#33 - 🤖 Gemini 3 lands, devs push back

Plus: Anthropic's agent tooling upgrade

November 25, 2025

Hey readers! 👋 Google dropped Gemini 3 this week and the AI coding world collectively lost its mind, while Anthropic quietly shipped some genuinely useful tooling updates that might actually change how we build agents. Meanwhile, the data keeps rolling in on whether all these AI tools are actually helping us, and the answer remains a resounding "it depends."

🚀 This Week's Highlights

Gemini 3 Is Here—and Google Says It Will Make Search Smarter — Google launched its most advanced multimodal model with a 1M token context window, 64K token output, and claims of superior reasoning and coding abilities. – WIRED

The hype machine is running hot on this one. Early testers report impressive results in Cursor, with some claiming it "one-shots everything" without errors. GitHub has already rolled out Gemini 3 Pro in public preview for Copilot, reporting 35% higher accuracy compared to Gemini 2.5 Pro. Simon Willison put it through its paces with audio transcription and his infamous pelican-on-a-bicycle benchmark, finding it's essentially "Gemini 2.5 upgraded to match the leading rival models." The model stumbled on timestamp accuracy for long audio files, which matters if you're trying to verify transcripts against source material.

Claude advanced tool use

Introducing advanced tool use on the Claude Developer Platform — Anthropic shipped three capabilities that make Claude agents dramatically more efficient with large tool libraries. – Anthropic

This is the kind of infrastructure work that doesn't make headlines but genuinely moves the needle. The Tool Search Tool lets Claude discover tools on-demand instead of loading everything into context, preserving nearly 70,000 tokens compared to the traditional approach. Programmatic Tool Calling allows Claude to orchestrate multiple tool calls through code, keeping intermediate results out of context and enabling parallel execution. Tool Use Examples provide sample calls that teach correct usage patterns for complex APIs. If you're building agents that need to work with dozens or hundreds of tools, this is worth your attention.

Claude Opus 4.5 is now rolling out to GitHub Copilot — Early testing shows Opus 4.5 surpasses internal coding benchmarks while cutting token usage in half, with particular strength in code migration and refactoring. – GitHub

The promotional 1x premium request multiplier runs through December 5, so now's the time to test it. CodeRabbit's analysis positions Opus 4.5 between Sonnet 4.5's verbose style and GPT-5.1's surgical precision, delivering higher per-comment precision and more meaningful findings.

📊 The Adoption Reality Check

Report: AI-assisted engineering boosts speed, quality mixed — A DX developer analytics report finds widespread adoption (~90%) coexists with modest measured impact: only about 22% of merged code is classified as AI-authored. – AI Native Dev

"Adoption doesn't equal impact," the report states. "Simply handing out licenses is not enough."

Heavy AI users ship more pull requests (median 2.3/week vs 1.4 for non-users), but quality outcomes vary wildly across organizations. The takeaway: structured training, governance, and measurement matter more than just turning on the tools.

I stopped using copilot and didn't notice a decrease in productivity — A developer switched laptops, forgot to enable Copilot, and was surprised to find no measurable slowdown. – Reddit

"I actually thought it was speeding things up quite a bit. But to my surprise, I found I wasn't going any slower with it off."

Good IntelliSense covers most needs, and avoiding AI-generated bugs has its own value. They still use Gemini for rubber-ducking and tests, but autocomplete might be optional for experienced devs.

Devs gripe about having AI shoved down their throats — Many developers feel pressured by employers to use AI coding tools, which they say can introduce bugs and impede skill development for junior engineers. – The Register

"The best way to learn is still with hands-on coding and getting feedback from someone who knows more. AI is short circuiting that entire cycle."

Companies are tracking usage and tying it to performance reviews to justify enterprise licenses. The tension between corporate mandates and developer autonomy is real.

🔧 Tools & Releases

OpenHands LM 32B — A new open-source coding model achieving 37.2% on SWE-Bench Verified, comparable to much larger models, available on Hugging Face for local deployment. – OpenHands

Olmo 3 — Allen Institute for AI released fully open 7B and 32B models with all training data, code, and checkpoints publicly available. – Ai2

New Codex model — OpenAI dropped an updated Codex model. Internal stats claim 95% of OpenAI engineers use Codex weekly, shipping roughly 70% more pull requests since adoption. – OpenAI

💡 Practical Guidance

Use AI to Boost Developer Productivity — Docker published a pragmatic four-phase workflow for agentic AI tools: Prompting, Planning, Producing, and Refining. – Docker

The key insight: manage context strictly, decompose work into small prompts, and use steering documents like CLAUDE.md to teach AI your project conventions. AI isn't autonomous; output quality reflects input quality.

The quality crisis in AI code — With 82-92% developer adoption and ~3x productivity boosts, 67% report maintaining quality is harder than ever. – QodoAI

Teams need review systems that are IDE-integrated, provide cross-repo context, and enforce policy-aware checks. The call to move from "vibes" to evidence for every change resonates.

📝 Quick Hits

Made with ❤️ by Data Drift Press

Hit reply with questions, comments, or feedback - we read everything!