โ† All digests
AI Developer Digest

Mon, Nov 16, 2026

5 signals that cleared the gate19 min read
The Signal โ€” start here
Two closely related SDK releases today mark a housekeeping inflection for Anthropic's API layer: Python SDK v2.9.0 and TypeScript SDK v1.9.0 formally deprecated seven beta headers that have been accumulating since 2024. The API still accepts them โ€” but the SDKs are signalling that a future release will remove them, and the ones most likely to bite are prompt-caching-2024-07-31 and token-efficient-tools-2025-02-19, which are frequently hardcoded in older agent setups. Separately, Anthropic extended text watermarking to four more Claude models today (Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5). The pattern is consistent: systematic pre-IPO platform tightening โ€” retire old API surface, harden the security model, expand compliance primitives. NVIDIA/SkillEvaluator is the standout from today's GitHub Trending scan: a new tool for validating, deduplicating, and live-testing agent skills before you deploy them โ€” timely as skills become a first-class platform primitive and the trending list this week shows five separate skills libraries all gaining traction simultaneously.
Must-reads today
1
anthropic-sdk-python v2.9.0 / sdk-typescript v1.9.0: 7 beta headers deprecated with explicit warning they'll be removed in a future SDK release โ€” scan your code for prompt-caching-2024-07-31, token-efficient-tools-2025-02-19, and 5 others today
2
NVIDIA/SkillEvaluator: a new CLI + framework for validating agent skills before deployment โ€” pip install skillevaluator and skillevaluator validate ./my-skill --tiers 1 runs the full static check with no API key required

๐Ÿ† Repo of the Day

1

NVIDIA/SkillEvaluator โ€” validate your agent skills before they bite you in production

Install

pip install skillevaluator

or

uv tool install skillevaluator

Tier 1 static check โ€” no API key needed

skillevaluator validate ./my-skill --tiers 1

Full quality check including rubric eval (needs NVIDIA Build API key)

skillevaluator quality-check ./my-skill

Set up CI integration

skillevaluator ci init


Breaking Changes

No hard breaking changes this period.

(The beta header deprecations in SDK v2.9.0 / v1.9.0 are soft deprecations โ€” the API still accepts the old headers. They are staged for removal in a future SDK release, not today. See API & SDK Changes for the action required.)


Model Releases

No model releases verified from primary sources within the 24-hour window.

(blog.google DNS-blocked for the 17th consecutive day โ€” Gemini 4 Hydrogen pricing and model ID still unverifiable from primary source. See near-misses.)


API & SDK Changes

2
Medium

Anthropic text watermark extended to Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5

What changed
Four additional Claude models โ€” Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 โ€” now carry Anthropic's text watermark on supported platforms. Previously (November 3), watermarking was live only on Mythos 5, Opus 4.7, and Sonnet 4.6. This more than doubles the set of watermarked models.
TL;DR
Text output from Claude Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 on supported platforms now embeds a machine-readable watermark; non-default temperature, top_p, or top_k values may suppress it.
Developer signal
If your application relies on detecting or attributing Claude-generated text โ€” content moderation pipelines, compliance logging, provenance tracking โ€” you should update your tooling to expect watermarks from these four model IDs. The key caveat: watermarking may not be applied when temperature, top_p, or top_k are set to non-default values. If your sampling parameters vary per-request, don't rely on watermark presence as a guarantee of model identity. The "supported platforms" language is unchanged from earlier entries โ€” Claude API, Claude Platform on AWS, Anthropic-managed Google Cloud, and Microsoft Foundry, based on prior release note framing; check the docs for the definitive platform list.


Affects you ifYou use Claude Opus 4.6, Opus 4.5, Sonnet 4.5, or Haiku 4.5 on supported platforms; you have downstream pipelines that detect, log, or rely on model provenance; you use non-default sampling parameters and expected watermarks.EffortQuick โ€” no code changes unless you're actively reading or depending on watermark absence.
Anthropic Platform Release Notes | Date: November 16, 2026 | Link: https://platform.claude.com/docs/en/release-notes/overviewhttps://platform.claude.com/docs/en/release-notes/overview (November 16, 2026 entry)
Medium

anthropic-sdk-python v2.9.0 + sdk-typescript v1.9.0 โ€” seven beta headers deprecated, grep/file tool security hardening

What changed
Both SDKs deprecate seven legacy anthropic-beta header values and remove the automatic message-batches-2024-09-24 header from batch requests. Security hardening was applied to the grep, file, and memory tools โ€” rg, tar, and unzip no longer receive ANTHROPIC_* environment variables, and symlink handling tightened. New: Update Bundle endpoint and RBAC Role endpoints return error codes.
TL;DR
Seven beta headers deprecated across Python v2.9.0 and TypeScript v1.9.0 โ€” the API still accepts them, but a future SDK release will hard-remove them; message-batches-2024-09-24 is already gone from batch requests.
Developer signal
Run a codebase search now for any of these string values โ€” they're the headers to remove:

  • prompt-caching-2024-07-31

  • pdfs-2024-09-25

  • token-counting-2024-11-01

  • token-efficient-tools-2025-02-19

  • extended-cache-ttl-2025-04-11

  • cache-diagnosis-2026-04-07

  • ce-user-management-2026-07-13

Do not remove files-api-2025-04-14 or skills-2025-10-02 until you've followed the Files API and Skills migration guides โ€” these two still select beta response shapes and removing them changes what the API returns. If you set message-batches-2024-09-24 manually in your batch requests, delete it. C# users: AnthropicBeta enum member numbers changed โ€” rebuild any code compiled against v2.8.x to avoid sending wrong header values silently. The grep tool security changes (no longer inherits ANTHROPIC_* env vars, only reads rg from absolute PATH entries) are fixes that do not require code changes but may affect custom tool setups that relied on environment forwarding.


Affects you ifYou send any of the seven deprecated beta headers in SDK or raw API requests; you use the Python SDK on Windows with Python โ‰ค 3.12 and hit credential file bugs; you use C# and compile against the AnthropicBeta enum; you rely on the Message Batches API without pinning your own beta header.EffortModerate โ€” search and remove deprecated headers; validate Files API and Skills API migration before touching files-api and skills headers; rebuild C# bindings.

Research

arxiv.org DNS-blocked (17th consecutive day). huggingface.co/papers DNS-blocked. No research items confirmed from primary sources this period.


Tooling

No new framework releases in the 24-hour window that weren't already covered.

(pydantic-ai v2.68.0 โ€” November 11, outside window. gemini-cli v0.69.0-preview โ€” November 10, outside window. OpenHands v1.40.0 โ€” November 12, outside window.)


Rising Dev Tools

1
Notable

NVIDIA/SkillEvaluator โ€” validate, deduplicate, and live-test agent skills before deployment

Developer signal
Three tiers of checks you can run selectively. Start with Tier 1 (no API key): pip install skillevaluator && skillevaluator validate ./my-skill --tiers 1 โ€” runs safety, format, and harness compatibility checks statically. Add --tiers 2 for semantic deduplication against your existing skills catalog. Tier 3 (--tiers 3) runs live agent comparisons and needs a provider key plus Docker. The ci init command drops a CI template into your repo for automated checks on skill PRs.


Affects you ifYou write, ship, or consume agent skills for Claude Code, Gemini CLI, or other harness-compatible agents; you manage a skills catalog across a team and need to prevent duplicate guidance; you've had skill-related security concerns (the grep/file tool security fixes in today's SDK release show why this matters).
GitHub Trending (all languages, daily, Nov 16) | Stars: ~2.5k (โ†‘ 953 today) | Link: github.com/NVIDIA/SkillEvaluator

Benchmarks & Leaderboards

lmarena.ai DNS-blocked (fifteenth consecutive day). swebench.com DNS-blocked. No benchmark data confirmed from primary sources this period.

(Third-party sources show Gemini 4 Carbon leading the LMArena Text board with ~1528 score per an October 28 evaluation snapshot โ€” unverifiable from primary source, excluded per quality gate.)


Technical Discussions

hnrss.org DNS-blocked. simonwillison.net DNS-blocked. Nothing cleared the quality bar this period from confirmed primary sources.


Quick Hits

  • โš ๏ธ Gemini API preview model retirements โ€” TOMORROW (Nov 17): gemini-3.1-flash-tts-preview, Live API preview models, and gemini-2.5-tts-preview all retire tomorrow. Migrate to gemini-3.8-live and gemini-3.8-flash-tts now. Source: ai.google.dev/gemini-api/docs/deprecations (DNS-blocked โ€” verify at official URL)
  • LangGraph 1.2.22 (Nov 14) โ€” subgraph child thread streaming, performance fixes. Already in Nov 14/15 Quick Hits โ€” carry-forward for visibility: github.com/langchain-ai/langgraph/releases/tag/0.2.75

Worth Watching (Announced, Not Yet Shipped)

  • โš ๏ธ Gemini API preview model retirements โ€” TOMORROW, November 17 (1 day): gemini-3.1-flash-tts-preview, Live API preview models, and gemini-2.5-tts-preview all retire tomorrow. Migrate to gemini-3.8-live and gemini-3.8-flash-tts. If you are on any of these model IDs, this is a live production risk. Source: ai.google.dev/gemini-api/docs/deprecations (DNS-blocked โ€” verify at official URL)
  • Claude Opus 4.8 fast mode removal โ€” November 30 (14 days): speed: "fast" with claude-opus-4-8 will error. Migrate to Claude Opus 5.5. Source: platform.claude.com/docs/en/release-notes/overview
  • Claude Sonnet 4.5 retirement โ€” November 30 (14 days): claude-sonnet-4-5-20250929 retires. Migrate to Claude Sonnet 5.5. Source: platform.claude.com/docs/en/release-notes/overview
  • Gemini 4 Argon pricing doubling โ€” December 1 (15 days): Introductory rates ($2 input / $10 output per million tokens) confirmed by multiple secondary sources to end November 30. Standard rates ($4/$20) take effect December 1. Audit your Gemini 4 Argon usage now.
  • Upcoming SDK beta header hard removal: Python v2.9.0 and TypeScript v1.9.0 deprecated 7 beta headers today with the note "a later release removes them from the SDKs." No date given, but this is a time-boxed action โ€” act before the next minor SDK version drops.


Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff โ€” only what changes what you build.