Claude Sonnet 5

Claude Sonnet 5 is my default model. It handles roughly two-thirds of my daily work, typically 400–600 API requests a day across local automation and IDE tooling: multi-step debugging in existing codebases, agent loops that run terminal and browser tools, writing complete pull requests, drafting architecture, and working through large codebases or document sets in its 1,000,000-token context window.

The case for it is economics. At $2 per million input tokens and $10 per million output, it costs about 40% of Opus 4.7 per token, and in my experience delivers around 90–95% of Opus's agentic reasoning for everyday engineering work. On a typical workload of 10,000 requests a month at 1,500 input and 500 output tokens, that's about $80 a month. Cached input at $0.20 per million makes long, repeated system prompts and codebase context much cheaper.

Two things make it better than the alternatives I use in agent loops. First, adaptive thinking: it scales its reasoning to each turn's difficulty without the budget tuning older models needed, which keeps end-to-end latency down without hurting tool accuracy. Second, follow-through. It sticks to a plan and verifies its own work. When a terminal command or test fails silently, it will re-read the logs, write a regression test that reproduces the bug, or stash changes to isolate the cause, all unprompted, rather than guessing. GPT-5.5, by comparison, sometimes stops a multi-step task early.

Two gotchas if you're migrating. The tokenizer is noticeably denser: in my usage, the same text comes out at roughly 30% more tokens than on Sonnet 4.6, so a low per-token price can still surprise you if your context budgets were sized for the older model. And the old controls are gone: setting temperature, top_p or a fixed thinking budget_tokens now returns a 400 error. Depth is controlled with the effort setting instead, from low for quick tasks up to max.

I deliberately switch away in two situations. For hard mathematical proofs, large refactors across heavy legacy code, and high-stakes security or production-safety reviews, I escalate to Opus 4.7 or 4.8. For high-volume, latency-sensitive work such as simple RAG lookups, bulk translation or real-time streaming, where cost per token has to be negligible, I use Gemini 3.8 Flash or Kimi K3.

Bottom line: for developers and technical founders who need reliable, autonomous coding and tool-calling agents without an enterprise-sized API bill, Sonnet 5 is the default to start with.