Claude Code 2.1.316 sunsets Opus 4.8 fast mode: move to 5.5 by Nov 30
The same build adds a per-subagent effort dial — while two fresh open-weight drops, Nanbeige4.3-33B and Baidu one-pass OCR, give you more to self-host.
Copy markdown
Fast mode retires Nov 30 — repin to Opus 5.5 now
Claude Code 2.1.316 (Oct 31) deprecated Opus 4.8 fast mode, with full removal on November 30. Hunt down any scripts, CI jobs, or hooks that pin Opus 4.8 by name and switch them to Opus 5.5 while both still run, so you can diff latency and output before the cutoff rather than after.
The same build adds a subagent effort dial
A new CLAUDE_CODE_SUBAGENT_EFFORT env var sets how hard subagents reason independently of the parent session — dial it down for cheap fan-out like checking many files, up for hard investigation. The build also lands CLAUDE_CODE_DISABLE_ULTRATHINK, pooled group budgets, and renames the Max 20x plan to "Ultra"; 5 improvements and 34 bug fixes in all.
Nanbeige4.3-33B: an open agent coder that tops Qwen
BOSS Zhipin's lab dropped a dense 33B model under Apache 2.0 with a 512K context window. Lab-reported scores beat Qwen3.8-27B on Terminal-Bench (80.9 vs 73.0) and DeepSWE (65.5 vs 42.5) and lead 17 of 20 tests; a community 4-bit build runs in ~22GB on a single card at ~64 tok/s. Numbers are vendor-reported and not yet independently reproduced.
Baidu's Unlimited-OCR-2 parses any-length PDFs in one pass
MIT-licensed open weights with Sparse Reference Sliding Window Attention that holds decode memory near-constant as a document grows — Baidu's demo parsed all seven Harry Potter books (3,600+ pages) in a single run, each step ~7x faster than full attention. Self-host via SGLang or vLLM to feed RAG and document pipelines without per-page stitching.