---

title: "Anthropic weekly watch (March 16-23, 2026)" type: "weekly-watch" date: "2026-03-23" score: 3 action: "partial-integration" sources:

• "Perplexity research synthesis (secondary source — no direct URL)"
• "https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md"
• "https://platform.claude.com/docs/en/release-notes/overview"
• "https://www.anthropic.com/81k-interviews"

---

Anthropic weekly watch (March 16-23, 2026)

Summary

Perplexity synthesis covering five event clusters in the March 16-23 window: Claude Code 2.1.81, Python SDK v0.85.0/v0.86.0, two Platform API release notes entries (thinking.display + model capabilities fields), and the "What 81,000 people want from AI" research blog.

Score: 3/5

Justification: Secondary source (Perplexity synthesis, not primary docs). CC 2.1.81 is already tracked in our releases YAML/MD. The real value is two undocumented API platform features (thinking.display: "omitted" + model capabilities fields) that belong in the extended thinking section. Python SDK changes are out of scope for a Claude Code CLI guide. The 81k study is too far from the guide's technical audience.

Gap Analysis

Integration Recommendations

P1 — thinking.display: "omitted" (Platform API, March 16)

Where: guide/ultimate-guide.md — Extended Thinking section (search for "extended_thinking" or "thinking budget")

What to add: Small subsection or note explaining:

• thinking.display: "omitted" hides the thinking text in the response while preserving the signature field for multi-turn continuity
• Use case: faster streaming, cleaner UX when thinking content isn't needed by the end user
• Billing unchanged: charged at full thinking token cost regardless of display setting
• Tradeoff: internal auditing or labeling workflows that rely on thinking text must use a different display mode

P2 — Model capabilities API fields (Platform API, March 18)

Where: guide/ultimate-guide.md — API Integration section or a new "Model Discovery" note

What to add: GET /v1/models and GET /v1/models/{model_id} now return max_input_tokens, max_tokens, and a capabilities object. This enables dynamic model selection in orchestrators instead of hardcoded capability maps.

P1 — --bare flag scripted mode tradeoffs (missed in initial pass)

Where: guide/ultimate-guide.md — scripted / CI/CD usage section (search for -p flag, headless, pipeline)

What to add: --bare vs standard -p tradeoff table. --bare skips hooks, LSP, plugin sync, and skill directory walks; requires ANTHROPIC_API_KEY or apiKeyHelper (no OAuth). Fast and minimal for CI scripts, but loses all observability that hooks provide. Team environments needing audit trails should use standard -p instead.

Fact-Check

Note on fact-check scope: The synthesis is a Perplexity digest of release notes and blog posts. The CC 2.1.81 claims are verifiable against our tracked YAML. The API platform features need cross-checking against https://platform.claude.com/docs/en/release-notes/overview before integration into the guide.

Challenge (technical-writer)

Score maintained: 3/5 (two API gaps + one missed CLI gap don't clear the 4/5 threshold of "major integrable improvement")

Key corrections from challenge:

• thinking.display: "omitted" is more significant than flagged. Guide covers extended thinking from CLI perspective only (lines 13539-13632). Multi-turn API patterns with chain continuity are entirely absent. Relevant to orchestrator builders. Shifts priority from "gap" to "absent concept."
• Model capabilities API is underweighted. GET /v1/models with capabilities object enables runtime-dynamic model selection in multi-agent routing — a pattern the guide documents as manual-only today. Not a metadata footnote; it is a pattern shift.
• --bare flag scripting implications missed in initial pass. Releases tracking says "covered" but guide does not document --bare vs standard -p tradeoffs for CI/CD. Added as P1.
• Python SDK: skip, not P3. Guide has always scoped away from the Python SDK. No hedge needed.
• 81k study: reject (score 1), not "P4 or skip." Study covers Claude.ai consumer users wanting from AI generally. Off-topic for a Claude Code CLI guide.

Risks of non-integration: Orchestrator builders pay unnecessary token costs or break chain continuity from not knowing thinking.display: "omitted". CI pipeline authors get no guidance on when --bare is appropriate vs risky.

Decision

• Score final: 3/5
• Action: Partial integration (3 gaps: thinking.display P1 + model capabilities P2 + --bare P1)
• Confidence: Medium (secondary source; CC releases verified, API platform claims need verification)
• Next step: Fetch official platform release notes to confirm thinking.display and model capabilities fields before editing the guide