Frontier models

Claude Sonnet 5

Speed-and-intelligence balance for everyday coding and agents.

Training

Reliable knowledge and training data cutoff: January 2026. Parameter count, FLOPs, dataset mix, and any RLHF / RLVR / RLAIF recipe tied to this SKU: UNKNOWN.

Inference

Adaptive thinking. Default effort is high. Latency class is fast relative to the current Claude lineup.

Context

1M token context window. 128K max output tokens. Modalities: text and image in, text out; multilingual; vision; tool use.

Methods

Effort parameter and tools are what Anthropic names for this SKU. Exact train-time RL recipe at SKU level: UNKNOWN.

Benchmarking

UNKNOWN. No lab-published score table for Claude Sonnet 5 appears in our sources (models overview and the Fable announce table cover Fable / Opus columns, not Sonnet 5). Do not invent scores.

Hardware

Anthropic trains and runs Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs. Amazon is the primary cloud and training partner. More than 1 million Trainium2 chips in use as of April 2026. Per-SKU serving silicon: UNKNOWN. Buyers purchase API tokens, not a named Claude GPU SKU.