Runway Agent, Grok Build CLI, ChatGPT Personal Finance, Copilot App Desktop and 2028 AI competition scenarios

A packed week: Runway Agent turns video production into a conversation, xAI joins the race for CLI coding agents with Grok Build, ChatGPT tackles personal finance via Plaid, and GitHub ships its Copilot App desktop in technical preview. Meanwhile, Anthropic publishes a policy paper on U.S./China AI competition on the 2028 horizon, and Perplexity Computer connects to Snowflake for enterprise analytics.

Runway Agent — AI creative partner for end-to-end video production

May 13 — Runway launches Runway Agent, a conversational creative partner capable of producing complete videos — sound design and editing included — in a single session from a simple conversation. The demo published on May 15 shows the creation of a full ad from a product shot, with no manual intervention between steps.

Capability	Description
End-to-end production	Ideation + sound design + editing in one conversation
Content types	Ads, shorts, social content
Entry point	app.runwayml.com/agent
X engagement	499,800 views, 237 bookmarks

The difference from usual generative tools is structural: Runway Agent does not just produce isolated clips but orchestrates the entire creative pipeline. The user describes their need (“an ad for this product, energetic tone, 30 seconds”), and the agent plans, generates, assembles, and delivers.

Meet Runway Agent. Your new AI creative partner that helps you ideate and execute fully finished, sound designed and edited videos. All with just a simple conversation. — Discover Runway Agent. Your new AI creative partner that helps you imagine and produce fully finished videos, with sound design and editing. All with a simple conversation.

The launch comes one week after the announcement of the expansion into Japan (see next section), signaling a strategy of simultaneous geographic and product growth.

🔗 Runway Agent

Grok Build Early Beta — terminal-native CLI coding agent by xAI

May 14 — xAI launches Grok Build in open beta: a terminal-native CLI agent for software development, a direct competitor to Claude Code and Gemini CLI. Available first to SuperGrok Heavy subscribers ( $99/month introductory price, normally$ 299/month), the CLI installs via curl -fsSL https://x.ai/cli/install.sh | bash.

Feature	Detail
Underlying model	Grok 4.3 beta, Heavy architecture with 16 agents
Context window	2M tokens
Plan/review/approve	Plan mode to validate before execution, readable diffs
Parallel sub-agents	Delegation to specialists in parallel, worktree support
Integrations	AGENTS.md, plugins, hooks, skills, MCP servers compatible
Headless mode	Flag `-p`, full ACP support for scripts and automations
Feedback	Built-in `/feedback` command

Grok Build enters direct competition in the CLI coding agent segment, which now has three players: Claude Code (Anthropic), Gemini CLI (Google), and Grok Build (xAI). The stated differentiator is out-of-the-box compatibility with existing tools (AGENTS.md, MCP servers) and Grok 4.3’s parallel multi-agent architecture. The announcement generated 53 million views on X.

🔗 Grok Build

ChatGPT Personal Finance — bank account connection via Plaid

May 15 — OpenAI launches, in preview for Pro subscribers in the United States, a new personal finance feature in ChatGPT. Users can now connect their financial accounts via Plaid, access a unified dashboard, and ask natural-language questions based on their real data.

Plan	Access	Region
Pro	Immediate preview	United States only
Plus	Coming soon	—
Free	Long-term goal	—

Key features:

Connection to more than 12,000 financial institutions via Plaid (Intuit integration coming)
Unified dashboard: portfolio, spending, subscriptions, upcoming payments
Financial memories (Financial memories): ChatGPT remembers goals, debts, shared projects
Default model: GPT-5.5 Thinking (GPT-5.5 Pro for Pro subscribers)
ChatGPT cannot make transactions or see full account numbers
Account disconnection triggers data deletion within 30 days

In an internal benchmark evaluated with more than 50 finance professionals: GPT-5.5 Thinking scores 79/100 and GPT-5.5 Pro 82.5/100 on complex personal finance tasks. The partnership with Intuit will make it possible to go beyond answers — credit card recommendation with eligibility check and online application, tax implications with expert consultation.

🔗 ChatGPT Personal Finance

GitHub Copilot App — desktop app in technical preview

May 14 — GitHub launches the GitHub Copilot App in technical preview: a native GitHub desktop application for agentic development. Copilot Business and Enterprise get access as soon as the organization enables it (rollout during the week of May 14); Copilot Pro and Pro+ must sign up for early access with gradual rollout.

Feature	Description
Isolated sessions	Each session has its own branch, files, conversation, and task state
Start from GitHub context	Start from an issue, PR, prompt, or previous session
Unified inbox	View of issues/PRs to handle across all connected repositories
Pause/resume	Leave a session and return exactly where it was left off
Automated workflows	Skills + prompts for triage, dependencies, release notes, cleanup
Integrated terminal + browser	Validate changes directly in the app
Agent Merge	The agent handles review comments, fixes checks, and merges automatically

Admin prerequisites: previews enabled + Copilot CLI enabled in policy settings.

The app represents the convergence of Copilot’s agentic capabilities in a dedicated desktop environment, distinct from the IDE plugin and the web interface. Agent Merge is the most advanced feature: the agent reads review comments, applies fixes, resolves blocking checks, and merges — without human intervention in the intermediate steps.

🔗 Copilot App Technical Preview

Copilot CLI agent and unified sessions in JetBrains IDEs

May 13 — GitHub ships a major update to the GitHub Copilot for JetBrains IDEs plugin with the introduction of the Copilot CLI agent in public preview.

Feature	Description
Copilot CLI agent	Delegate long tasks from JetBrains to a local CLI agent
Worktree isolation	The agent works in a separate Git worktree, without touching the current branch
Workspace isolation	The agent applies changes directly for fast iteration
Unified sessions view	All agent sessions with status, type, duration in one place
Ask question tool	The agent can ask targeted clarification questions
Global .agent.md	Custom agents in `~/.copilot/agents`, available everywhere
GHES sign-in support	GitHub Enterprise Server in the authentication flow

Deprecation: the “Edit” mode is removed in this update.

Worktree isolation is the key point: the CLI agent works in a dedicated, isolated Git branch, allowing the user to keep working on their main branch while the agent performs long tasks in parallel.

🔗 Copilot CLI agent JetBrains

Anthropic — 2028 scenarios for global AI competition

May 14 — Anthropic published a policy paper titled “2028: Two scenarios for global AI leadership”, presenting its positions on AI competition between the United States and China. The article generated 3.8 million views on X in 24 hours.

The paper describes two hypothetical scenarios for 2028:

Scenario 1 — Consolidated democratic lead: The United States and its allies maintain a 12- to 24-month lead on the model frontier. Semiconductor export controls are strengthened, distillation attacks are countered, and American AI dominates the global economy.

Scenario 2 — Neck-and-neck race with China: The Chinese Communist Party manages to stay at the frontier level through export-control circumvention (chip smuggling, offshore data centers) and massive distillation attacks on American models.

Indicator	Value
Views on X	3.8 million
Recommended U.S. target lead	12-24 months in 2028
Huawei vs NVIDIA compute 2026	4% in total performance
Huawei vs NVIDIA compute 2027	2% (estimated)

Anthropic identifies four fronts of competition: intelligence (model capabilities), domestic adoption, global distribution, and resilience. Three policy action areas are recommended: close loopholes in export controls, counter illicit distillation attacks, and promote global export of American AI.

We’ve published a paper that explains our views on AI competition between the US and China. The US and democratic allies hold the lead in frontier AI today. — We published a paper that explains our positions on AI competition between the United States and China. The United States and its democratic allies hold the lead in frontier AI today.

🔗 2028: Two scenarios for global AI leadership

Perplexity Computer + Snowflake — enterprise workflows on live data

May 14 — Perplexity Computer now connects to Snowflake, making it possible to run end-to-end workflows directly on live data from a data warehouse. Documented use cases cover pipeline analysis, product usage tracking, and customer analytics.

The integration positions Perplexity Computer in the enterprise self-service analytics segment: an analyst can ask natural-language questions about their Snowflake data and get automated workflows — extraction, transformation, visualization — without code. The announcement generated 60,761 views and 99 bookmarks on X, with significant enterprise traction.

🔗 Perplexity Computer + Snowflake

PayPal — 74,000 weekly tasks on Perplexity Enterprise

May 13 — Perplexity publishes the PayPal customer case: the company runs 74,000 weekly tasks via Perplexity Enterprise for model validation and channel analytics use cases. This is, to date, the highest publicly disclosed enterprise usage figure Perplexity has communicated for an individual customer. The announcement, accompanied by a blog post detailing PayPal’s workflows, generated 61,442 views and 81 bookmarks on X — significant engagement for a B2B news item. This customer case, combined with the Snowflake integration announced the following day, outlines a coherent enterprise upmarket strategy for Perplexity Computer.

🔗 PayPal + Perplexity Enterprise

Codex in the ChatGPT mobile app (iOS and Android)

May 14 — Codex is now available in the ChatGPT mobile app (iOS and Android), in preview on all plans including Free and Go. OpenAI says 4 million users use Codex every week.

Feature	Description
Multi-machine connection	Laptop, Mac mini, managed remote environment
Real-time state	Screenshots, terminal output, diffs, test results
Actions from mobile	Start, unblock, approve, change model
Remote SSH	Now generally available
Hooks	Generally available — prompt scans, validators, custom memories
Programmable tokens	For CI pipelines, available on Enterprise and Business
HIPAA compliance	For Codex on local ChatGPT Enterprise workspaces

You’ve been asking for this one… Now in preview: Codex in the ChatGPT mobile app. Start new work, review outputs, steer execution, and approve next steps, all from the ChatGPT mobile app. Codex will keep running on your laptop, Mac mini, or devbox. — You were waiting for it… Now in preview: Codex in the ChatGPT mobile app. Start new work, review results, guide execution, and approve next steps, all from the ChatGPT mobile app. Codex will continue running on your laptop, Mac mini, or devbox.

🔗 Codex on mobile

Anthropic — Claude for Small Business, Gates Foundation, PwC

Claude for Small Business

May 13 — Anthropic launches Claude for Small Business: 15 agentic workflows and 15 preconfigured skills for small businesses, via Claude Cowork. Connected tools include QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365. Features cover payroll, monthly close, cash tracking, marketing campaigns, and invoice follow-up. A “Claude SMB Tour” kicks off on May 14 in Chicago (100 SMB leaders per stop, one month of Claude Max included).

🔗 Claude for Small Business

$200 million partnership with the Gates Foundation

May 14 — Anthropic and the Bill & Melinda Gates Foundation announce a $200 million partnership over four years. Claude credits, grants, and technical support cover four areas: global health (vaccines and therapies for low-income-country diseases — polio, HPV, preeclampsia), life sciences, K-12 education in sub-Saharan Africa and India, and economic mobility for solo founders and farmers. The target population on the health side represents 4.6 billion people without access to essential care.

🔗 Gates Foundation Partnership

Expanded alliance with PwC

May 14 — PwC will deploy Claude Code and Claude Cowork to its U.S. teams, then globally. The alliance creates a joint Center of Excellence and a certification program for 30,000 PwC professionals. Current production deployments report gains of up to 70%: insurance underwriting cut from 10 weeks to 10 days, cybersecurity incident response from several hours to a few minutes. A new “Office of the CFO” unit is launching, targeting regulated sectors.

🔗 PwC + Anthropic

Claude Code v2.1.141-142 — terminalSequence hooks, claude agents flags + Fast mode Opus 4.7

Two releases published on May 14 and 15, 2026 bring substantial improvements to Claude Code, especially in hooks management, background sessions, and Fast mode.

New feature (v2.1.142)	Details
New flags `claude agents`	`--add-dir`, `--settings`, `--mcp-config`, `--plugin-dir`, `--permission-mode`, `--model`, `--effort`, `--dangerously-skip-permissions`
Fast mode → Opus 4.7	Now the default (variable `CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1` to revert to Opus 4.6)
Root SKILL.md plugins	A plugin with `SKILL.md` at the root without a `skills/` subfolder is now exposed as a skill

New feature (v2.1.141)	Details
`terminalSequence` in hooks	Desktop notifications, window titles, ringtones without a control terminal
`claude agents --cwd <chemin>`	Filter the session list by directory
`ANTHROPIC_WORKSPACE_ID`	Workload identity federation
Rewind “Summarize up to here”	Compress earlier context from the Rewind menu
Amber spinner	Indicates that Claude is still working after 10 seconds

Notable fixes (v2.1.142): MCP_TOOL_TIMEOUT no longer limited HTTP/SSE calls to 60 seconds; daemon crash loop after brew upgrade; Windows deadlock on network drives. v2.1.141 also included 40+ fixes for background sessions, MCP plugins, hooks, and terminal rendering.

🔗 Claude Code CHANGELOG

Claude Code — Weekly limits +50% until July 13

May 13 — @ClaudeDevs announces a 50% increase in Claude Code weekly limits until July 13, 2026 at 6 PM PDT, for all Pro, Max, Team, and Enterprise plans. The increase applies to the CLI, IDE extensions, desktop, and web — no user action required. It stacks with the doubling of 5-hour limits announced on May 6.

🔗 ClaudeDevs on X

GitHub — Additional Copilot updates

Copilot Memory — user preferences (Pro and Pro+)

May 15 — Copilot Memory expands its scope to the user level (early access). Previously, memory only stored repository-level information. Now, personal preferences (commit style, pull request structure, communication tone) follow the user across all their repositories and agents. Manageable in Copilot Memory personal settings.

🔗 Copilot Memory User Level

Copilot cloud agent — automatic model selection

May 14 — The “Auto” mode in the cloud agent model picker lets Copilot intelligently choose the best model based on system health and performance. Benefits: 10% discount on the normal multiplier, no weekly rate limit with this mode.

🔗 Auto model selection

Agent tasks REST API (Business/Enterprise)

May 13 — Copilot Business and Enterprise users can now launch cloud agent tasks through the new Agent tasks REST API (public preview). Use cases: large-scale refactoring across multiple repositories, bootstrapping repositories from an internal developer portal, automatic preparation of release notes. Authentication via PAT (classic and fine-grained) and OAuth tokens. GitHub App installation tokens and Copilot Pro/Pro+ support coming soon.

🔗 Agent tasks REST API

Copilot metrics by team via API

May 14 — The Copilot metrics API exposes a new user-teams report via two endpoints (/enterprises/{enterprise}/copilot/metrics/reports/user-teams-1-day and /orgs/{org}/copilot/metrics/reports/user-teams-1-day). It can produce team adoption metrics (active users, completions, chats, by language/IDE/model). Limitation: teams with fewer than 5 users are excluded.

🔗 Team-level Copilot metrics

GitHub Actions — June 2026 image migrations

May 14 — Three GitHub Actions migrations to plan for:

Migration	Date
Arm64 images taken over by GitHub (from Arm Limited)	In progress
`windows-latest` / `windows-2025` → VS 2026	June 8-15, 2026
`macos-latest` → macOS 26	Starting June 15, 2026, 30 days

To prepare: runs-on: windows-2025-vs2026 (test VS 2026), runs-on: windows-2022 (stay on VS 2022), runs-on: macos-15 (stay on macOS 15).

🔗 Upcoming image migrations

Runway expansion Japan — Tokyo office + 40 million USD

May 15 — Runway is opening an office in Tokyo with an initial investment of 40 million USD. Japan has become Runway’s 3rd largest market worldwide, with enterprise customers tripling in 12 months. Customers: Yamaha, NHN, SoftBank — in marketing, advertising, and creative content sectors. The expansion is aimed primarily at Japanese leaders in gaming, media, and robotics. The launch comes in the same week as Runway Agent (see section above), signaling a simultaneous growth strategy across product and geography. The announcement generated 86,600 views and 139 bookmarks on X.

🔗 Runway in Japan

AI Media — NVIDIA OpenShell, FLUX Outpainting, ElevenCreative

NVIDIA OpenShell v0.0.41

May 15 — NVIDIA releases version v0.0.41 of its open source compute tool OpenShell. New features: agent-driven policy management, sandbox resource flags in the CLI, custom CA support for OIDC TLS verification, sandbox downloads with workspace boundary controls. Follow-up to v0.0.37 (May 11).

🔗 OpenShell v0.0.41 on GitHub

Black Forest Labs — FLUX Outpainting

May 14 — Black Forest Labs launches FLUX Outpainting, which makes it possible to extend any image to any aspect ratio. The key breakthrough: common edge consistency problems (light cutoffs, texture drift, visible seams) are handled at the model level — not in post-processing. The tool takes an image and canvas geometry, and produces a coherent scene extension.

🔗 Black Forest Labs on X

ElevenLabs — Templates in ElevenCreative

May 13 — ElevenLabs launches Templates in ElevenCreative: ready-to-use creative workflows (input upload → template selection → output). Use cases: producing hundreds of ad variants, turning selfies into animated cartoon characters. Available on web and mobile with the same number of steps regardless of the project. 625,254 views on the X announcement.

🔗 ElevenCreative Templates

xAI — API consolidation and Hermes Agent

Retirement of 8 API models — redirect to Grok 4.3

May 15 — xAI is retiring 8 models from its API at 12 PM PT: grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-4-fast-reasoning, grok-4-fast-non-reasoning, grok-4-0709, grok-code-fast-1, grok-3, grok-imagine-image-pro. The slugs remain valid but redirect automatically.

Redirect	Pricing
Text models → `grok-4.3`	$1.25/1M input tokens, $2.50/1M output tokens
Reasoning models → `grok-4.3` with `low reasoning effort`	—
Non-reasoning models → `grok-4.3` with `none reasoning effort`	—
`grok-imagine-image-pro` → `grok-imagine-image-quality`	—

Developers are encouraged to migrate explicitly to control the billed reasoning level.

🔗 xAI migration — May 15

Grok in Nous Research’s Hermes Agent

May 15 — xAI integrates with Hermes Agent, the open-source persistent-memory agent from Nous Research. Grok subscribers (any tier) can use their subscription in Hermes, which runs continuously on desktop, sandbox, or VPS. Hermes accesses Grok 4.3 (text and reasoning), Grok TTS (voice responses), and Grok Imagine (images/videos). Compatible with WhatsApp, Discord, Telegram, Signal. Installation: curl -fsSL .../install.sh | bash + hermes model → select xAI Grok OAuth.

🔗 Grok + Hermes Agent

Perplexity Computer — security by default architecture

May 13 — Perplexity publishes the technical details of the security architecture of Perplexity Computer. Each task runs in its own hardware-isolated sandbox with VPC network isolation. External content is scanned in parallel by ML classifiers and the BrowseSafe model before agents act. A multilayer architecture designed for computer use agents in enterprise environments.

🔗 Perplexity Computer security

ChatGPT — Security in sensitive conversations

May 14 — OpenAI details ChatGPT Safety summaries: short factual notes on relevant security context from previous conversations, generated by a model trained for safety reasoning. The mechanism addresses the case where a message seems harmless in isolation but reveals a risk in the context of a long conversation.

Scenario	Improvement
Suicide/self-harm (long conversation)	+50% safe responses
Harm to others (long conversation)	+16%
Harm to others (GPT-5.5 Instant, multi-conversation)	+52%
Suicide/self-harm (GPT-5.5 Instant, multi-conversation)	+39%

Relevance score for summaries: 4.93/5 across more than 4,000 evaluations. Developed with psychiatrists and psychologists specializing in suicide prevention and forensic psychology. No degradation detected in ordinary conversations.

🔗 Safety summaries ChatGPT

OpenAI — Response to the TanStack npm supply chain attack

May 13 — OpenAI was affected by the “Mini Shai-Hulud” attack targeting the open-source TanStack npm library, identified on May 11. Two employee devices were compromised, with limited access to internal source code repositories. No user data compromised, no intellectual property impacted.

macOS users must update their applications before June 12, 2026 — after that date, versions signed with the old certificates will no longer work:

macOS application	Last version with old certificate
ChatGPT Desktop	1.2026.125
Codex App	26.506.31421
Codex CLI	0.130.0
Atlas	1.2026.119.1

No action required for Windows and iOS.

🔗 OpenAI response — TanStack

Gemini CLI v0.42.0

May 13 — Gemini CLI releases version v0.42.0 with three major features:

New feature	Description
Auto Memory Inbox	Canonical skill management via the “canonical-patch” contract; review and validation of memories extracted during session
Gemma 4 by default	Enabled for all users via the Gemini API, with no additional configuration
Vocal mode improvements	Visual wave animations, privacy and compliance warnings for Gemini Live

Other changes: --delete flag for the /exit command, new /commands list command, Unicode LaTeX rendering in the terminal.

🔗 Gemini CLI v0.42.0

Cohere Project Pursue — Compass on declassified UFO archives

May 14 — Cohere launches Project Pursue, a fictional declassified UFO/UAP archive (90s style, “Netscape Navigator 4.7” aesthetic), based on real declassified U.S. government documents published on May 8, 2026 (FBI, CIA, State Department). The demo is a showcase for Cohere Compass, their intelligent search product for unstructured data — scans of handwritten and typed documents. Creative campaign with 35,320 views and 126 bookmarks.

🔗 Project Pursue

Briefs

Grok Code Fast 1 deprecated on Copilot — deprecation effective May 15 across all GitHub Copilot experiences (Chat, inline edits, ask/agent, completion). Suggested alternatives: GPT-5 mini or Claude Haiku 4.5. 🔗 Announcement
Timestamp fields in GitHub Projects — three new native fields (Created, Updated, Closed) can be added to any view to sort and filter by creation, modification, or closure dates. 🔗 Announcement
Enterprise Installation API — public preview — a GitHub App can now discover its installation on an enterprise and retrieve its installation ID via the new API, closing a gap versus the existing org/repo/user APIs. 🔗 Announcement
Runway AI Festival 4th edition — 10 finalist films — 10 films selected for Runway’s 4th international AI festival. NY (June 11) and LA (June 18, 2026) events. 🔗 Announcement
Suno — enriched user profiles — cover image upload, bio editing, adding social links, pinning 5 songs with captions, playlist display. 🔗 Suno on X
Codex on Windows — engineering article — David Wiesen (OpenAI) details Codex’s Windows sandbox architecture: synthetic SIDs, write-restricted tokens, dedicated users (CodexSandboxOffline, CodexSandboxOnline), Windows Firewall rules. Conclusion: Windows does not provide a single primitive; the solution is a composition of several mechanisms. 🔗 Article

What this means

The agentic coding CLI is becoming a three-way battleground. In a few weeks, Claude Code (Anthropic), Gemini CLI (Google), and Grok Build (xAI) find themselves in direct competition in the same segment: the development agent controlled from the terminal. Grok Build brings 2M tokens of context and a parallel multi-agent architecture; Claude Code responds with +50% weekly limits and Fast mode on Opus 4.7; GitHub Copilot unifies the experience in a dedicated desktop app and a strengthened JetBrains plugin. The question is no longer “can an agent code?” but “which agent integrates best into my existing workflow?”

AI is moving into financial and enterprise workflows at scale. ChatGPT Personal Finance with Plaid marks OpenAI’s entry into a highly sensitive sector — 12,000 financial institutions, GPT-5.5 Thinking by default, an Intuit partnership to go as far as credit requests. At the same time, Perplexity Computer connects Snowflake (60,761 views on the announcement) and cites PayPal at 74,000 weekly tasks. PwC and Anthropic announce 30,000 certifications and 70% gains. AI no longer just “drives” experiments — it is being integrated into financial systems and production data pipelines.

Creative content production is shifting toward conversational orchestration. Runway Agent does not generate clips — it produces complete videos (sound + editing) from a conversation. FLUX Outpainting solves edge consistency problems at the model level. ElevenCreative adds standardized templates to industrialize ad production. The common trend: creative AI is moving from a one-off tool to a full production chain, with fewer manual interventions between steps.

The geopolitics of AI is entering the institutional register. Anthropic’s paper on 2028 scenarios (3.8M views) and the Gates Foundation partnership (USD 200M) signal that frontier labs are no longer positioning themselves only as technology companies but as actors in global public policy. The U.S./China competition over semiconductors and frontier models is becoming as much a commercial argument as a national security concern. The PwC partnership, certifications for 30,000 professionals, and Runway’s expansion in Japan (USD 40M) illustrate that large organizations are integrating AI into their permanent structures — not as an experiment.