Anthropic ships Claude Sonnet 5, a mid-tier agentic model at $2/M input scoring 63.2% on agentic coding

Claude Sonnet 5 slots into Anthropic's mid-tier as an agentic workhorse. The headline numbers: $2 per million input tokens and a 63.2% agentic-coding score, a meaningful jump over Sonnet 4.6's 58.1%. Anthropic positioned it as the price-performance option for production agent workloads rather than a frontier flagship.

Distribution was fast. Microsoft's Foundry platform brought Sonnet 5 (and restored Fable 5) to GA on Azure with both global and US data-zone options inside Microsoft's governance boundary, which the community praised as 'a game-changer for production reliability.' That speed of enterprise availability is itself a competitive signal in a market where deployment friction, not raw capability, is often the bottleneck.

There's a catch developers surfaced quickly: Sonnet 5's new tokenizer reportedly emits 1.0-1.35x more tokens than its predecessor, meaning the sticker price doesn't fully capture migration cost. Teams moving existing prompts over reported surprise increases in effective spend — a reminder that per-token pricing comparisons across model generations can mislead.

Competitively, Sonnet 5 is Anthropic's answer to a crowded mid-tier where DeepSeek-V4-Flash ($0.28/M output) and open-weight Chinese models are undercutting Western pricing aggressively. At $2/M input, Sonnet 5 is far from cheapest, so Anthropic is betting on coding quality and enterprise trust. What to watch: independent agentic-coding benchmarks and whether the tokenizer overhead erodes the value proposition against cheaper rivals.