 that do.
- 06A cap changed mid-bench
The prompt cap went from 4 to 8 during the bench, and each run keeps its own. With the original cap of 4, the median number of prompts does not change. After an outside review, the review loop is counted per task, each task once, rather than per run; the verdict is the same either way.
Runs of November 12 and 13, 2026. Mios measured: the build of November 12, 2026, commit 251d8392. Claude Sonnet 5.5 for every setup. Only what we measured is on this page, no figures from other vendors. How Mios compares, feature by feature
Give your team a memory.
Your people ask it. Your agents read it before they edit.
Mios opens to a few teams first. Book a call and we’ll get yours in.