Insights
Real telemetry from a real AI workforce.
amoOS measures itself: sessions, dispatches, tokens per model, cache hits, merged PRs, interventions and subscription capacity, rolled up every hour. The numbers below are straight from those rollups.
920
Agent sessions
951
Tasks completed
342
PRs merged
12
Models in use
3
Machines
9
Weeks measured
Throughput
Weekly throughput
Tasks completed and pull requests merged per week
Agent sessions by engine
Sessions started per week across the fleet
View the weekly data as a table
| Week of | Tasks | PRs merged | Claude sessions | Codex sessions |
|---|---|---|---|---|
| 27 Jul | 36 | 11 | 18 | 0 |
| 3 Aug | 62 | 41 | 41 | 2 |
| 10 Aug | 106 | 19 | 28 | 7 |
| 17 Aug | 174 | 48 | 206 | 17 |
| 24 Aug | 47 | 13 | 29 | 12 |
| 31 Aug | 103 | 26 | 56 | 1 |
| 7 Sep | 119 | 35 | 58 | 0 |
| 14 Sep | 113 | 6 | 135 | 7 |
| 21 Sep | 157 | 102 | 193 | 16 |
Last 30 days
Rolling 30 days to 29 Sep 2026, compared with the 30 days before.
566
Dispatches
+138%
210
PRs merged
+74%
526
Tasks completed
+30%
134
Phone decisions
answered remotely
98%
Prompt-cache hit rate
9.18 billion tokens in 30 days, and 98% of input came from cache instead of being paid for again.
7.8%→2.6%
MCP tool error rate
Down by two thirds in a month. Each failure the fleet reports becomes a tracked gap with evidence and then a fix.
Model mix
Twelve models from two vendors ran in September. Routing follows difficulty: the strongest models coordinate and review, faster ones handle routine work, and Codex takes bulk.
Tokens by model
September 2026 · every model the fleet ran, measured per session
Claude Opus 5
Anthropic · Coordination, deep work
4.57B
Claude Opus 5.5
Anthropic · Hard tasks, review
2.75B
Claude Fable 5.1
Anthropic · Frontier experiments
716M
GPT-6 Astra
OpenAI · Codex bulk execution
397M
Claude Sonnet 5
Anthropic · Standard tasks
340M
Claude Opus 4.7
Anthropic · Legacy sessions
204M
GPT-5.6 Sol
OpenAI · Codex routine work
119M
Claude Sonnet 5.5
Anthropic · Mechanical tasks
52M
Codex auto-review
OpenAI · Automated review
6M
Claude Haiku 4.5
Anthropic · Utility calls
4M
Hard
Opus 5.5
Coordination, architecture, security-sensitive changes and every review.
Standard
Sonnet 5.5
Well-specified features, tests and mechanical refactors.
Bulk
Codex (GPT-6 Astra, GPT-5.6 Sol)
Parallel batch work on a separate subscription pool, which keeps Claude capacity free.
Utility
Haiku 4.5 · local models
Classification, summaries and background jobs at near-zero cost.
Capacity
Claude and Codex usage is probed continuously. The coordinator paces work against weekly limits and routes bulk to the separate Codex pool, so paid capacity gets used instead of expiring.
Where the work runs
Share of agent sessions per machine since late July
amonodeAlways-on Linux engine room
60% · 548
amom4proMacBook cockpit
22% · 202
amominiMac mini engine room
18% · 170
100%
Peak weekly subscription use
The fleet reached its weekly Claude limit this month. That’s the goal: capacity turned into shipped work.
848
Claude Code sessions
72
Codex sessions
Aggregates only, taken from amoOS’s own metric rollups on 29 Sep 2026. Project names, prompts and client data are excluded.