Perplexity Computer for Counsel, Gemini 3.5 Flash adds computer use, OpenAI and Broadcom unveil Jalapeño

June 24, 2026 marks a particularly dense day of major announcements: an in-house AI chip at OpenAI, computer use becoming a native capability in Gemini 3.5 Flash, a legal agent at Perplexity, and a new world model from Qwen. Mistral strengthens its enterprise connectors, Genspark launches a design tool powered by Claude, and Cursor integrates directly with Notion. Fourteen notable items and seven brief updates round out an overview spanning nine domains.

Perplexity Computer for Counsel — the Computer agent targets legal professionals

June 24 — Perplexity launches Computer for Counsel, a vertical version of its Computer agent aimed at law firms and legal professionals. The integration connects Computer to legal databases, document management systems, and matter-management systems used every day.

Launch partners

Partner	Role
midpageAI	Legal research with citable sources
LegalZoom	Generation of legal documents
Docusign	Electronic signature
netdocuments	Document management for law firms

Perplexity Computer for Counsel is available to all Pro and Max subscribers. Perplexity’s official changelog does not yet include a dedicated entry for this announcement — the rollout happened via X on June 24.

This vertical version illustrates Perplexity’s strategy: deploying its Computer agent into professional markets with direct integrations into existing business tools, rather than asking users to change their workflows.

🔗 Perplexity Tweet — Computer for Counsel · Follow-up thread

Computer use integrated into Gemini 3.5 Flash — native tool with enterprise safeguards

June 24 — Google announces that computer use is becoming a native tool in Gemini 3.5 Flash. Until now, this capability was only available in a dedicated model (Gemini 2.5 computer use). It is now built directly into the main Flash model, alongside function calling, Search grounding, and Maps grounding.

Gemini 3.5 Flash can therefore see, reason, and act across browser, mobile, and desktop environments from a single model. Google’s priority use cases are continuous software testing and knowledge-work tasks on enterprise applications. A demo is available via Browserbase.

Built-in safety measures

Targeted adversarial training against indirect prompt injection
User confirmation required for irreversible actions
Detection and automatic shutdown when indirect injection is detected
Recommended “defense in depth” approach: sandboxing, human verification, access controls

Available via the Gemini API and the Gemini Enterprise Agent Platform. The blog post is signed by Mateo Quiros, Product Manager at Google DeepMind.

🔗 Google Blog — Computer use in Gemini 3.5 Flash

OpenAI and Broadcom unveil Jalapeño — first AI chip dedicated to LLM inference

June 24 — OpenAI and Broadcom (NASDAQ: AVGO) present Jalapeño, the first hardware accelerator designed entirely in-house by OpenAI, built around the specific needs of LLM inference. Co-developed in nine months — an ASIC development cycle claimed to be the fastest for high-performance semiconductors — Jalapeño is not an adapted general-purpose GPU, but a from-scratch design optimized for LLM inference patterns.

Architecture and partners

Component	Role
Broadcom	Silicon implementation + Tomahawk networking
Celestica	Board/rack integration
Microsoft	Gigawatt data center deployment (late 2026)

Engineering samples are already running ML workloads in the lab, including GPT-5.3-Codex-Spark. Preliminary tests indicate watt-for-watt performance above the current state of the art, according to OpenAI; a detailed technical report is promised in the coming months. The initial gigawatt-scale rollout is planned for late 2026.

“Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems.” — “Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in faster, more reliable, more affordable AI for individuals and businesses, and able to be used to solve bigger problems.” — Greg Brockman, President of OpenAI

Jalapeño fits into OpenAI’s strategy of controlling the entire stack — products, models, and now chips — and marks the first step toward a multi-generational compute platform.

🔗 OpenAI — Jalapeño inference chip · OpenAI Tweet

Qwen-AgentWorld — World model simulating 7 agent environments

June 24 — Qwen (Alibaba) presents Qwen-AgentWorld, a native-language world model (Language World Model, LWM) simulating seven agent environments within a single model: MCP, Search, Terminal, SWE, Web, OS, and Android. The key point: environment modeling is the training objective from the outset, not a post-hoc adaptation.

Two research directions are presented in arXiv paper 2606.24597:

Foundation model for environment simulation — according to Qwen, Qwen-AgentWorld outperforms Claude Opus 4.8 and GPT-5.4 on AgentWorldBench. This claim comes from Qwen, based on their own benchmark; it has not been independently validated at the time of the announcement.
Agent training strengthened by world modeling — Controllable Sim RL (agentic reinforcement learning with the LWM as the simulated environment) outperforms training in real environments. Predictive environment learning transfers to agentic tasks with zero additional fine-tuning.

The main tweet (@Alibaba_Qwen, June 24, 419,000 views) links to arXiv, an official Qwen blog, GitHub, and HuggingFace.

🔗 Qwen Tweet · arXiv 2606.24597 · Qwen Blog · GitHub

Mistral Connectors — Advanced control over enterprise connectors

June 24 — Mistral publishes an engineering post detailing several new capabilities for its Connectors in Mistral Studio, with a focus on security and control in production.

Available new features

Feature	Status	Description
Enriched admin controls	GA	Workspace-level access control, individual enable/disable
API keys with connector scopes	GA	API keys limited to connectors, anti-impersonation in automated workflows
Multi-account connectors	GA	Authentication with multiple accounts, configurable default account
Connectors Debugger	Public Preview	End-to-end analysis of MCP failures in 11 steps
Connectors in Vibe Code	GA	Reuse developer connectors via the `/connectors` command
Connectors in Workflows	Public Preview	Persistent connections for long or scheduled tasks

The catalog now covers more than 60 integrations: Atlassian, Google Drive, Notion, GitHub, Slack, Salesforce, Stripe, and others. The Connectors Debugger covers 11 analysis steps, from server access all the way to opening the MCP session — a diagnostic tool intended for teams deploying agents in production.

🔗 Mistral — More control over connectors · Connectors documentation

Cursor × Notion — task delegation via Cursor SDK

June 24 — Cursor announces that users can now delegate tasks to Cursor directly from Notion. The integration is built on the Cursor SDK: Notion built the integration itself, publishing a technical article on cursor.com explaining the approach. Each cloud agent runs on the same models, the same harness, and the same runtime as Cursor.

The workflow: mention @Cursor in any Notion spec, or assign it a task. Cursor opens a PR that the whole team can review.

This launch illustrates the openness of the Cursor SDK to third-party partners: Notion did not have to wait for an official Cursor integration to bring cloud agents into its own product. The SDK exposes enough primitives for an external partner to build a complete integration and distribute it to its own users.

🔗 Cursor Tweet · cursor.com

Genspark Design — powered by Claude Opus 4.7, merging Build Preview and AI Designer

June 24 — Genspark launches Genspark Design, its visual design tool powered by Claude Opus 4.7. This launch merges two existing products from the platform: Genspark Build Preview and Genspark AI Designer.

Genspark Design capabilities

UI prototypes, videos, HTML animations, and posters, production-ready
Brand design system: Figma file import, reuse across all projects, team sharing
One-click code: turn any app or website design into functional code via Genspark Code
Launch pricing available on genspark.ai/agents?type=design

The pinned announcement tweet by @genspark_ai reached 87,000 views in eight hours. The Claude Opus 4.7 power is mentioned explicitly in the official announcement.

🔗 Genspark Design Tweet · Genspark Design

Notable items

OpenAI: GPT-5.5 Instant updated and DevDay 2026 announced

June 24 — OpenAI announces a new version of GPT-5.5 Instant, its most widely used model. The improvements focus on understanding the intent behind a question, handling complex constraints, and shopping and local recommendations. Rollout is underway for paid users on June 24, free users on June 25. 🔗 OpenAI Tweet

June 23 — OpenAI Developers announces that applications are open for OpenAI DevDay 2026, on September 29 in San Francisco. Applications are open until July 10 on devday.openai.com. 🔗 @OpenAIDevs Tweet

NVIDIA — four announcements in one day

June 24 — NVIDIA concentrates four announcements into a single day:

Cosmos 3 Nano — #1 ranking on MolmoSpaces: a post-trained policy model on Cosmos 3 Nano reaches the top spot on Ai2’s open MolmoSpaces ranking for robot policies, in environments requiring interpretation, reasoning, and action. 🔗 NVIDIA Robotics Tweet
Metropolis Blueprint VSS 3: version 3 of the blueprint for video search and summarization brings 16 new agent skills — search, summarization, alerts, reports — and live stream analysis in natural language. 🔗 NVIDIAAI Tweet
NeMo AutoModel + Transformers v5: NeMo AutoModel now relies on HuggingFace Transformers v5 for first-class support of MoE (Mixture of Experts) models, with 1.3x throughput gains (TPS/GPU) and a 30% reduction in peak memory. 🔗 NVIDIAAI Tweet
BioNeMo Agent Toolkit: an open toolkit that gives any AI agent callable tools for protein structure prediction, molecular docking, generative chemistry, and genomic analysis. 🔗 NVIDIA Health Tweet

Hugging Face — Moon Bot and Ai2 MolmoSpaces

June 24 — Hugging Face publishes the complete architecture of Moon Bot, their internal Slack assistant. Moon Bot relies on the Pi SDK (open-source coding agent) with a configurable LLM (Kimi K2, Claude, or another). Each Slack session’s memory is serialized in JSONL in a private HF Buckets bucket (huggingface/moon-bot-memory). Three Okta-based access tiers grant access to Elasticsearch, MongoDB, GitHub CLI, and other tools depending on the profile. The article aims to serve as a replicable model for other teams. 🔗 Hugging Face Article

June 24 — Ai2 announces MolmoSpaces, its open ranking for robot policies. Any team can submit a policy and try to dethrone the current leader — NVIDIA Cosmos 3 Nano. 🔗 Ai2 Tweet · Leaderboard

Coding tools: Zed v1.8, Copilot Free/Student, Manus Hosting

June 24 — Zed v1.8: the new agent.terminal_init_command setting automatically runs a command when each new terminal thread opens in the agent panel. Git performance is improved (HEAD resolution, uncommit, remote listing) with fewer spawned processes. Two new editor commands: editor: select {inside,around} delimiters and workspace: reset pane sizes. Diff scrolling, multi-cursor editing, and search in the Markdown preview are also faster. 🔗 Zed Tweet

June 24 — GitHub Copilot switches to exclusive auto mode for Free and Student plans: manual model selection is removed in favor of dynamic routing to the best available model. The (Preview) labels for Microsoft models are removed. 🔗 GitHub Changelog

June 24 — Manus Website Builder introduces two hosting modes: Autoscale (0 to 5 instances depending on traffic, zero cost when idle) and Reserved (persistent 24/7 instance, no cold start, suited to real-time dashboards and bots, up to $36/month at maximum usage). Each account includes $10 in free monthly credit. Mode switching happens with no service interruption and no code changes. Note: Manus is part of Meta. 🔗 Manus Blog

Media: ElevenLabs × FOX Sports, xAI MongoDB, Kimi on AWS

June 24 — ElevenLabs and FOX Sports are launching a voice agent powered by Colin Cowherd’s voice in the FOX Sports app. The agent answers questions about the FIFA World Cup group stage, fantasy football picks, and offers sports analysis. It can argue if the user is wrong. The collaboration is based on an official agreement between ElevenLabs and Colin Cowherd. 🔗 ElevenLabs Tweet

June 24 — xAI Grok Build is welcoming the official MongoDB plugin in its marketplace: data querying, index optimization, and database management directly from the Grok Build CLI (329,000 views on the announcement tweet). 🔗 xAI Tweet

June 24 — The Kimi API (Kimi-K2.7 Code model) is now available on AWS Marketplace with consolidated billing. Eligible customers can apply usage against their EDP commitments (Enterprise Discount Program). 🔗 Kimi Tweet · AWS Marketplace

Project Genie — Cannes Lions AI Craft Grand Prix

June 23 — Google DeepMind’s Project Genie prototype (generative Playable Worlds) wins the Grand Prix in the AI Craft category at Cannes Lions 2026. Jack Parker-Holder, who leads the project, cites the ability to achieve results unattainable without cutting-edge AI. 🔗 @jparkerholder Tweet

Briefs

Claude Code v2.1.190 (June 24) — Maintenance release: bug fixes and reliability improvements, published after v2.1.187. 🔗 Claude Code Changelog
Hugging Face — weekly huggingface_hub release (June 23) — The HF team describes its automated release pipeline combining AI agents and human-in-the-loop validation for the Hub’s Python client library. 🔗 HF Article
Hugging Face — Cross-Origin Storage in Transformers.js (June 23) — Experimentation with the COS API proposal to share model weights across distinct web origins, reducing browser downloads. 🔗 HF Article
Google DeepMind Podcast — agentic economies (June 24) — An episode about what happens when millions of AI agents negotiate, transact, and delegate tasks, with a focus on diversifying decision-making to avoid AI groupthink. 🔗 @GoogleDeepMind Tweet
Manus — automatic URL suggestions (June 23) — Manus now suggests relevant web pages from a natural-language description in the Files & sources tab, saved as references for subsequent tasks. 🔗 Manus Tweet
Runway Localize Ads (June 24) — New feature: a single advertising image as input, localized versions for all markets as output, with one click. 🔗 Runway Tweet
Luma Connectors (June 24) — Airtable, Dropbox, and Google Drive integrations directly in Luma to pull files into boards on demand. 🔗 Luma Tweet

What this means

The verticalization of AI agents is accelerating. Perplexity Computer for Counsel is not a general-purpose version with legal documentation bolted on — it is a direct integration with the business tools (LegalZoom, Docusign, netdocuments) that lawyers use every day. The same pattern appears at ElevenLabs with FOX Sports: rather than a generic AI tool, a voice agent for a real sports commentator, distributed inside the app of an existing media company. The value is no longer in the base model but in vertical integration and distribution.

Computer use is moving from experimental to standard capability. Google is integrating computer use directly into Gemini 3.5 Flash — its fastest general-purpose model — rather than into a dedicated model. This move signals that the capability is mature enough to be embedded in the main model without extra cost. The emphasis on safeguards (automatic stop, confirmation for irreversible actions) indicates that Google is aiming for real enterprise deployment, not just demos.

OpenAI is moving up the hardware stack. Jalapeño represents a shift in OpenAI’s strategy: from NVIDIA customer to designer of its own accelerators. Built in nine months with Broadcom — a short cycle for a high-performance ASIC — and aimed at gigawatt-scale deployments by late 2026, Jalapeño is targeting lower inference costs to make AI “faster, more reliable, more affordable.” This evolution aligns with the trend seen at Google (TPU), Amazon (Trainium), and Apple (Neural Engine): major platforms are trying to control their own silicon for AI.

Environment modeling as a line of agentic research. Qwen-AgentWorld is pushing in a less-explored direction: training a model to simulate agent environments (terminal, browser, OS, Android) rather than acting in them directly. The idea is that predictive knowledge of the environment transfers to real agentic tasks without additional fine-tuning. If the claim of outperforming Claude Opus 4.8 and GPT-5.4 on AgentWorldBench is confirmed by third parties, this LWM paradigm will deserve sustained attention.

Production-grade agent tooling is becoming professionalized. Mistral Connectors, with its 11-step MCP Debugger, API key scopes, and persistent connections for long-running workflows — these are not demo features. They are responses to concrete problems teams face when deploying agents in production: connection failures that are hard to diagnose, identity spoofing risks in automated workflows, long-session management. The same operational maturity can be seen in Cursor SDK opening up to Notion, HF’s Moon Bot with its three Okta access levels, and Zed v1.8 with agent.terminal_init_command.