5G exits for IPv6 and latency-critical mobile eval

What 5G actually changes

Most "5G proxy" marketing is 4G LTE with a 5G SIM slapped on a non-standalone network. There's a real distinction worth making:

• 5G NSA (non-standalone). Uses the LTE core with a 5G radio. Latency behaves like 4G. Most "5G" plans in the US, UK, and large parts of Europe are NSA as of early 2026.
• 5G SA (standalone). Uses the 5G core. Sub-30ms RAN latency is achievable, edge compute integrations are possible, and network slicing is real. Commercial SA coverage is patchy but expanding.

SquadProxy runs both. When you request 5G exit, you can pin to SA where the carrier supports it.

When 5G is the right AI tool

• Latency-critical eval. Voice AI, real-time agent eval, and streaming model-response benchmarks where the mobile-path latency matters to the measurement.
• IPv6 path testing. A growing fraction of 5G traffic is IPv6-primary. If your eval targets advertise different content or apply different routing on IPv6 (some mobile-first social platforms do), you need a real IPv6 mobile origin.
• Mobile-first model agent evaluation. Evaluating how an AI agent behaves on a mobile network-typical path — retries, TCP behaviour, TLS resumption — from within that network.

Where 5G is wrong

• SquadProxy collection at any meaningful scale. Mobile bandwidth economics make 5G unviable for TB-scale training-corpus work.
• Cost-sensitive long-session eval. 5G sessions cost roughly the same per minute as 4G. The premium goes on latency. If your eval is tolerant of 100–200ms, use 4G or ISP.

Coverage

• US — Verizon and T-Mobile 5G SA in major metros. AT&T 5G NSA.
• UK — EE 5G SA (where deployed), Vodafone 5G, Three 5G.
• Germany — Deutsche Telekom 5G SA, Vodafone DE 5G.
• Japan — NTT Docomo 5G SA, KDDI au, SoftBank, Rakuten.
• Singapore — Singtel 5G SA, StarHub 5G.

Other markets are 4G-only for now. We'll expand as SA rollouts reach stable commercial coverage.