AI Industry
News Digest

Claude Code Rate Limits Double and Opus API Limits Surge, Backed by SpaceX's Colossus Cluster

If you live inside Claude Code or run heavy Opus workloads through the API, your throttle just got considerably looser. Anthropic has signed for the entirety of SpaceX's Colossus 1 data centre in Memphis (over 220,000 NVIDIA GPUs and 300+ megawatts), and is using the new capacity to double the five-hour Claude Code limits for Pro, Max, Team, and seat-based Enterprise users while removing peak-hour reductions on Pro and Max.

API customers on Tier 1 are reportedly seeing roughly a 1,500% lift on Opus input tokens per minute and 900% on output. The practical takeaway: workflows that previously bumped into limits in the afternoon should now run cleanly through the day, and there is real headroom to push longer agentic runs on Opus without architecting around quotas.

Code with Claude Ships Multi-Agent Orchestration, Outcomes, and Dreaming for Managed Agents

Anthropic's developer day in San Francisco is the bigger story for anyone shipping with Claude Code. Three additions to Claude Managed Agents stand out: multi-agent orchestration for spinning up coordinated fleets to chew through complex work; Outcomes, so you can declare what success looks like and let Claude iterate to the goal; and Dreaming, where Claude inspects past sessions, identifies what it missed, and self-improves between runs.

If you have been hand-rolling supervisor agents, retry loops, or post-hoc evaluation tooling, expect to discard meaningful chunks of glue code over the next few weeks. London follows on 19 May and Tokyo on 10 June, with virtual seats free.

ChatGPT Workspace Agents Move to Credit-Based Pricing, Replacing Custom GPTs for Organisations

If your team has been quietly piloting workspace agents in ChatGPT Business, Enterprise, Edu, or Teachers, the meter started turning today. OpenAI's Codex-powered persistent agents, which can connect to Slack, Salesforce, and the rest of an organisation's stack, were free through 5 May and moved to credit-based billing on 6 May. OpenAI is also positioning workspace agents as the explicit successor to custom GPTs for organisations, with a deprecation date for org-level GPTs to follow.

Two practical implications: budget owners need to look at credit consumption quickly before invoices arrive unexpectedly, and if you have org GPTs in production it is worth starting the migration now rather than waiting for the deprecation notice.

Grok Voice Think Fast 1.0 and Standalone Speech APIs Go Live for Developers

xAI has put a credible voice stack on the table. The flagship grok-voice-think-fast-1.0 model is aimed at multi-step workflows, customer support, and sales, with low latency, tool use, and support for 25 or more languages. Alongside it are standalone Grok speech-to-text and text-to-speech APIs, with batch transcription priced at $0.10 per hour and streaming at $0.20 per hour, plus speaker diarisation, timestamps, and expressive speech tags.

The pricing is sharply below most incumbents on the speech-to-text side, and the same stack already runs Grok Voice, Tesla in-car assistants, and Starlink customer support, which gives the underlying reliability story a real reference. Worth a benchmark this week if you have voice agents or call-centre automation in production.

Nvidia Ising: The First Open AI Models for Quantum Processor Calibration

Quantum is normally a slow-burn topic for AI builders, but Nvidia Ising deserves a look because the weights are open. The family targets two specific jobs that have until now been bespoke and slow: AI-based calibration of quantum processors, and quantum error-correction decoding, which Nvidia claims runs up to 2.5 times faster and 3 times more accurately than traditional approaches. For most product teams this is not immediately actionable, but if you are touching scientific computing, materials, drug discovery, or any partner wiring quantum into their roadmap, this is the first credible open baseline you can fork rather than reinvent.

Nvidia and Corning Sign a Multi-Year Deal to Scale US AI Optical-Connectivity Manufacturing

Less glamorous than a model launch but arguably more consequential for cost curves: Nvidia and Corning have signed a multi-year manufacturing partnership to expand US production of advanced optical connectivity for next-generation AI infrastructure. Corning shares jumped 14% on the news, and Nvidia is taking warrants for up to 18 million shares for an aggregate $500 million. Optical interconnect is currently one of the binding constraints on training and inference cluster scale, and a domestic supply ramp moderates one of the supply-chain risks that has been pushing GPU lead times around.

IBM Think 2026: watsonx Orchestrate, Bob, Concert, and Content Cortex Form an Agentic Control Stack

IBM Think 2026 in Boston produced the most concentrated agentic-product launch of the day. The next generation of watsonx Orchestrate (private preview) is positioned as a control plane: bring agents from any source and apply consistent policy and accountability across them. IBM Bob is an end-to-end software development agent covering generation through testing, security, and deployment, distinct from a simple coding assistant.

IBM Concert pulls signals from Instana, Turbonomic, SevOne, and Cloud Pak for AIOps into a single agentic operations platform, and Content Cortex tackles the prepare, govern, and activate problem for unstructured content feeding agents. For product teams: if you have been treating agent governance, observability, and SDLC tooling as separate procurements, IBM is now offering a unified pitch that makes it the most plausible enterprise alternative to assembling this layer yourself.

Dragon Copilot for Nurses Ships Ambient Documentation and Flowsheet Automation

Microsoft used Nurses Week to ship the next slice of Dragon Copilot aimed specifically at nursing workflows, including ambient AI capture and flowsheet population. This is a meaningful capability expansion because flowsheet entry is one of the highest-friction administrative tasks in nursing software, and it has been the gap that prevented the previous physician-focused Dragon Copilot from generalising across a hospital. If you build healthcare integrations or sit anywhere near clinical informatics, expect this to set the floor for what hospital procurement teams now expect from any ambient AI vendor.

DeepMind Opens an AI Research Collaboration with the Makers of EVE Online

DeepMind is taking a research stake in CCP Games, the studio behind EVE Online, to train AI systems on a 20-plus year corpus of player-driven economic and political behaviour. EVE is unusual among games because it has functional markets, alliances, scams, and emergent geopolitics at scale, which makes it a far cleaner sandbox for multi-agent research than most curated benchmarks. There is nothing for builders to use today, but if you work on agents that need to negotiate, coordinate, or recover from adversarial behaviour, the eventual papers and any open data drops are worth tracking. CCP plans more detail at Fanfest 2026.

Mistral Pivots towards Owning Its Own Infrastructure with a Dedicated French Data Centre

At a Marseille event, Mistral CTO Timothée Lacroix confirmed the company's first dedicated data centre will go live before the end of the year at the Eclairion site in Essonne, near Paris. The framing is a deliberate move beyond being a pure model provider towards becoming a full European AI infrastructure player, supported by ASML's 1.3 billion euro investment from last September.

For European customers worried about sovereignty, vendor concentration, or US export-control exposure, Mistral now has a credible end-to-end story rather than just open weights. For everyone else, the relevance is competitive pricing pressure: in-house compute is one of the few levers that can close the cost gap with the US hyperscalers.

Scale AI Lands a $500 Million Pentagon Contract, a Fivefold Increase on the Prior Deal

The Pentagon awarded Scale AI a $500 million contract for data labelling and decision-support work, up from a $100 million deal in September 2025. The relevance to builders is twofold: this anchors Scale AI's position as the default labelling and evaluation vendor for high-stakes deployments, keeping competitive pricing pressure on alternative annotation providers; and it signals continued government willingness to spend at scale on AI tooling rather than full-system deployments. If you are sourcing labelled data or human evaluation services, expect Scale's enterprise tier to be busier and to push longer SLAs over the next quarter.