Sonnet 5 — Speed-Value Protocol

Previous generation. Sonnet 5.5 (sonnet-5-5) succeeds it at the same price with breaking API changes — prefer that skill for new work; use this one for projects pinned to claude-sonnet-5.

Model ID: claude-sonnet-5 · 1M context · 128K max output · $2 in / $10 out per MTok (live pricing table, 2026-09-25; the earlier $3/$15 base with intro pricing through 2026-08-31 is no longer listed — re-verify before cost-critical decisions). Near-Opus quality on coding and agentic work at a fraction of the cost — the best default when Opus-tier depth isn't required.

Staleness guard

Model facts here (ID, pricing, limits, API rules) were verified 2026-07. Intro pricing expires 2026-08-31. If a newer Claude model generation exists or the API contradicts this file: verify against current docs (Models API / platform.claude.com or the claude-api skill), follow the live source, and log the correction in learnings.md.

Self-improvement (do this first and last)

1. At start: read learnings.md in this skill's folder if it exists. Apply relevant lessons.
2. At end of every use: append one dated bullet — what worked, what wasted tokens, what you'd prompt differently. Merge instead of duplicating; delete disproven bullets.

Hard API rules

• Adaptive thinking is ON when thinking is omitted (opposite of Opus 4.8). Set {type: "disabled"} explicitly only for latency-critical simple paths — prefer adaptive + effort: "low".
• budget_tokens → 400. Non-default temperature/top_p/top_k → 400. Assistant prefill → 400.
• New tokenizer: ~30% more tokens than Sonnet 4.6 for the same text — re-baseline max_tokens and budgets with count_tokens.
• thinking.display defaults to "omitted" — set "summarized" if you show reasoning.

Where it shines

• Interactive coding sessions, medium-complexity agent loops, high-volume pipelines where Opus cost doesn't pay for itself.
• First Sonnet with xhigh effort and high-res vision (2576px) — use xhigh for its hardest coding tasks.
• Rough mapping: Sonnet 5 medium ≈ Sonnet 4.6 high; Sonnet 5 high ≈ Sonnet 4.6 max.

Prompting patterns that matter on Sonnet 5

• It is very literal. It won't generalize an instruction across items or infer unstated requests. State scope explicitly ("apply to every section, not just the first"). Re-baseline holdover style directives — they now apply at face value.
• More agentic by default — reaches for tools and self-verification readily. With thinking disabled it's less tool-eager; add an explicit tool-triggering nudge if you disable thinking.
• Drop forced progress scaffolding — its native interim updates are good.
• Effort over prompting: if reasoning is shallow at low/medium, raise effort rather than adding "think carefully" prose.
• Design variety: no temperature — use "propose 4 distinct visual directions, pick one" instead.
• Code review: report-everything-filter-downstream, same as Opus.

Token discipline

• Leave max_tokens headroom at xhigh/max — adaptive thinking shares the output budget; stop_reason: "max_tokens" with a wall of thinking means the budget was too tight.
• Cache-first layout; verify cache_read_input_tokens.
• Route true fan-out grunt work down to Haiku 5.5; route work Sonnet repeatedly fails at up to Opus 4.8 rather than retrying.