## Sources

1. [OpenAI Ships Codex Mobile App for iOS and Android](https://awesomeagents.ai/news/openai-codex-mobile-app-ios-android/)
2. [Raindrop Workshop Gives AI Agents a Local Debugger](https://awesomeagents.ai/news/raindrop-workshop-open-source-agent-debugger/)
3. [Thinking Machines Builds AI That Listens While Talking](https://awesomeagents.ai/news/thinking-machines-interaction-models/)
4. [Microsoft Drops Claude Code for GitHub Copilot Desktop](https://awesomeagents.ai/news/microsoft-drops-claude-code-copilot-desktop/)
5. [GAIA Benchmark Leaderboard: Best AI Agents May 2026](https://awesomeagents.ai/leaderboards/gaia-benchmark-leaderboard/)
6. [Claude Haiku 4.5](https://awesomeagents.ai/models/claude-haiku-4-5/)
7. [GPT-4.1](https://awesomeagents.ai/models/gpt-4-1/)
8. [Claude 3.7 Sonnet](https://awesomeagents.ai/models/claude-3-7-sonnet/)
9. [Cisco Bets $9B on AI Networking, Cuts 4,000 Jobs](https://awesomeagents.ai/news/cisco-9b-ai-networking-4000-jobs/)

---

### **Cisco Bets $9B on AI Networking, Cuts 4,000 Jobs | Sophie Zhang**

*   **Main Arguments**: Cisco is aggressively pivoting its business model to focus on **AI infrastructure**, specifically high-speed switching and silicon, to meet the massive demand from hyperscale cloud providers [1, 2]. Despite achieving **record quarterly revenue of $15.8 billion**, the company is simultaneously undergoing its fourth major restructuring since early 2024 to reallocate resources toward AI-centric roles [3, 4].
*   **Key Takeaways**:
    *   Cisco raised its **FY2026 AI infrastructure order target from $5 billion to $9 billion**, with over $5.3 billion already booked [1, 3].
    *   The company announced **4,000 layoffs**, bringing its total job cuts since early 2024 to approximately 9,750 roles [1, 4].
    *   Revenue grew by **12% year-over-year**, driven by a 40%+ surge in data center switching orders and 50%+ growth in networking product orders [1, 2].
*   **Important Details**:
    *   The primary drivers of this demand are **hyperscalers** like Microsoft, Google, Amazon, and Meta, who require low-latency interconnects for AI training clusters [1, 5].
    *   Cisco is emphasizing a **vertical play** by building its own Application-Specific Integrated Circuits (ASICs) and high-speed optical transceivers to maintain better margins than commodity hardware [6, 7].
    *   There is a significant **customer concentration risk**, as Cisco’s growth is now heavily dependent on the capital expenditure cycles of just four major customers [8, 9].

### **Claude 3.7 Sonnet | James Kowalski**

*   **Main Arguments**: Released in February 2025, Claude 3.7 Sonnet was Anthropic's **first hybrid reasoning model**, designed to allow users to choose between near-instant responses and deep, step-by-step reasoning from a single set of weights [10]. It established a new frontier for **agentic coding** and complex problem-solving before being succeeded by the Claude 4.x series [11, 12].
*   **Key Takeaways**:
    *   It introduced a **togglable "extended thinking" mode** with a configurable reasoning budget of up to 128K tokens [10, 11].
    *   At launch, it set a **record on SWE-bench Verified at 70.3%** (with scaffolding), significantly outperforming OpenAI’s o1 [11, 13].
    *   The model featured a **200K input context window** and was priced at $3 per million input tokens and $15 per million output tokens [12].
*   **Important Details**:
    *   Unlike some competitors, Claude 3.7’s **chain-of-thought is fully visible** to the user rather than being hidden behind the API [10, 14].
    *   The model showed a **45% reduction in unnecessary refusals** compared to its predecessor, Claude 3.5 Sonnet, making it more reliable for autonomous pipelines [15, 16].
    *   The model was **deprecated in late 2025** as Anthropic transitioned to the Claude 4.x naming scheme and newer model identifiers [12, 17].

### **Claude Haiku 4.5 | James Kowalski**

*   **Main Arguments**: Claude Haiku 4.5 represents a strategic shift for Anthropic by bringing **advanced agentic features**—like extended thinking and computer use—to its most affordable and fastest model tier [18]. It aims to provide "near-frontier" intelligence at a fraction of the cost of flagship models [19, 20].
*   **Key Takeaways**:
    *   It achieved an impressive **73.3% on SWE-bench Verified**, nearly matching the performance of much more expensive models like Claude Sonnet 4.5 [18, 21].
    *   Haiku 4.5 is the **first in its line to support Computer Use**, outperforming the previous-generation Sonnet 4 on the OSWorld benchmark (50.7% vs 42.2%) [18, 22].
    *   Pricing is set at **$1 per million input tokens and $5 per million output tokens**, making it roughly one-third the cost of the Sonnet tier [19, 20].
*   **Important Details**:
    *   The model is optimized for speed, producing approximately **93 tokens per second**, which is double the pace of Sonnet-tier models [21, 23].
    *   It features **"context-awareness" training** specifically designed to reduce "agentic laziness" and keep the model from abandoning long, multi-step tasks [24, 25].
    *   Anthropic has made Haiku 4.5 the **default model for free Claude.ai users**, signaling a competitive move against OpenAI's free-tier offerings [20, 26].

### **GAIA Benchmark Leaderboard: Best AI Agents May 2026 | James Kowalski**

*   **Main Arguments**: The GAIA benchmark, which tests real-world multi-step tasks, reveals that the **scaffolding or framework** surrounding a model is often more important for success than the underlying model itself [27, 28]. For 2026, the data shows that well-built agent loops can boost performance by as much as **30 percentage points** compared to direct API calls [28, 29].
*   **Key Takeaways**:
    *   **Claude Sonnet 4.5 with the HAL Generalist framework** leads the leaderboard with a score of 74.55% [30, 31].
    *   Anthropic models dominate the top of the rankings, holding the **first six positions** on the standardized HAL leaderboard [31].
    *   The **"scaffold effect"** is the single largest performance lever; for example, Claude Sonnet 4.6 scores 45.5% raw but earlier versions reach 74%+ when properly orchestrated [29, 32].
*   **Important Details**:
    *   **Level 3 tasks**, which require 20+ steps and long-horizon state maintenance, remain the hardest for AI, with most models staying below 55% while humans score 92% [33, 34].
    *   **Cost efficiency** does not scale linearly with accuracy; for instance, o4-mini Low with HAL is considered a "Pareto-best" value at $73 per run compared to over $600 for some flagship configurations [35, 36].
    *   Direct API calls of 2026 frontier models, like **Claude Mythos Preview (52.3%)**, still underperform older 2025 models that use advanced scaffolding [37, 38].

### **GPT-4.1 | James Kowalski**

*   **Main Arguments**: GPT-4.1 was released by OpenAI as a **coding-optimized API model** designed to address developer frustrations with instruction-following and "diff-format" outputs in earlier versions [39, 40]. It prioritizes **high reliability and massive context** over the deep reasoning found in the "o" series [41, 42].
*   **Key Takeaways**:
    *   It features a **1 million token context window**, an 8x increase over GPT-4o [41, 43].
    *   The model saw a significant jump in coding performance, reaching **54.6% on SWE-bench Verified**, up from 33.2% in GPT-4o [42, 44].
    *   It is priced competitively at **$2 per million input tokens and $8 per million output tokens**, which was 26% cheaper than GPT-4o for median queries at launch [41, 45].
*   **Important Details**:
    *   A major technical highlight is the reduction in the **extraneous edit rate** from 9% to 2%, making it highly efficient for tools that update code via diffs [40, 46].
    *   Performance on the **needle-in-a-haystack test** reached 100% across all context lengths, though accuracy on complex tasks can degrade at the very end of the 1M window [47, 48].
    *   OpenAI launched the model in three sizes—**Full, Mini, and Nano**—with Nano being the fastest and cheapest at $0.10/$0.40 per million tokens [41, 45].

### **Microsoft Drops Claude Code for GitHub Copilot Desktop | Sophie Zhang**

*   **Main Arguments**: Microsoft is consolidating its internal engineering tools by canceling thousands of licenses for Anthropic's **Claude Code** in favor of its own **GitHub Copilot Desktop** [49, 50]. This move is driven by a desire for **direct strategic control** over internal workflows, security requirements, and repository integration [51, 52].
*   **Key Takeaways**:
    *   Engineers in the **Experiences + Devices division** (Windows, Office, Teams) must transition away from Claude Code by June 30, 2026 [50, 53].
    *   **GitHub Copilot Desktop** launched in technical preview on May 14 as the designated replacement, offering a standalone agentic coding environment [49, 54].
    *   While the Claude Code interface is being removed, **Claude models remain available** within Microsoft via Azure and the Copilot model selector [50, 55].
*   **Important Details**:
    *   GitHub Copilot Desktop introduces **parallel session management**, allowing engineers to work on multiple branches or issues simultaneously using distinct "git worktrees" [54, 56].
    *   The app features three operating modes: **Interactive, Plan, and Autopilot**, with Autopilot handling the full cycle of coding, testing, and CI conflict resolution [56, 57].
    *   Internal feedback suggests that **Claude Code was highly popular** due to its speed and context handling, leading to a "mixed" and frustrated response among some engineers regarding the forced switch [55, 58].

### **OpenAI Ships Codex Mobile App for iOS and Android | Sophie Zhang**

*   **Main Arguments**: OpenAI has released mobile apps for its **Codex coding agent**, positioning the smartphone as a **"remote control"** for desktop sessions [59, 60]. This allows engineers to monitor and steer long-running agent jobs while away from their primary workstation [60, 61].
*   **Key Takeaways**:
    *   The app uses **QR code pairing** to link a mobile device to a macOS host machine in under a minute [59, 62].
    *   It features **live terminal streaming**, allowing users to see real-time output, test results, and browser screenshots from active sessions [61, 63].
    *   The security model ensures that **sensitive data never leaves the host Mac**; the mobile app only has "read" access to output and "write" access to approvals [64].
*   **Important Details**:
    *   As of launch, the app requires a **macOS host machine**, with no specific date given for Windows support [60, 65].
    *   Users can **switch models mid-session** from their phone, choosing between frontier models like GPT-5.4 [63, 66].
    *   The mobile app **cannot function standalone**; if the host machine goes to sleep or the desktop app crashes, the session terminates [67].

### **Raindrop Workshop Gives AI Agents a Local Debugger | Sophie Zhang**

*   **Main Arguments**: **Workshop** is an open-source, MIT-licensed tool designed to provide the **observability** necessary for debugging complex AI agents that standard terminal logs can no longer handle [68, 69]. It aims to replace "printf-based" debugging with a real-time, visual trace of tokens, tool calls, and decision trees [70, 71].
*   **Key Takeaways**:
    *   The tool streams every token and span to a **local browser dashboard** at `localhost:5899` [69, 72].
    *   It includes an **MCP (Model Context Protocol) server**, enabling coding agents like Claude Code to read their own traces and **automatically fix failing evaluations** [69, 73].
    *   Workshop supports **13+ AI frameworks** (including LangChain and OpenAI Agents SDK) across multiple languages like Python, TypeScript, and Rust [69, 74].
*   **Important Details**:
    *   A standout feature is **Trace Replay**, which allows developers to reproduce production failures locally without incurring additional API costs [72].
    *   The system uses a **local SQLite database**, ensuring that trace data remains private and under the developer's control [71, 75].
    *   While powerful, the tool currently has gaps in **team sharing** and lacks clear documentation for tracing distributed, multi-process agent systems [75, 76].

### **Thinking Machines Builds AI That Listens While Talking | Elena Marchetti**

*   **Main Arguments**: Thinking Machines, the startup founded by Mira Murati, has introduced a **full-duplex conversation model** called TML-Interaction-Small [77, 78]. Unlike traditional voice assistants that process turns sequentially, this model **listens and generates speech concurrently**, aiming for a more human-like, non-robotic interaction [78, 79].
*   **Key Takeaways**:
    *   The model achieves a response latency of **0.40 seconds**, which is nearly 3x faster than GPT-4o Realtime-2 [78, 80].
    *   It uses a **276-billion-parameter Mixture-of-Experts (MoE)** architecture that processes audio, video, and text together in **200ms micro-turns** [78, 79].
    *   The model demonstrates **proactive capabilities**, such as noticing visual changes (CueSpeak) or following timing-based instructions (TimeSpeak), where competitors score near zero [81].
*   **Important Details**:
    *   The architecture is **encoder-free**, merging raw audio dMel representations and image patches directly into the transformer [82].
    *   The system splits work between an **Interaction Model** for live presence and a **Background Model** for asynchronous tasks like web searches or tool calls [83, 84].
    *   Despite the strong benchmarks, the model is currently a **research preview only**, with no public API, pricing, or independent validation of the company's internal "FD-bench" scores [85, 86].