AI Dev Essentials #35: Anthropic Acquires Bun + TypeScript 7 is 10x Faster

John Lindquist
Instructor

John Lindquist

Hey Everyone 👋,

John Lindquist here with the 35th issue of AI Dev Essentials!

Anthropic acquired Bun as Claude Code hit $1 billion in run-rate revenue just six months after public launch. AWS re:Invent 2025 dominated the week with Amazon Nova 2 models, the Kiro autonomous coding agent, and the blazing-fast Trainium3 chip. Mistral AI dropped their entire Mistral 3 family under Apache 2.0, including a massive 675B parameter model. DeepSeek released V3.2 with groundbreaking sparse attention that rivals Gemini 3 Pro. TypeScript 7 is now ready for testing with its native Go-based compiler delivering 10x faster builds. And Anthropic is reportedly preparing for a 2026 IPO at a potential $300 billion valuation.

The AI video space also heated up with Runway Gen-4.5 "Whisper Thunder" claiming the top spot on benchmarks, while Kling 2.6 introduced native audio generation for synchronized speaking, singing, and sound effects.

I've been focusing on a tool for benchmarking CLAUDE.md files (sharing soon!), optimizing "code mode" for lazy loading tools, and preparing for my guest speaking spot at Microsoft's AI Dev Days next week. I'll be speaking about "Scripting AI Agents" and how you can combine deterministic scripts with non-deterministic agents to create powerful workflows. It will be published after and I'll make sure to share the link here.

⚡ Claude Code Power User Workshop — 📅 Dec 12 🚨 ENDS TONIGHT • 🔥 Early Bird: $300 (was $375) — Register Now

🚀 Major Announcements

Breaking Anthropic Acquires Bun as Claude Code Hits $1B Milestone

Anthropic announced on December 2, 2025 that it is acquiring Bun, the breakthrough JavaScript runtime, as Claude Code reached $1 billion in run-rate revenue just six months after becoming publicly available.

Claude Code Milestone:

  • $1B Run-Rate Revenue: Achieved in only 6 months since May 2025 GA launch
  • Enterprise Adoption: Netflix, Spotify, KPMG, L'Oreal, Salesforce among customers
  • Origin: Started as internal engineering experiment, now critical enterprise tool

Bun Acquisition Details:

  • Founded: 2021 by Jarred Sumner
  • Capabilities: All-in-one JavaScript toolkit (runtime, package manager, bundler, test runner)
  • Adoption: 7+ million monthly downloads, 82,000+ GitHub stars
  • Users: Midjourney, Lovable, and many others
  • License: Will remain open source and MIT-licensed

Integration Benefits:

  • Faster Performance: Bun's speed optimizations coming to Claude Code
  • Improved Stability: Enhanced reliability for AI-led development workflows
  • Native Installer: Bun already powered Claude Code's recent native installer launch
  • New Capabilities: Deeper integration between runtime and AI coding workflows

"Bun represents exactly the kind of technical excellence we want to bring into Anthropic. Jarred and his team rethought the entire JavaScript toolchain from first principles while remaining focused on real use cases." — Mike Krieger, Chief Product Officer, Anthropic

(Anthropic News)

💬 I use Bun for all of my personal projects and even scripts. I'm excited to see it get the support it needs to continue to grow. I just hope Anthropic is a good steward of the project since ironically Claude Code is one of the only AI CLIs that is closed-source. The most interesting theory I've heard is that now Anthropic can add a standard lib to Bun, train Claude Code on it, then use it as a much more powerful and reliable foundational tool that works in every ecosystem.

Breaking AWS re:Invent 2025: Nova 2 Models, Kiro Agent, and Trainium3

AWS unveiled a comprehensive AI strategy at re:Invent 2025 (December 2-5), introducing the Nova 2 model family, three frontier AI agents, and the next-generation Trainium3 chip.

Nova 2 Model Family:

  • Nova 2 Pro: Flagship reasoning model with multimodal capabilities (text, images, video, speech) for complex problem-solving and agentic workflows
  • Nova 2 Lite: Cost-effective, low-latency model for high-volume tasks
  • Nova 2 Omni: Native multimodal model processing text, audio, and video simultaneously with image generation
  • Nova 2 Sonic: Speech-to-speech model for real-time voice interactions with emotion understanding
  • Context Window: 1 million tokens across all models
  • Performance: ~66% on SWE-bench Verified via Amazon Q Developer

Frontier AI Agents:

  • Kiro: Autonomous coding agent that works independently for hours to days, learns team patterns, integrates with Jira/GitHub/Slack, and manages multi-repository changes
  • AWS Security Agent: Virtual penetration tester with autonomous security reviews and CI/CD validation
  • AWS DevOps Agent: Always-on operations expert for incident triage, root cause analysis, and IaC generation

Trainium3 Chip Specifications:

  • Performance: 2.52 PetaFLOPS (FP8) per chip, up to 10 PetaFLOPS with sparsity
  • Memory: 144 GB HBM3e per chip with 4.9 TB/s bandwidth
  • Efficiency: 40% more energy efficient, 3x throughput vs Trainium2
  • UltraServer Scale: 144 chips per server (~362 PetaFLOPS total)
  • Cost: 50% lower training costs vs comparable GPU instances

(AWS News Blog, TechCrunch, VentureBeat)

💬 I was wondering where AWS was in the AI space. Looks like they're diving in full throttle. My "way too early" take is that the early benchmarks for Nova 2 Pro lag a little behind the other frontier models, so no one is really going to rush out to try it. I'm certainly way too busy with so many other tools to dedicate any time to anything that's not the "best". Opus 4.5 has stolen my heart.

Breaking Mistral 3 Family: 675B Parameters Under Apache 2.0

Mistral AI released their third-generation model family on December 2, 2025, open-sourcing their entire frontier lineup including a 675B parameter flagship model under the Apache 2.0 license.

Mistral Large 3 Specifications:

  • Total Parameters: 675 billion (41B active via Mixture-of-Experts)
  • Context Window: 256,000 tokens
  • Training: 3,000 NVIDIA H200 GPUs
  • License: Apache 2.0 (unrestricted commercial use)
  • Performance: #2 on LMArena for open-source non-reasoning models

Ministral 3 Series (Edge-Focused):

  • Sizes: 3B, 8B, and 14B parameters
  • Variants: Base, Instruct, and Reasoning versions
  • Target: Edge devices, local deployment, consumer hardware

Key Capabilities:

  • Native Vision: All models include image understanding
  • Tool Use: Optimized for agentic workflows and function calling
  • Multilingual: Strong European language support
  • Cost Efficiency: Claims ~80% lower cost than top-tier proprietary models

Availability: Mistral AI Studio, Amazon Bedrock, Azure AI Foundry, Google Cloud Vertex AI, Hugging Face, OpenRouter, vLLM

(Mistral AI Blog, TechCrunch, VentureBeat)

💬 Another solid win for open source. It's great to see them staying close to closed models and it gives me some hope that open source will win out in the end. That cost-efficiency is a huge deal for many scenarios though.

Breaking DeepSeek V3.2 Launches with Revolutionary Sparse Attention

DeepSeek released V3.2 and V3.2-Speciale on December 1, 2025, introducing DeepSeek Sparse Attention (DSA) that reduces attention complexity from quadratic to near-linear while rivaling Gemini 3 Pro performance.

Technical Breakthrough:

  • DeepSeek Sparse Attention: 90% sparsity with near-linear O(kL) complexity
  • Performance Gains: 3.5x faster inference for 128k context, 50-70% cost reduction
  • Architecture: Lightning Indexer for token ranking, Fine-Grained Selector for relevance

Model Variants:

  • V3.2: Daily driver with "thinking in tool-use" capability, OpenAI-compatible API
  • V3.2-Speciale: Maxed reasoning (2-5x more output tokens), rivals Gemini 3 Pro

Performance Benchmarks:

  • MMLU: ~91% (Speciale)
  • Codeforces: Gold Level (outperforms GPT-5 High in competitive coding)
  • IMO/IOI: Gold-level performance
  • ICPC World Finals: 2nd place

Pricing (Extremely Aggressive):

  • Input: $0.27-0.28/M tokens (Cache hit: $0.028)
  • Output: $0.41-0.42/M tokens
  • 10-20x cheaper than GPT-5 equivalents

Availability: DeepSeek platform, mobile app, API, OpenRouter. V3.2-Speciale free until December 15, 2025.

(DeepSeek Blog, OpenRouter)

💬 DeepSeek has full integration with Claude Code, so it's a model I've thrown a few tests at last week. It does feel a little slower than Opus 4.5, but good enough (and cheap enough) that I'll keep it in the mix and throw some more long-running tests with it. I did see some bugs around compaction which might be a blocker for me, but I'm not sure if they were random or if I'll keep running into them.

Tool TypeScript 7 Native Port Ready for Testing with 10x Speed Gains

Microsoft announced on December 2, 2025 that TypeScript 7's native Go-based compiler (Project Corsa) is now ready for public testing, delivering 10x faster build times with full --build and --incremental support.

Performance Improvements:

  • 10x Faster Builds: Full builds complete in a fraction of the time
  • Reduced Memory: Addresses OOM crashes in large codebases
  • Improved Parallelism: Efficient multi-project builds for monorepos
  • Faster Editor Startup: Drastically improved VS Code project loading

Features Implemented:

  • --build Support: Full project references working
  • --incremental Support: Incremental compilation functional
  • Feature-Complete Type-Checking: Works for most projects
  • Native Language Service: Powers IntelliSense, Go-to-Definition, Refactoring
  • VS Code Extension: Daily-updated extension available now

Current Gaps:

  • JavaScript emit for older runtimes still in development
  • Full --watch mode feature parity not yet complete

Release Strategy:

  • TypeScript 6.0: Final JavaScript-based release (transition point)
  • TypeScript 7.0: First native release with strict: true as default

(Microsoft DevBlogs, GitHub: microsoft/typescript-go)

💬 TypeScript is quickly becoming the language for writing AI Agents. So many SDKs are based on it (the AI SDK, Claude Agent SDK, etc). I'll take any speed improvement I can get to help me build faster.

Breaking Anthropic Preparing for 2026 IPO at $300B+ Valuation

Anthropic is laying groundwork for a potential 2026 IPO that could value the company at over $300 billion, according to Financial Times reporting from December 2025.

IPO Details:

  • Target Timeline: As early as 2026
  • Target Valuation: Over $300 billion (some reports suggest $350B)
  • Legal Advisor: Wilson Sonsini Goodrich & Rosati engaged
  • Current Valuation: ~$183 billion (September 2025 Series F)

Financial Projections:

  • 2026 Revenue: ~$26 billion ARR projected
  • Customer Base: Over 300,000 business customers
  • Major Backers: Amazon, Google, Microsoft, NVIDIA

Official Statement: "We have not made any decisions on when, or even if, we will go public."

Context: Move seen as strategic response to OpenAI's rumored late-2026 IPO with potential $1 trillion valuation.

(Financial Times, Reuters, Bloomberg)

💬 $300 billion would make this one of the largest tech IPOs ever. I know many people want to make this about OpenAI and "competitors", but it seems like Anthropic mostly just stays in their own lane and focuses on quality agents. It will be curious as models evolve and how they drive their revenue. Especially if an open-source model starts to dominate the space.

🛠️ Developer Tooling Updates

Tool Runway Gen-4.5 "Whisper Thunder" Claims Video Benchmark Crown

Runway released Gen-4.5 on December 1, 2025, achieving the #1 ranking on the Artificial Analysis Video Arena leaderboard with breakthrough physics simulation and director tools.

Technical Capabilities:

  • Physics-Based World Model: Realistic liquid flow, object weight, momentum, and interactions
  • Complex Choreography: Multi-stage prompts in single generation
  • Director Mode 2.0: Advanced camera control tools
  • Motion Brush 3.0: Independent movement control for specific objects
  • Resolution: Native 1080p
  • Duration: 5-10 second clips

Benchmark Performance:

  • Artificial Analysis Elo: 1247 (#1, ahead of Veo 3.1 at 1226 and Sora 2 Pro at 1206)

Pricing:

  • Gen-4.5 Base: 12 credits/second ($0.12/sec)
  • Gen-4.5 Turbo: 5 credits/second ($0.05/sec)
  • Subscriptions: Standard $12/mo, Pro $28/mo, Unlimited $76/mo

Limitation: No native audio (unlike Veo 3 and Sora 2)

(Runway ML, TechCrunch)

Tool Kling 2.6 Introduces Native Audio-Visual Generation

Kling AI released version 2.6 on December 3, 2025, introducing native audio generation that synchronizes speaking, singing, and sound effects with video in a single pass.

Audio Capabilities:

  • Speaking & Dialogue: Multilingual (English, Mandarin, Spanish), frame-accurate lip-sync, emotional intonation
  • Singing: Lyric integration, genre versatility (Pop, Opera, Rap, Country), rhythm control
  • Sound Effects: Frame-level foley synchronization, prompt-driven SFX, environmental audio

Technical Specs:

  • Resolution: Native 1080p
  • Duration: 5-10 second clips
  • Audio Output: High-fidelity stereo
  • Generation: Single-pass joint attention for audio-visual synchronization

Workflow Options:

  • Text-to-Audio-Visual: Describe scene and sound together
  • Image-to-Audio-Visual: Animate reference image with appropriate audio

(Kling AI, Kuaishou Technology)

Tool Google Workspace Studio: No-Code AI Agent Builder

Google launched Workspace Studio on December 3-4, 2025, a no-code platform for building autonomous AI agents that work across Gmail, Drive, Docs, Sheets, Calendar, and third-party apps.

Core Capabilities:

  • Natural Language Agent Creation: Build agents by describing tasks in plain English
  • Cross-App Orchestration: Agents work across entire Google Workspace
  • Third-Party Integration: Salesforce, Jira, Asana, Mailchimp
  • Shareable Agents: Share like Google Docs for team standardization
  • Powered By: Gemini 3 for multi-step planning and autonomous execution

Example Workflow: Monitor Gmail for invoices → Extract data to Sheets → Request approval via Chat → Email confirmation (all automatic)

Rollout: Generally Available, rolling out to business customers. Alpha participants completed 20+ million tasks in last 30 days.

(Google Workspace Blog, Google Blog)

💬 Google is massively integrating AI into every step of their product suite. I personally switched to a Pixel phone just so I could have a physical button to trigger AI. Allowing users to create custom workflows is going to be a fascinating UX pattern to watch. I know many of us set these up in Zapier and other tools, but this is a much more integrated approach.

Tool Claude Opus 4.5 Now Available in Claude Code for Pro Users

Anthropic announced on December 3, 2025 that Claude Opus 4.5 is now available in Claude Code for Pro subscribers via the /model command.

Access Details:

  • Availability: Claude Code Pro users
  • Selection: Use /model command in terminal
  • Capabilities: Full Opus 4.5 intelligence including effort parameter, 200K context

(Claude Code Changelog)

💬 My favorite model coming to the Pro tier. I'm excited that more people get to experience it.

Tool Gemini CLI v0.19 Ships with ElevenLabs Extension Support

Google released Gemini CLI v0.19.0 in late November 2025, featuring enhanced MCP support that enables the official ElevenLabs audio extension.

New Features:

  • Interactive Session Browser: UI for browsing past sessions
  • Model Availability Service: Track model health across providers
  • Lenient MCP Validator: Enables complex extensions like ElevenLabs

ElevenLabs Extension Capabilities (via MCP):

  • Text-to-Speech generation
  • Voice Design and Cloning
  • Sound Effects and Music Generation
  • Install: gemini extensions install https://github.com/elevenlabs/elevenlabs-mcp

Bug Fixes: CJK character cursor positioning, authentication flow, bash prompt detection

(GitHub: google-gemini/gemini-cli)

💬 The ElevenLabs integration is a nice addition for anyone building audio workflows. Seeing the Gemini CLI embrace extensions is a big win for the ecosystem that might help differentiate it from the competition.

🤖 AI Ecosystem Updates

Tool ByteDance Releases Seedream 4.5 Image Generation Model

ByteDance released Seedream 4.5 on December 3, 2025, featuring multi-reference consistency for maintaining character and scene coherence across multiple images.

Key Capabilities:

  • Multi-Reference System: Up to 10-14 reference images for consistency
  • Typography: Native support for complex text layouts and logos
  • Resolution: Native 4K (4096×4096) output
  • Speed: ~1.8 seconds for 2K images
  • Unified Architecture: Single model for generation and editing

Performance: Outperforms Gemini 2.5 Flash Image in prompt adherence; leads in multi-image consistency vs Midjourney v6.1 and Flux.1

Availability: BytePlus ModelArk (enterprise), Jimeng AI/Dreamina (consumer), third-party APIs (WaveSpeedAI, Fal.ai, Kie.ai)

Pricing: ~$0.03-0.04 per image

(BytePlus, Jimeng AI)

Research Apple Releases CLaRa-7B for Efficient RAG

Apple released CLaRa-7B-Instruct on December 3, 2025, a 7B parameter model that retrieves compressed vector representations instead of raw text for 16-128x more efficient RAG.

Core Innovation:

  • Latent Retrieval: Retrieves compressed vectors instead of text chunks
  • Compression: 16x to 128x compression while retaining semantic meaning
  • Efficiency: Process significantly more documents within same context window

Model Variants:

  • apple/CLaRa-7B-Instruct (instruction-tuned for QA)
  • apple/CLaRa-7B-Base (foundation model)
  • apple/CLaRa-7B-E2E (end-to-end retriever and generator)

Availability: Open source on Hugging Face. Deployable on Apple Silicon.

(arXiv:2511.18659, Hugging Face: apple)

Ecosystem OpenAI Partners with Accenture for Massive Enterprise Rollout

OpenAI and Accenture announced a strategic partnership on December 1, 2025, deploying ChatGPT Enterprise to tens of thousands of Accenture professionals and creating the largest OpenAI certification program.

Partnership Scope:

  • ChatGPT Enterprise: Deployed to tens of thousands of Accenture staff
  • OpenAI Certifications: Largest group of certified professionals
  • AgentKit Focus: Building custom AI agents for enterprise workflows
  • Target Industries: Financial Services, Healthcare, Public Sector, Retail

Long-term Goal: OpenAI committed to certifying 10 million Americans by 2030 through OpenAI Academy

(OpenAI Blog, Accenture Newsroom)

💬 This signals OpenAI's serious enterprise push. Accenture as "Client Zero" gives them a massive proof point for security and scalability concerns. The certification program could create a whole ecosystem of OpenAI-trained implementation partners.

Education Anthropic Partners with Dartmouth for Claude for Education

Anthropic announced a partnership with Dartmouth College and AWS on December 3, 2025, bringing Claude for Education to the entire Dartmouth community.

(Anthropic Newsroom)

⚡ Quick Updates

OpenAI "Code Red" Response to Gemini (Rumor)

  • Status: Unconfirmed reporting from The Information
  • Claims: Internal "Code Red" declared December 1, feature freeze on non-core products
  • Context: ChatGPT reportedly lost ~12M daily active users post-Gemini 3 launch
  • Response: New model codenamed "Garlic" targeting early 2026

(The Information)

Cursor Composer Deep-Dive Talk Released

  • Content: Technical journey behind Cursor's first in-house coding LLM
  • Speakers: @srush_nlp with @leerob
  • Platform: AI Engineer YouTube channel

(YouTube: AI Engineer)

Vercel AI Gateway Adds Image-Only Models

  • Models: 7 BFL/Flux models, 3 Google Imagen models
  • Integration: Supported by AI SDK's generateImage
  • Pricing: Per megapixel or per image

(Vercel AI Gateway)

Vercel Workflow Builder Hits 8,600+ Workflows

  • First Week Stats: 580 stars, 7,600+ users, 330 commits, 11 contributors
  • Status: Actively accepting contributions

(GitHub: vercel/workflow)

Gemini 3 + Structured Outputs Now Combined

  • Feature: Google Search tool works with Structured Outputs
  • Use Case: Grounded search results in typed schemas

(Google AI Developers)

✨ Workshop Spotlight

Claude Code Power User Workshop

Learn the essential skills to ship reliable AI-generated code with confidence. This hands-on workshop covers everything from foundational prompting to advanced automation using the Claude Code SDK and custom integrations.

Friday, Dec 12 @ 9:00 AM PST

🚨 ENDS TONIGHT! • 🔥 Early Bird Pricing — Standard $375 (save $75) | Pro Member $300 (save $150)

What you'll learn:

  • Context Engineering: Control what context Claude sees for reliable, consistent results
  • TypeScript SDK: Script Claude programmatically to build custom workflows
  • Custom Hooks: Automate repetitive tasks with Claude Code hooks
  • Model Context Protocol: Integrate APIs securely to extend Claude's capabilities
  • Claude Code Skills: Build custom skills for Claude Code
  • Live Q&A: Get your specific questions answered by John Lindquist

Register Now ($300) — $225 for egghead Pro Members

Did you enjoy this issue? Share it with a friend.

© 2025 John Lindquist • egghead.io

Share with a coworker