Developers are choosing older AI models, and the data explains why (www.augmentcode.com)

Augment Code’s production telemetry shows developers increasingly pick models by task profile rather than version — older models aren’t being abandoned, they’re being repurposed. Over a week in early October 2025 Sonnet 4.5’s request share fell from 66% to 52% while Sonnet 4.0 rose from 23% to 37%; GPT‑5 held steady around 10–12%. Rather than a straight upgrade path, teams are assembling “model alloys”: Sonnet 4.5 for deep multi-file reasoning and autonomous planning, Sonnet 4.0 for deterministic, tool-centric automation, and GPT‑5 for explanatory fluency and documentation. The data reveal measurable behavioral and system-level trade-offs. Sonnet 4.5 issues fewer tool calls per user message (12.33 vs 15.65 for 4.0) but produces ~7.5k tokens per interaction vs ~5.5k for 4.0 (≈37% more), suggesting heavier internal reasoning and longer contexts. That reasoning increases latency and cache reads (Sonnet 4.5 ≈240B reads vs 135B for 4.0), shifting compute from raw token generation to context management and retrieval-augmented workflows. The implications for builders: choose models for cognitive style (depth vs action), optimize cache and throughput for long-context workloads, and consider ensembles or routing logic rather than a single “best” model. Ongoing work will test whether deeper internal reasoning materially improves completion success rates.