grok-claude-bridge · performance report
Agent speed on the Oracle VM
How fast the bridge and its agents answer after moving to a larger Oracle machine, next to the first measurements.
Summary
The bridge is not a bottleneck: it adds about 2 ms. Claude answers a one-word prompt in 3.7 s and a 250-word reply in 12.7 s. Several Claude jobs at once overlap well, and Claude and Grok together do not slow each other. Grok's speed changes from run to run (10 to 29 tok/s when run directly), so the ~65 tok/s measured on 13 Oct was a good moment, not a constant. Codex is still over its usage limit.
Output speed (approx. tokens per second)
Scale 0 to 100 tok/s. The faded 13 Oct bars came from longer replies (a Claude file write and a 450-word Grok essay), which spread the fixed start-up cost over more tokens, so compare 14 Oct with 14 Oct.
Latency and concurrency (14 Oct)
| Test | Result |
|---|---|
| Bridge overhead (30 calls, no agent) | 1.7 ms median, 2.2 ms p95 |
| Claude, one-word reply (3 runs) | 3.7 s median (3.7 to 3.8) |
| Grok, one-word reply (3 runs) | 7.1 s median (5.4 to 16.9) |
| 3 Claude jobs at once | 8.5 s in total, CPU 54% |
| 4 Claude jobs at once | 8.6 s in total, CPU 62% |
| Claude and Grok together | 6.7 s in total (3.9 s and 6.7 s) |
Four jobs at once cost no more than three, and memory stayed at about 22 GB free throughout.
Grok head to head
| Run | Reply | Time | Words | Speed |
|---|---|---|---|---|
| 13 Oct · VM 1.0.52, direct | 400-word essay | 15.8 s | 456 | ~65 tok/s |
| 13 Oct · Mac 1.0.50, direct | 400-word essay | 22.3 s | 409 | ~35 tok/s |
| 14 Oct · VM 1.0.52, direct, run 1 | 400-word essay | 22.6 s | 476 | ~28 tok/s |
| 14 Oct · VM 1.0.52, direct, run 2 | 400-word essay | 53.9 s | 422 | ~10 tok/s |
| 14 Oct · VM 1.0.52, direct, run 3 | 400-word essay | 18.4 s | 404 | ~29 tok/s |
| 14 Oct · VM 1.0.52, via bridge | 250-word essay | 75.6 s | 255 | ~5 tok/s |
The slow Grok runs are the provider's variance, not the machine or the bridge: the VM reaches the Grok service in a few tens of milliseconds, and the same CLI ran from 18 s to 54 s for the same prompt.
Status per agent
| Agent | Version | State | Notes |
|---|---|---|---|
| Claude Code | 2.1.299 | OK | 3.7 s for a one-word reply, 12.7 s for a 250-word reply; several jobs at once overlap. |
| Grok Build | 1.0.52 | Variable | Works, but its speed varies a lot from run to run; benchmark it several times before drawing conclusions. |
| Codex | 0.162.1 | Limited | Usage limit still hit; the quota fallback hands its jobs to Claude. Its sandbox works on the new VM (verified). |
Method
- Run on the VM against the bridge on its own loopback address, so the network and the proxy in front are not counted.
- Baseline: the reply to "reply with the single word pong", the fixed start-up cost of each agent.
- Speed: (essay time − baseline) ÷ estimated tokens, at about 1.33 tokens per word. Through the bridge both agents write shorter essays (about 250 words) than when run directly, so short replies carry more of the start-up cost.
- Single runs unless a count is given. Treat the numbers as rough; Grok in particular varies a lot.
What changed
- The bridge moved to a larger Oracle machine (4 cores, 24 GB, shared with other services) and the old 1-core machine was retired.
- The VM's keep-alive load now tops the CPU up only when the machine is idle and steps aside for real work: a fixed CPU job took 1.76 s with it running and 1.74 s without, where the old fixed version cost about 25%.
- Codex's sandbox was re-enabled on the new machine with a profile that allows user namespaces for
/usr/bin/bwraponly.
Static report: the numbers are recorded by hand from the 14 Oct 2026 measurements and are not live. Nothing on this page reads from the bridges.