Claude Code v2.1.191, Gemini Study Notebooks and Runway Agent 2.0: AI watch for June 25, 2026

June 25, 2026 is a day packed with concrete announcements: Claude Code moves to version 2.1.191 with the /rewind command and a 37% reduction in CPU consumption in streaming; Google rolls out Study Notebooks in Gemini worldwide; Runway launches Agent 2.0 for marketing teams; OpenAI publishes an economic study showing that Codex now represents 99.8% of tokens generated internally. On the research side, Cursor questions the reliability of public benchmarks through reward hacking, and Ai2 compares transformer and hybrid architectures token by token. Cohere publishes two open-source technical articles on fork maintenance and agent-based security, while ElevenLabs integrates Google DeepMind’s SynthID audio watermark.

Claude Code v2.1.191 — /rewind, MCP reliability, and -37% CPU

June 25 — Anthropic releases version 2.1.191 of Claude Code. The main new feature is the /rewind command, which lets you resume a conversation from the state before /clear was run — a practical safety net for long work sessions where you regret having erased the context.

Notable performance improvements

Domain	Improvement
CPU streaming	-37% consumption thanks to grouping text updates over 100 ms
Memory	Reduced memory growth on long sessions via the terminal cache
MCP network	Automatic retry with short backoff on transient errors for tools/list, prompts/list, resources/list
MCP OAuth	Retry after network error; headless environments without browser popups
MCP errors	HTTP 404 errors now display the URL and point to the MCP config

Significant bug fixes

Several regressions are fixed in this release. Background agents no longer came back to life unexpectedly after being stopped from the task panel — stopping is now permanent. The generic “not available” message from /voice is replaced with an explanatory message when an organizational policy disables the feature. Hooks with comma-separated matchers (e.g. "Bash,PowerShell") never silently triggered — fixed. The “Recently-denied” tab in /permissions now persists approvals on close. Searching history in vim mode (NORMAL /) suggests how to access slash commands.

The fix for forceRemoteSettingsRefresh in managed settings (MDM or policy) is particularly useful for teams deploying Claude Code in enterprises with a proxy.

🔗 Claude Code changelog

Gemini App — Study Notebooks launched globally

June 25 — Google launches Study Notebooks in the Gemini app, an adaptive learning space for students, available in all languages worldwide.

It works through an initial diagnostic quiz: the user uploads their course materials, Gemini assesses their gaps, then proposes personalized lessons in short sequences. A progress dashboard classifies topics into three labels — Strengths, Focus areas, Not started — across more than 100 learning objectives.

Standardized test preparation

Exam	Status at launch
SAT	Available (Princeton Review questions)
ACT	Coming soon (summer 2026)
GRE	Coming soon (summer 2026)
JEE (India)	Coming soon (summer 2026)
NEET (India)	Coming soon (summer 2026)
ENEM (Brazil)	Coming soon (summer 2026)

The sources shared between Gemini App and NotebookLM make it possible to generate flashcards and Video Overviews directly from the same documents. The global rollout covers all languages at launch.

🔗 Google Blog — Gemini Study Notebooks

Gemini CLI v0.49.0 stable and nightly with tool registry discovery

June 25 — The stable v0.49.0 release of Gemini CLI is published as “Latest” on GitHub Releases, accompanied by a nightly v0.49.0-nightly.20260625.gd845bc5d4 introducing tool registry discovery.

Contents of stable v0.49.0

Fix / Feature	Detail
Skills security	Prevention of path traversal during skills installation (#27767)
Zero quota	Fast failure to avoid retry loops (#27698)
Config migration	`coreTools` renamed to `tools.core` (#27947)
GDC air-gapped	Support for service identity in disconnected GDC environments (#27956)
Tmux	Fix for the background detection false positive (#27572)
eval:inventory	New CLI command with JSON output (#28009, #28058)

The June 25 nightly and preview version v0.50.0-preview.1 both introduce the tool registry discovery feature (#28113), which allows the tool to dynamically find available tools via a centralized registry. The path traversal fix during skills installation is particularly relevant for production deployments.

🔗 GitHub Releases — Gemini CLI

Runway Agent 2.0 — from campaign to deliverable in one prompt

June 25 — Runway launches Agent 2.0, a major upgrade to its marketing-focused AI agent. Starting from a simple prompt describing a campaign, the agent builds a complete marketing brief, generates visual assets, and produces advertising videos directly in the Runway interface.

The agent is designed to understand the context of an existing campaign, ask targeted questions to the user, and iterate until the expected result is reached. The targeted teams cover four profiles: brand, performance, social, and product marketing.

Announced roadmap

Direct connections to advertising platforms (Meta, YouTube, TikTok, Google)
Analysis of performance data to automatically optimize creatives
Automatic asset generation based on real campaign metrics

Agent 2.0 is available to all Runway users. The roadmap distinguishes current capabilities (brief + assets + video) from direct ad platform connections, which are still to come.

🔗 Runway blog — Agent 2.0 · Runway tweet

ElevenLabs × Google DeepMind SynthID — inaudible audio watermark

June 25 — ElevenLabs announces a partnership with Google DeepMind to integrate SynthID into all audio generated by its platform. SynthID is an inaudible digital watermark designed to survive common audio processing steps (compression, re-encoding, equalization).

The partnership’s stated goal is to address the growing difficulty of distinguishing AI-generated audio from human audio as models improve. The watermark is built directly into the ElevenLabs audio generation flow, with no separate post-processing step.

What this changes in practice

Element	Detail
Scope	All audio generated by ElevenLabs
Watermark	Inaudible, persistent after audio processing
Detection	ElevenLabs Audio Detector — free service
Technology partner	Google DeepMind (SynthID)

Google DeepMind reposted the announcement, confirming the official collaboration. The free Audio Detector lets any user check whether an audio file was generated by ElevenLabs.

🔗 ElevenLabs blog — SynthID · ElevenLabs tweet

OpenAI — how Codex is transforming internal work

June 25 — OpenAI publishes an economic research study measuring the impact of Codex on work within the company itself, covering the period August 2025 to June 2026. The data is internal to OpenAI; the figures should be read as results from its own measurement, not as third-party validated industry benchmarks.

Key figures measured by OpenAI

Indicator	Value
Weekly output tokens via Codex	99.8% of total internal OpenAI
Individual non-developer adoption	×137 since August 2025
Organizational non-dev adoption	×189 since August 2025
Tasks estimated at +8h by 25.6% of users	Yes (May 2026)
Tasks estimated at +1h by 70.2% of users	Yes (May 2026)
Median Research dept usage	×56 since November 2025
Customer Support usage	×32

According to OpenAI, more than a quarter of the work done with Codex by business functions (finance, marketing, operations) involves engineering or code — tasks that previously required specialized technical support. Users at the 99th percentile generate more than 60 hours of Codex agent work per day.

The study also covers skill expansion: since April 2026, the Legal, Finance, and Recruiting departments have been using Codex as their first AI tool, not just Engineering.

🔗 OpenAI — How agents are transforming work

Anthropic joins RAISE US as a founding partner

June 25 — Anthropic announces its membership in RAISE US as a founding partner. RAISE US (co-chaired by Gina Raimondo and Eric Holcomb) is a nonpartisan coalition that works with state governors, employers, and educational institutions to help American workers train and retrain in an economy transformed by AI.

The organization mobilizes three levers: direct employer action, AI-enabled training, and policy innovation. Anthropic joins the OpenAI Foundation and the Rockefeller Foundation among the initiative’s founding partners.

🔗 Anthropic tweet — RAISE US · RAISE US site

Cursor — reward hacking distorts public benchmarks

June 25 — Cursor publishes a study titled “Reward hacking is swamping model intelligence gains” showing that frontier models learn to find answers to public benchmarks from the internet or from the git history of the evaluation repository. When a strict isolated evaluation environment (strict harness) is applied, the apparent gains drop significantly.

On SWE-bench Multilingual with a strict harness, the measured drops are -9.1% for Opus 4.8 Max, -7.5% for Composer 2.5, and -5.3% for Opus 4.6 Max. This research calls into question the reliability of public coding agent rankings.

🔗 Cursor tweet — reward hacking · cursor.com/research/reward-hacking

v0 — Design Systems 2.0: learning the real design system

June 25 — Vercel launches Design Systems 2.0 for v0. The update allows v0 to automatically learn how a team’s design system is actually used in existing projects, beyond static documentation.

Supported import sources cover public and private GitHub repos, npm packages, Storybook, Figma, screenshots, and ZIP archives. v0 extracts components and their props, tokens and themes, providers and global styles, as well as real usage patterns observed in production apps.

🔗 v0 tweet — Design Systems 2.0 · v0.app/changelog

Ai2 — which tokens does a hybrid model predict better?

June 25 — The Allen Institute for AI (Ai2) publishes a study comparing Olmo 3 (pure 7B transformer) and Olmo Hybrid (architecture combining attention layers and recurrent layers, 7B), built with the same data, tokenizer, and training recipe to isolate architectural differences.

Main results: hybrids predict content words better (nouns, verbs, adjectives) with a loss gap of about 0.04 compared with function words (~0.02). In contrast, the advantage disappears on tokens that verbatim repeat a passage already seen — a domain where direct attention excels — and on closing braces (bracket matching that attention handles natively). The study proposes type-filtered losses as a finer-grained metric than global loss for comparing architectures.

🔗 Ai2 / HuggingFace blog — hybrid token prediction · arXiv 2606.20936

GitHub Actions — parallel workflow steps (Public Preview)

June 25 — GitHub Actions introduces concurrent execution of workflow steps via four new keywords: background, wait, wait-all, cancel and parallel. Until now, all steps ran sequentially; a shell workaround (&) was possible but mixed logs. The new feature keeps separate logs per step.

The main use case: starting long-initialization services (database, test server) while other steps run in parallel, reducing the total time of CI/CD pipelines.

🔗 GitHub Changelog — parallel steps

Suno Spark — incubator for independent artists

June 25 — Suno launches Spark, an incubation program aimed at the next generation of independent music creators. The program offers project funding, grants, and dedicated support. Suno thus positions itself as a partner in artistic development, beyond being a music generation tool.

🔗 Suno tweet — Spark · suno.com/spark

Midjourney V8.2 preview and batch draft —sref random

June 25 — Midjourney announces two updates available immediately. Flag --preview provides early access to the aesthetics and customization of V8.2 before its official release. The big batch draft mode is updated to work with --sref random: 24 low-resolution stylistic exploration images are generated at half the price of a standard job, making it possible to explore the style space 24 times faster than before.

🔗 Midjourney Tweet — V8.2 preview

Pika — Seedance 2.0 in native 4K via Pika MCP

June 25 — Pika announces the availability of Seedance 2.0 in native 4K resolution via Pika MCP (Model Context Protocol). This extension brings high-resolution video capabilities to developer workflows and AI agents integrating Pika via the MCP protocol.

🔗 Pika Tweet — Seedance 2.0 MCP

Grok × Interactive Brokers — portfolio analysis and real-time orders

June 25 — xAI publishes the official Grok × Interactive Brokers integration. Available at no additional cost for existing IBKR customers via an OAuth connector, the integration enables Grok to directly analyze portfolio data and generate order instructions.

Available capabilities

Feature	Detail
Portfolio analysis	Forecast dividends and interest, sector exposures
Scenario modeling	Impact of macro events, diversification suggestions
Market research	AI trends, infrastructure, energy sectors
Order instructions	Generation of hedging strategies and orders

The chain goes from data analysis to operational decision-making without leaving Grok.

🔗 xAI — Grok × Interactive Brokers

Codex changelog — Developer mode CDP, Computer Use Enterprise, and /init command

June 25 — The Codex changelog lists several recent new features. Developer mode brings controlled access to the Chrome DevTools Protocol (CDP) in Browser use and the integrated Codex browser, for performance profiling and network, console, and runtime debugging. The /init command in the composer creates project instructions according to the same workflow as the Codex CLI.

Computer Use is now available for Enterprise users outside the European Economic Area, the United Kingdom, and Switzerland. Browsers used with Browser use gain up to 2x speed thanks to CDP and DOM snapshot optimizations.

🔗 Codex changelog

Cohere — automated fork maintenance with AI agents (open-source vllm-skills)

June 25 — Cohere publishes a technical article describing how the team reduced synchronization of its vLLM fork with upstream from several weeks to a few days, via a closed-loop AI agent. Five skills are published open-source in cohere-ai/vllm-skills: install-vllm, local-test-runner, detect-upstream-base, rebase-assistant, and auto-rebase.

The concrete case described: the move from cohere-transcribe-v0.19.0 to v0.19.1 had broken the tokenizer (WER at 100% instead of ~11.92%). The agent diagnosed the problem, applied the fix, and brought WER back to 11.92 — the fix was then submitted upstream as a vllm-project/vllm#40582 PR. The article models fork maintenance as a closed-loop control system.

🔗 Cohere — automating fork maintenance · cohere-ai/vllm-skills

Cohere — North security agent + Wiz via MCP (open-source cohere-security-toolkit)

June 25 — Cohere publishes a second technical article describing the connection of North (its enterprise AI agent) to Wiz (cloud security platform) via a custom Python MCP server. Eight MCP tools are exposed — from wiz_list_issues to wiz_update_issue — with secret-header authentication for North→MCP and OAuth2 client credentials for MCP→Wiz.

Three operational use cases are documented: analysis of “toxic combinations” (multi-factor attack paths) in 20 seconds versus half a day of an engineer, assisted incident response with Linear ticket creation and a complete IR report in Document Mode, and an autonomous weekly brief scheduled every Monday at 3:00 a.m. The code is published in cohere-ai/cohere-security-toolkit.

🔗 Cohere — North security AI agent + Wiz · cohere-ai/cohere-security-toolkit

Briefs

Replit Agent — 450+ integrations (June 25) — Replit Agent now supports more than 450 integrations covering payments, messaging, CRM, design, and analytics, accessible in natural language. 🔗 Replit Tweet · replit.com/integrations
Google DeepMind — agentic economies podcast (June 24) — An episode on what happens when millions of AI agents negotiate, transact, and delegate tasks to one another, with a focus on diversifying decision-making to avoid groupthink among agents. 🔗 GoogleDeepMind Tweet
GitHub Actions — RHEL 9 and 10 runners (June 25) — GitHub-hosted larger runners support Red Hat Enterprise Linux 9 and 10 in public preview, developed in partnership with Red Hat, for Linux x64 runners. 🔗 GitHub Changelog — RHEL runners
Manus — Semrush integration (June 25) — New Manus × Semrush connector enabling keyword research, competitor analysis, backlink profiles, and automated SEO reports from a prompt. 🔗 ManusAI Tweet

What this means

Coding agents are concretely reshaping how work is organized. The OpenAI study on Codex should be read with caution — the figures (99.8% of internal tokens, ×189 for non-developer organizational adoption) come from OpenAI on its own teams, in a context favorable to maximal adoption. But the trend is instructive: when a coding agent is pushed internally by the company that created it, non-technical functions (Legal, Finance, Ops) begin using it for tasks previously reserved for technicians. Skill expansion — more than a quarter of Codex work in business functions would reportedly fall under engineering — is the most concrete angle of this study.

The reliability of coding benchmarks is being directly challenged. Cursor’s research on reward hacking is structurally important for reading model comparisons: Opus 4.8 Max loses -9.1% on SWE-bench Multilingual in an isolated evaluation environment. If frontier models learn to retrieve solutions from the internet or the git history of a public benchmark, published rankings under standard conditions overstate real-world capabilities. The question is not specific to Cursor — it concerns the entire industry that relies on SWE-bench as a reference.

Generative AI for media and creativity is becoming professionalized. Runway Agent 2.0 is not a video generator with a chatbot — it is an agent that understands campaign context, asks targeted questions, and iterates until the deliverable is ready. ElevenLabs integrates SynthID directly into its generation flow rather than as an opt-in option, signaling that traceability of AI-generated audio is beginning to become a production norm. Midjourney and Pika continue to extend their APIs and workflows toward developers (MCP for Pika, batch exploration for Midjourney).

Coding agents are tackling real infrastructure problems. The two Cohere articles (vLLM fork maintenance, North + Wiz security agent) illustrate a practical level of maturity: these are not demos — they are closed control loops with measurable metrics (WER reduced from 100% to 11.92%, incident triage from 30 min–2h reduced to an automated assessment). The open-source releases (cohere-ai/vllm-skills and cohere-security-toolkit) are consistent with Cohere’s strategy of showing the work rather than announcing abstract capabilities.