Anthropic dropped the cache-read price from $1.00 to $0.25 per million tokens for both Fable 5.1 and Mythos 5.1, per Anthropic's release page. The cut makes prompt-caching substantially more attractive for applications that reuse long system prompts or document context across many requests. Model capabilities are otherwise unchanged.

Meta's new speech model handles streaming transcription, speaker diarization, and voice-activity endpointing in a single pass, priced at $3 per 1,000 minutes per Meta AI Research. The unified architecture targets real-time meeting transcription and call-center workloads that currently chain three separate models.

HydraFusion is a new Copilot routing layer that dispatches each task to the cheapest capable model in real time. GitHub's offline evaluation shows 67% cost reduction compared with routing every request through Anthropic's Opus 5, with no measured quality regression on the benchmark suite.

Cursor's new self-hosted machine pools let enterprise teams run Cursor's cloud agents on their own infrastructure, keeping code and context within their network perimeter. The feature is aimed at organizations with data-residency requirements or security policies that prohibit sending source code to external servers.