Claude Sonnet 5

Claude Sonnet 5 is the mid-tier model of Anthropic's Claude 5 family and, from launch (July 2026), the default model on the Claude Free and Pro plans. Like earlier Claude models it is a decoder-only transformer built through large-scale pretraining followed by instruction tuning and preference optimisation (RLHF / RLAIF), with a training-data cutoff of January 2026. It is served with a 1M-token context window at standard pricing.

Anthropic positions Sonnet 5 as offering near-Opus-4.8 quality on most tasks at Sonnet-tier cost, with the recommended path being to step up to Opus 4.8 for the hardest agentic or reasoning work and drop to Haiku 4.5 for latency- or cost-sensitive, high-volume work.

Two release-history facts matter for cost planning. First, on 2026-08-10 Anthropic made the $2 / $10 per-million-token pricing permanent, cancelling a previously scheduled step-up to $3 / $15 — so this is the standing price, not an introductory rate. Second, Sonnet 5 uses an updated tokenizer (shared with Opus 4.7 and later) that produces roughly 30% more tokens than Sonnet 4.6 for the same text; per-task cost comparisons against pre-tokenizer-change Claude models need to account for that.

As of Modelglass's 2026-08-19 capability review, the coding rating has independent corroboration (Vals AI's SWE-bench Verified leaderboard scored it ~75%, below Anthropic's system-card figure of 85.2%), while the reasoning, tool-use, and other ratings remain vendor-reported with no independent cross-check found. Modelglass records the overall capability confidence for this entry as medium.