Networking - LLM Comparison

Comparison of LLMs with networking tasks


Project maintained by gtom84 Hosted on GitHub Pages — Theme by mattgraham

Prompt

Create a working, fully functional production grade Cisco IOS-XE 17 configuration of a router ISR8000 which can be copy pasted into a production device.

Properties

Rating

Add +1 point for:

Extra credit for identfying configuration issues, ie. IPv6 syslog server without any IPv6 routing, address, etc.

Result

Score Summary Table

Model IP/VLAN BGP/Traffic Eng. OSPF NAT Shaping Sec/AAA Total
GPT-5.3-chat 1 0.5 1 0 0 1 3.5
GPT-5.5 1 1 1 1 1 1 6.0
Opus 4.7 1 1 1 0 1 1 5.0
Qwen 3.6-Plus 1 1 0.5 0 0 1 3.5
Gemini 3.1 Pro 1 1 1 0 1 1 5.0
Gemma 4 31B IT 1 1 1 0 0 1 4.0
Sonnet 4.6 1 1 1 1 1 1 6.0

Detailed Evaluation

1. Sonnet 4.6 & GPT-5.5 (The Winners)

These two models provided the most complete “copy-paste” production configurations.

2. Gemini 3.1 Pro & Opus 4.7

3. GPT-5.3-chat & Qwen 3.6-Plus


Critical Configuration Identification (Extra Credit)

Most models correctly identified the IPv6 Syslog Paradox:

The prompt asks for a SYSLOG server at 2000:DEAD:BEEF::1, but the provided interface requirements are strictly IPv4. Without an IPv6 address on the WAN/LAN or a tunnel, the router cannot route the syslog packets.

Sonnet 4.6 and GPT-5.5 were the most vocal about this, with GPT-5.5 explicitly adding a “Note” that IPv6 reachability must be established for the syslog to function.

Common Miss: Only Sonnet 4.6 and GPT-5.5 addressed the Asymmetric WAN2 (100 Down / 20 Up). In Cisco IOS, the bandwidth command is mostly for metric calculation; to actually respect the ISP’s 20Mbps policing on the upload, a service-policy with a shape average 20000000 is mandatory to prevent packet drops.