`anthropic-sdk-typescript` `v0.92.0` improves Managed Agents and env-based headers

// related

• #0001

  #0001Models & APISimon Willison7 hours ago

  `llm-gemini` `0.32` adds `gemini-3.5-flash`

  50radar

  llm-geminiLLM CLI plugin — calls Gemini models from `llm`

  Simon Willison's llm CLI can now call Google's new Flash model through the Gemini plugin. Small update, but useful if your scripts already depend on llm.

  • —llm-gemini now exposes gemini-3.5-flash, so existing llm CLI workflows can test the model without custom API glue.
  • —Scope is one model alias in a plugin release. This is practical plumbing, not a new app-building capability by itself.
  • —Best fit is quick model comparison for summaries, extraction, and generation jobs already wired around Simon Willison's llm ecosystem.

  Source: simonwillison.net/2026/May/19/llm-gemini-2/#atom-everythRead original →
• #0002

  #0002Models & APISimon Willison8 hours ago

  `Gemini 3.5 Flash` ships broadly with a 3-6x price jump

  90radar

  Google is putting the new default-grade model into Search, Gemini, Android Studio, and the API. The API math changed: high-output features need a fresh margin check.

  • —Model ID is gemini-3.5-flash; it supports 1,048,576 input tokens and 65,536 output tokens. Strong fit for long-context document flows.
  • —Pricing is $1.50/M input and $9/M output, 3x 3 Flash Preview and 6x 3.1 Flash-Lite. Flash is no longer the obvious cheap default.
  • —Interactions API is in beta with server-side history management, echoing the Responses API pattern. Agent backends may get simpler state handling.
  • —No computer-use feature in this release. If your workflow depends on browser/desktop control, this is a model upgrade, not a full agent-runtime replacement.
  • —3.5 Pro is slated for next month, likely pricier. Build model routing now instead of hard-coding one Gemini tier.

  Source: simonwillison.net/2026/May/19/gemini-35-flash/#atom-everRead original →
• #0003

  #0003Models & APIGeekNews10 hours ago

  `Gemini 3.5 Flash` targets long-running agents and coding

  100radar

  Google is pushing the fast tier into frontier-agent territory. Recheck Gemini automation stacks where latency matters and quality was the blocker.

  • —First Gemini 3.5 model combines frontier-level intelligence with execution ability, aimed at long-running agent and coding tasks.
  • —Keeps the Flash-series speed profile, so it competes for production flows where Pro-class latency was too expensive.
  • —Scores 76.2% on Terminal-Bench 2.1 and 1656 Elo on GDPval-AA, beating Gemini 3.1 Pro in the cited benchmarks.

  Source: news.hada.io/topic?id=29670Read original →