Claude Sonnet 5

PROVISIONAL CALIBRATION per /NM310-A1 (2026-09-07) — 0 BAND / 0 BAND-COMPOSITE / 19 INHERITED-PROVISIONAL / 1 PROXY; re-calibrate when per-dimension evals land. Released 2026-06-30; Anthropic's most agentic Sonnet, default model for Free + Pro plans. Model ID claude-sonnet-5. Context 1M default (also max — no smaller variant), 128K max output, adaptive thinking on by default (manual extended thinking removed; sampling params locked to defaults). NEW TOKENIZER: ~30% more tokens for the same text — per-token price unchanged vs 4.6 but equivalent-request cost differs. Pricing: $2/$10 intro through 2026-08-31, then $3/$15. Positioned near Opus 4.8 at ~40-60% lower cost: SWE-Bench Pro-class agentic coding 63.2% (Opus 4.8: 69.2%, Sonnet 4.6: 58.1%); slightly beats Opus 4.8 on a knowledge-work benchmark. First Sonnet-tier model with real-time cyber safeguards on by default (refusals return HTTP 200 + stop_reason 'refusal'). System card: much lower cyber capability than Opus-class (Firefox exploit dev 0.0% working, vs Opus 4.8 68.8%, Mythos 5 88.4%); lower hallucination + sycophancy than 4.6 but somewhat higher misalignment rates than Opus 4.8. Parent Model → Sonnet 4.6. PROVENANCE (-A1 lanes M1/M1b, per-dim): Cost Efficiency [PROXY:AA $6.00/M blended]; all remaining 19 dimensions [INHERITED-PROVISIONAL:claude-sonnet-4-6]. SWE-Bench Pro 63.2% is on file but its Benchmarks Calibration Stance excludes it from banding (), so Code Generation stays inherited.