This Week in One Line

OpenAI shipped GPT‑5.5 for do‑the‑work tasks, DeepSeek previewed a 1M‑context model with cut‑rate tokens, Google committed $10B (up to $40B with compute) to Anthropic, and Gmail added AI Overviews — everyday tools are getting faster, cheaper, and more agent‑like.

Week in Numbers

• 82.7% — GPT‑5.5’s score on Terminal‑Bench 2.0 for complex command‑line workflows. ¹
• $0.14/1M tokens — DeepSeek V4 Flash input price; output is $0.28/1M. ²
• $10B (up to $40B) — Google’s cash commitment to Anthropic, with additional compute access tied in. ³
• 5 GW — Compute capacity Anthropic secured from AWS alongside a $100B, 10‑year spend plan. ⁴
• ~40% faster — OpenAI reports agent loops run about 40% quicker using WebSockets in its Responses API (application programming interface). ⁵
• $510M — Cerebras’s 2025 revenue disclosed in its IPO (initial public offering) filing. ⁶
• 2–5× — Siemens says its new engineering agent completes workflows two to five times faster than manual work. ⁷

Top Stories

OpenAI launches GPT‑5.5 for agentic computer use

OpenAI released GPT‑5.5 in ChatGPT for Plus, Pro, Business, and Enterprise, positioning it to plan multi‑step tasks, operate software and tools, and finish work with less micromanagement — while matching GPT‑5.4’s per‑token latency and posting scores like 82.7% on Terminal‑Bench 2.0, 58.6% on SWE‑Bench Pro, 84.9% on GDPval, and 78.7% on OSWorld‑Verified. ¹

OpenAI frames this as a step toward more “agentic and intuitive computing,” with testing under internal/external safety frameworks and feedback from nearly 200 early‑access partners; rollout spans ChatGPT tiers now, with API access after additional safeguards. ⁸

DeepSeek V4 preview: 1M‑token context and lower token prices

DeepSeek previewed two mixture‑of‑experts (MoE) models — V4 Flash and V4 Pro — featuring 1,000,000‑token context windows; V4 Pro totals 1.6T parameters (49B active per request) and Flash 284B (13B active), with reporting noting competitive coding performance but text‑only I/O today. ²

Pricing undercuts peers: V4 Flash lists $0.14/1M input tokens and $0.28/1M output; V4 Pro lists $0.145/1M input and $3.48/1M output, with separate coverage pointing to remaining gaps versus top closed systems on some tasks. ⁹

Google commits up to $40B (cash + compute) to Anthropic

Axios reports Google committed $10B in cash, with the potential for an additional $30B, and bundled access to 5 gigawatts (GW) of compute on Google infrastructure at a $350B valuation for Anthropic — a sign that controlling compute has become as important as capital. ³

In parallel, Anthropic’s AWS agreement secures up to 5 GW of capacity with a $100B, 10‑year spend, illustrating a multi‑cloud strategy to lock in power and chips at scale. ⁴

ChatGPT Images 2.0 adds web‑grounded, consistent series generation

OpenAI introduced Images 2.0 in ChatGPT; paired with a Thinking model, it can consult the web, generate up to eight images per prompt with consistent characters/styles, offer more aspect ratios and up to 2K resolution, and improve non‑Latin text rendering. ¹⁰

Access details matter for teams: paid tiers can choose GPT‑5.3 Instant or GPT‑5.4 Thinking; manual Thinking usage is capped at up to 3,000 messages/week on Plus/Business, with larger context windows and a thinking‑time toggle to balance speed and depth. ¹¹

OpenAI lines up “DeployCo” and speeds up agent loops

Private Equity Wire reports OpenAI is negotiating up to $1.5B for a private‑equity–backed venture (“DeployCo”) to embed its tools across buyout‑firm portfolios, targeting a ~$10B valuation with majority OpenAI ownership and a forward‑deployed engineer model. ¹²

Separately, OpenAI cut agent‑workflow latency by ~40% using WebSockets in the Responses API (application programming interface), citing end‑user gains from ~65 to near 1,000 tokens/second and bursts to ~4,000 TPS (tokens per second). ⁵

OpenAI also highlighted Codex adoption and a systems‑integrator network (Accenture, Capgemini, CGI, Cognizant, Infosys, PwC, TCS) to move pilots into production for thousands of organizations. ¹³

Adobe launches CX Enterprise: a suite of marketing agents

Adobe unveiled CX Enterprise, a set of AI agents to automate and personalize digital marketing; Reuters notes shares rose 2.2% intraday on launch. ¹⁴

The “Coworker” agent can coordinate other agents and assemble relevant data, with partnerships across Amazon, Microsoft, Anthropic, OpenAI, and Nvidia — pitched as a governed enterprise path for agentic AI. ¹⁵

Siemens ships an AI engineering agent for PLC coding

Siemens released the Eigen Engineering Agent to automate PLC (programmable logic controller) coding, HMI (human–machine interface) visualization, and device configuration in TIA Portal; pilots span 100+ companies in 19 countries, with claims of 2–5× speed, up to 80% higher solution quality, and 50% greater engineering efficiency. ⁷

At Hannover Messe, Rockwell demonstrated an AI‑orchestrated workflow linking digital‑twin simulation, AI‑assisted code generation, and validation — reinforcing a shift toward AI‑native engineering. ¹⁶

Cerebras files for IPO on wafer‑scale AI chips

Cerebras filed to go public, disclosing 2025 revenue of $510M, GAAP net income of $237.8M, and a non‑GAAP net loss of $75.7M, and positioning its wafer‑scale hardware as an alternative to GPU clusters for training and inference. ⁶

AI Overviews arrive in Gmail at work

Google is rolling out AI Overviews to Gmail for work accounts: ask natural‑language questions and get summaries pulled from multiple messages without opening each thread; it’s on by default where admins enable Gemini for Workspace and related smart features. ¹⁷

Cohere to merge with Aleph Alpha, targeting “sovereign AI”

Cohere plans to combine with Germany’s Aleph Alpha, backed by Schwarz Group’s €500M (~$600M) commitment and a reported ~$20B combined valuation, to offer businesses and governments more control over data and deployment. ¹⁸

Reuters adds that Cohere shareholders would own about 90% and Aleph Alpha’s about 10%, with a focus on regulated sectors across Europe and beyond. ¹⁹

Trend Analysis

Agent‑first workflows moved into everyday tools. GPT‑5.5 aims to plan and execute multi‑step work, Gmail’s AI Overviews condense sprawling threads into answers, Adobe’s CX Enterprise coordinates marketing tasks, and OpenAI’s workspace agents (Business/Enterprise/Edu/Teachers) hint at embedded assistants that keep work moving. Together these point toward assistants that not only draft content but also operate software and tie into governed stacks. ^{1 17}

Compute strategy hardened. Google’s $10B (up to $40B) package for Anthropic and Anthropic’s separate $100B/5 GW AWS commitment illustrate a shift from sporadic capacity to multi‑cloud, multi‑gigawatt procurement; Cerebras’s IPO filing and data‑platform raises underscore that chips and storage are both chokepoints teams now budget explicitly. For non‑specialists, this signals steadier quotas, clearer SLAs, and faster responses as vendors scale underlying capacity. ^{3 6}

Open‑weight efficiency pushed costs down. DeepSeek V4’s 1M‑token context at cut‑rate prices targets long documents and codebases without elaborate chunking, while Google’s Gemma 4 and other open families emphasize local‑first agent workflows and long contexts across common runtimes. The throughline: more work can move to cheaper stacks or even to your own hardware, provided governance and evaluation keep pace. ^{2 20}

Watch Points

• “DeployCo” — If OpenAI’s PE‑backed venture closes, expect more embedded rollouts with forward‑deployed engineers across portfolio companies. ¹²
• “Inference chips” — Google’s push into dedicated inference silicon signals competition on latency and cost; watch for availability and partner wins. ²¹
• “Florida AG probe” — A criminal investigation into OpenAI over ChatGPT’s alleged role in a shooting raises accountability and policy questions around chatbot use. ²²

Open Source Spotlight

• browser‑harness — A minimal, self‑healing browser agent harness atop Chrome DevTools Protocol; agents can even edit the harness mid‑task. Useful for teams who prioritize reliability over heavy frameworks. browser-use/browser-harness
• screenpipe — “AI memory for your screen” with background agents that summarize meetings and surface context as you work; active desktop iteration. Good for personal productivity pilots. screenpipe/screenpipe
• Hermes Agent — Always‑on, server‑side assistant that learns reusable “skills,” with long‑term memory and multi‑platform chat gateways; MIT‑licensed. NousResearch/hermes-agent
• nanobot — Ultra‑light personal agent, easy to install from PyPI, now with Windows and Python 3.14 support; for users who want simple, local helpers. HKUDS/nanobot
• Harvard CS249r + TinyTorch — Hands‑on machine‑learning systems textbook and labs (multi‑language docs incl. Korean); latest TinyTorch adds a richer Tensor API and security fixes. harvard-edge/cs249r_book

What Can I Try?

1. Put GPT‑5.5 on a real multi‑step task: “find sources, draft slides, and build a spreadsheet,” then review its plan and checks. ¹
2. Draft a consistent 6–8 image set in one prompt using ChatGPT Images 2.0; test wide and tall aspect ratios. ¹⁰
3. Turn on Gmail AI Overviews at work (ask your admin to enable Gemini for Workspace and smart features), then ask, “What changed in last week’s SOW?” ¹⁷
4. Try an always‑on desktop agent: install screenpipe and test the updated “summarize meetings” flow. ²³