Grok Voice 3.0 takes the speech crown as Claude Code learns to multitask

Grok Voice 3.0 tops the speech-to-speech leaderboard at the same 8 cents a minute

If you are building a phone agent or any voice interface, this is the release to test this week. Grok Voice 3.0 is live now in xAI's Voice Agent API, and anyone already pointing at the grok-voice-latest alias has been moved across automatically, so check your agents still behave as expected. The price is unchanged at $0.08 per minute ($4.80 an hour), and xAI says replies arrive in under a second, with 40% lower latency than OpenAI's GPT-Live-1 on the τ-Voice test.

xAI says the model ranks first on Artificial Analysis's Speech-to-Speech Index with 86.9, ahead of GPT-Live-1 at 83.2 and Google's Gemini 3.8 Live at 82.6, and scores 97% on Full Duplex Bench for handling interruptions. It is not a clean sweep: OpenAI still leads on τ-Voice's noisy, accented support calls, and the claim that it resolves 70% of Starlink support calls on its own comes from xAI rather than independent testing. Last month a rival voice model briefly took the same crown, so this lead may not last long. Language coverage has not been published, which matters if you serve UK customers with regional accents.

Claude Code desktop is rebuilt around running many sessions at once

If you already juggle two or three Claude Code sessions, this update is aimed squarely at you. A new session screen keeps recent folders one click away, and you can save a setup as an environment (a few folders, an SSH host or the cloud) and launch sessions into it, optionally with a separate git worktree per folder so parallel agents do not trip over each other. It builds on Claude Code Projects, which opened to Pro and Max users on Monday.

Anthropic says it is now quicker to start sessions, easier to split a task across several and simpler to see what each one is doing, and early write-ups describe child sessions that report back to a parent plus a sidebar showing which sessions are waiting on you. It arrives as a normal desktop app update for paid plans. One practical warning: every parallel session draws on your usage allowance, so you will hit plan limits sooner if you run lots at once. If you are still choosing a coding assistant, our guide to the best AI coding tools compares the main options.

Claude Platform can now send OpenTelemetry traces to your own monitoring tools

For anyone running Claude in production, this closes a real gap. In a public beta, Claude Platform can emit OpenTelemetry traces from the Messages API, Claude Managed Agents and the Claude SDKs straight into whatever OpenTelemetry-compatible observability stack you already use. That means you can follow an agent's calls, latency and failures alongside the rest of your system instead of digging through separate logs. Because OpenTelemetry is an open standard, you are not locked into one vendor to get this visibility. As a beta, expect the trace format and coverage to change, so avoid hard-wiring alerts to it just yet.

Anthropic opens in Singapore, and Claude gets in-region inference there through Amazon Bedrock

The office opening is the headline, but the product detail is what matters if you work in Southeast Asia. Amazon Bedrock now offers in-region inference for Claude in Singapore, which helps banks, insurers and public bodies that need data processed locally. Anthropic also signed a memorandum of understanding with Singapore's Ministry of Digital Development and Information covering AI safety, model evaluation and support for small businesses. Customers include GovTech Singapore, where more than 1,500 engineers now use Claude Code, plus DBS, GIC, Grab and Shopee.

Ultrafast GPT-6 Astra and SOL 6.1 arrive in Perplexity Computer for Max subscribers

If you pay for Perplexity Max and use Computer for long, multi-step jobs, you can now ask it to run noticeably faster. Ultrafast versions of OpenAI's GPT-6 Astra and SOL 6.1 are available by choosing either model under Custom and setting the speed to Ultrafast. Perplexity says both generate tokens faster, so long Computer tasks finish sooner, but it has not published speed figures or said whether Ultrafast eats into your usage limits. Try it on a task you run often and compare. It is another sign that agent products now compete on speed as much as on model quality.

YouTube Music starts labelling tracks as Fully Gen AI, Partly Gen AI or No Gen AI

If you make music with AI tools, or simply want to know what you are listening to, this changes what you see on the player page. YouTube is rolling out three labels on YouTube Music and the main YouTube app, each expanding into a short note on what AI actually did in the track. Crucially, YouTube says AI involvement alone does not make a track ineligible for monetisation, so disclosed AI music can still earn money. That is a different line from rivals such as Deezer, which is moving towards letting listeners filter AI music out. YouTube has not said which countries get the labels first.

Runway video models can now run in private, dedicated environments on Baseten

This is aimed at businesses that like Runway's output but cannot send footage or prompts to a shared cloud. Through a new Baseten partnership, companies can deploy Gen-4.5 and Aleph 2.0 in single-tenant environments, fine-tune them on their own data and reserve GPU capacity so generation does not slow at busy times. For agencies and studios with confidentiality rules, that removes one of the main blockers to using AI video at scale. It is an enterprise deal rather than a change to the Runway app most creators use, so expect negotiated rather than published pricing.

Nvidia launches cuPotential and signs up the big chip toolmakers for AI-run factories

This one is mostly for engineers in semiconductors and materials science, but it shows where Nvidia sees its next wave of software demand. cuPotential is a new library, in early access, for building AI agents that simulate how atoms and molecules interact, and Nvidia says it adds a further two to three times speed-up on workflows already running on GPUs. Lam Research, KLA, Applied Materials, Synopsys and Cadence all reported large speed-ups from Nvidia's CUDA-X libraries, from eight times on package thermal modelling to 50 times on image processing.

Nvidia pledges $250 million of UK compute for agent safety research

For UK researchers this is a meaningful boost in access to serious hardware. Nvidia will donate compute worth about $50 million a year for five years, hosted in the UK through Nvidia Cloud Partners, with dedicated support for Oxford's BOLD lab, UCL's SOFAIR lab and Imperial College London. The research will use Nvidia's OpenShell secure runtime and Sentry reference design to study secure agents and open-model security. Provider selection and commercial terms are still to be agreed, so do not expect allocations straight away. Jensen Huang also announced investments in five UK start-ups, including Boltz, Riverlane and Hiverge.