Anthropic AI Updates: September 9, 2026

1. Anthropic Published Hard Numbers on Prompt Anti-Patterns That Inflate Token Spend

Anthropic. A Claude Platform post documents specific prompting habits that cost money without improving results, drawn from Anthropic’s own migration from Opus 4.8 to Opus 5. Verification rituals such as “verify twice” duplicated tokens, and emphasis boosters such as “be maximally thorough” triggered tool calls that were not needed; removing them produced a 14.6% cost reduction and a 5.3% accuracy improvement on the same workload. The post also reports effort-level calibration data, with Claude Fable 5 at max effort scoring 30.9% at $19 per task against 11.5% at $5.35 per task, and Fable 5.1 at low effort matching Fable 5 at high effort for a third of the cost.

Tooling shipped alongside it: /claude-api prompt-audit finds cache-breaking prefix divergence, /claude-api hillclimb searches model and effort combinations, and /claude-api cost-optimize profiles token spend and proposes cuts, with reported savings of 58% on LegalBench, 73% on tau2-bench, 55% on SWE-bench, and 52% on OfficeQA Pro. The platform additions are a 1-hour cache TTL, mid-conversation system messages that preserve cache integrity, a defer_loading option that keeps rarely-used tools out of the cached prefix, and automatic cache breakpoint placement for growing conversations. Source

2. Claude Code 2.1.265 Adds Multi-Plugin Directories and an MCP SSE Fallback

Anthropic. The release lets --plugin-dir point at a folder containing multiple plugin subfolders, with plugins loaded and unloaded dynamically at runtime rather than only at startup. On the protocol side, Claude Code now falls back to SSE as the MCP spec describes when a server is HTTP-only, which fixes connections against servers that never implemented streamable HTTP. Other additions include a 1 GB cap on tool results written to disk with truncation notices in conversation previews, streaming of forked skills’ kickoff prompts and text turns as stream-json progress events, automatic archiving of inactive VS Code sessions after 14 days, and plugin discovery that reads display names and descriptions from marketplace metadata when a plugin ships none of its own. A follow-up 2.1.266 reverted a gateway regression in which CLAUDE_CODE_USE_GATEWAY forced cloud-gateway sign-in; it is now ignored unless both ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN are set. Source