🤖 AI Weekly Recap (Week 23)

Audio-to-Audio, Not Text-to-Speech: Dubbing v2 conditions directly on the original vocal performance, capturing tone, pacing, emphasis, hesitation, and emotional intent. Earlier systems converted speech to text before translation, losing every nuance that made the original sound human.

Voice Clone of the Original Speaker: Each translation is delivered in an automatic voice clone of the original speaker, preserving identity and pitch with no manual setup. Background music and sound effects stay intact, so there's no re-mixing required.

One-Click Localization: Creators can localize YouTube videos and other content inside ElevenCreative with a single click. ElevenLabs is also launching a Creator Dubbing Partner Program giving eligible creators discounted access.

Sees the Screen Like You Do: Instead of brittle selector-based automation, agents use vision and reasoning to navigate live UIs. They adapt when layouts shift, fields move, or workflows branch, so they don't break every time a vendor portal updates its design.

Enterprise-Grade Governance: Sessions log straight into Microsoft Purview and Dataverse with full replay, credentials live in Azure Key Vault, and human-in-the-loop checkpoints can be configured for any step that needs approval or falls below a confidence threshold.

Model Choice Built In: Agents can run on models from OpenAI and Anthropic, with Windows 365 Cloud PC pool support for ephemeral execution. Microsoft is the first major platform to ship computer-use at production scale while Anthropic and Google are still in beta.

Claude Writes Its Own Orchestration Script: Describe a task in plain English and Claude writes a JavaScript orchestration script on the fly. A separate runtime executes it in the background while your chat session stays responsive, with a cap of 1,000 subagents per run.

Built for Codebase-Scale Work: Internal teams used Dynamic Workflows to port the Bun JavaScript runtime from Zig to Rust, producing 750,000 lines of Rust code with 99.8% test pass rate in just 11 days from kickoff to merge.

Self-Verifying by Design: Adversarial subagents critique each other before any result reaches you, with the existing test suite as the completion bar. Available in research preview on Claude Code CLI, Desktop, and VS Code for Max, Team, and Enterprise plans.

Plan-First Execution: Before touching any file, Grok produces a written plan in plain English listing the files to modify, commands to run, and intermediate checks. Approve, comment, or rewrite the plan, like a mental pull request, before the CLI moves to execution.

Up to 8 Parallel Subagents: Each subagent runs in its own isolated Git worktree, so concurrent edits never stomp on each other. One can rewrite auth, another updates tests, a third refactors a schema migration, all in parallel.

Drop-In Compatible: Grok Build reads your AGENTS.md, respects your existing hooks, plugins, MCP servers, and skills, and runs on Grok Build 0.1 (a purpose-built coding model) scoring 70.8% on SWE-Bench Verified.

Honest About Its Own Work: Opus 4.8 is roughly 4x less likely than 4.7 to let flaws in code it has written pass unremarked, and rates of deception and cooperation with misuse drop substantially, approaching the alignment levels of the unreleased Claude Mythos Preview.

Cheaper Fast Mode: Fast mode now runs at $10 input / $50 output per million tokens (down from $30/$150 on Opus 4.7) while keeping you on the full Opus model, not a smaller one. Activate it in Claude Code with /fast.

Frontier Coding and Computer Use: Scores 69.2% on SWE-Bench Pro (up from 64.3% on 4.7) and 84% on Online-Mind2Web for computer use, beating GPT-5.5 and Gemini 3.1 Pro on the majority of benchmarks. Standard pricing stays flat at $5 input / $25 output.

Persistent Cloud Execution: Unlike chatbots that end when you close the tab, Spark lives on its own cloud VM and runs continuously. It can monitor Gmail for school deadlines, parse statements for hidden subscriptions, or extract leads from inbound inquiries into a Sheet, all without you being present.

Native Workspace Access: Connects directly to Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Google Maps. Third-party services like Canva, OpenTable, and Instacart plug in via MCP, the open integration standard from Anthropic.

Tasks, Skills, and Schedules: Teach Spark how you want recurring jobs done, save them as reusable Skills, and schedule them to run on their own. Powered by Gemini 3.5 Flash on the Antigravity agent harness, with beta access starting now for Google AI Ultra subscribers in the US.