Anthropic AI Updates: September 23, 2026

1. Opus 5.5 Claims Fable 5.1 Performance at 40 Percent Lower Cost

Anthropic. Claude Opus 5.5 is the first model in the 5.5 family, and the pitch is efficiency rather than a new capability ceiling: Anthropic says it performs at the level of Claude Fable 5.1 on most work while costing about 40 percent less to run than Opus 5. Published numbers include 66.4 percent on Terminal-Bench 4.0, 54.4 percent on FrontierCode v1.1, 57.8 percent on CursorBench 4.0, 1846 Elo on GDPval-AA v2.1, 81.8 percent partial success on OSWorld 2.0, and 40.0 percent pass on AutomationBench. Source

2. The Price Cut Is Larger on Cache Reads Than on Tokens

Anthropic. Input drops to $4 per million and output to $20 per million, both down 20 percent from Opus 5, but cache reads fall 60 percent to $0.20 per million with cache writes at $5 per million. That asymmetry favors long-running agent loops that replay a large stable prefix, where cache reads dominate the bill. Anthropic also claims roughly 30 percent faster output generation and fewer tokens consumed per task, and one customer reported output about 40 percent less verbose. Source

3. Anthropic Says It Fixed the “Claudish” Writing Voice

Anthropic. Alongside the benchmark gains, the release targets the model’s prose: clearer and more concise responses that lead with the essential information. Anthropic reports Opus 5.5 scored highest on its automated behavioral audit, with improved prompt injection resistance, new biology and cybersecurity safeguards, and preserved anti-distillation measures on thinking output. EU AI Act watermarking compliance ships with it. Source

4. Anthropic Is Building a Wet Lab Where Claude Directs the Robots

Anthropic. Reuters reports Anthropic is standing up its own biology lab in the San Francisco area so AI-driven drug development can run physical experiments rather than simulations alone, with Claude guiding robots through protocols under human safety oversight. Two shipped pieces make it plausible: Claude Science as the researcher workspace, and the Model Hardware Standard for controlling lab equipment. Eric Kauderer-Abrams, who runs the life sciences division, points at diseases considered undruggable and at multi-target antibody design. The lab builds on April’s roughly $400 million acquisition of Coefficient Bio. Source

5. Claude Code 2.1.280 Makes Opus 5.5 the Default Opus Model

Anthropic. The release switches the default Opus model to claude-opus-5-5 with 1M context at $4/$20 per Mtok and $0.20/Mtok cache reads. It also adds CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH to lift the 2,048-character cap on MCP tool descriptions and server instructions for every server in a session, extends mouse support to the /skills list and /plugin skill state options in fullscreen mode, and reports hook output sizes plus the count of oversized outputs saved to file in the hook_execution_complete OpenTelemetry event. Source