The week Anthropic overtook OpenAI

Anthropic is now worth more than OpenAI on paper. Secondary shares priced the company at $1 trillion versus OpenAI's $880 billion, and Google added up to $40 billion on top at a $350 billion paper valuation. The numbers are getting hard to take seriously, but investors are clearly betting that Anthropic catches OpenAI on revenue too.

OpenAI is not standing still either. GPT-5.5 is a very good model that we have leaned on hard since it launched. It has been particularly impressive on creative writing and is now our daily driver inside Notion. People are in heated debates over whether Codex with GPT-5.5 has overtaken Claude Code with Opus 4.7 for agentic coding. Some have switched, while others stay with Claude Code because the total package still delivers. There is no doubt that both are incredibly capable now, at coding and at advanced non-coding agent work too. We are still Claude Code fans, but we have started comparing Codex on some projects to get a feel for the hype ourselves.

Last week we ran our second Claude Code Bootcamp and Claude Code Meetup back to back at Mesh Oslo. The bootcamp sold out this time, and 600 people showed up for the meetup with the queue around the block, making it one of the largest Claude Code meetups in the world.

Two more Claude Code bootcamps are coming in May and June.

Claude Code Bootcamp, 27 May

Our third Claude Code Bootcamp is happening on 27 May. As always, we are hosting it at Mesh Oslo. It is a full day, hands-on programme for anyone who wants to learn Claude Code in a single day.

This is for project managers, consultants, analysts, lawyers, researchers, marketers, founders, and anyone who wants to multiply their output. All you need is a laptop and the willingness to try something new. Newsletter subscribers get 20 percent off with the code CCGRITAI03.

AI Pulse

A bite-sized curation of this week's most important AI news.

• OpenAI releases GPT-5.5. Improved agentic reasoning and tool use. The per-token price doubles versus GPT-5.4, but token efficiency offsets most of it for an effective 20 percent cost increase. Better at coding and long-horizon tasks.
• The Pentagon excludes Anthropic from classified AI contracts. Deals went to OpenAI, Google, Microsoft, AWS, Nvidia, SpaceX, Oracle and Reflection AI. Anthropic refused to allow Claude for autonomous weapons or mass surveillance under the Pentagon's "all lawful purposes" terms.
• The White House blocks Anthropic's Mythos expansion. It opposes the plan to widen Mythos access from 50 to 120 firms, citing compute strain on government use. A separate national security memo in the works could eventually ease the supply-chain risk designation.
• Anthropic hits a $1 trillion valuation on secondary markets, surpassing OpenAI's $880 billion on Forge Global. Driven by scarce shares and growing demand from Claude Code adoption.
• Anthropic eyes a $900 billion valuation in a $50 billion primary round. It could close within two weeks, driven by investor demand and revenue approaching a $40 billion annual run rate.
• Google invests up to $40 billion in Anthropic at a $350 billion valuation. $10 billion in cash now, with up to $30 billion more tied to performance targets. Amazon also committed up to $25 billion at the same valuation, with $5 billion paid in now.
• Microsoft and OpenAI restructure their partnership. OpenAI's models, including Codex, land on AWS via Bedrock the day after Azure exclusivity ended. Microsoft keeps first-launch rights on Azure and revenue share through 2030, but cloud customers now have full choice.
• OpenAI ships major Codex updates. GPT-5.5 is now the recommended model for complex coding. The CLI adds an in-app browser for UI testing, persistent /goal workflows with pause and resume, a plugin marketplace, multi-environment runs and Bedrock support.
• xAI ships Grok 4.3. A higher Intelligence Index score than Grok 4.20, a one million token context window and roughly 40 percent lower input price. Strong on instruction following and agentic customer support.
• Claude agents negotiate real deals autonomously. In an internal marketplace experiment, Claude closed 186 deals worth over $4,000. Opus agents got better prices than Haiku agents, and users never noticed.
• Claude Security hits public beta for Enterprise. Powered by Opus 4.7, it scans codebases for vulnerabilities, generates patches and pushes findings to Slack or Jira. Microsoft Security, Palo Alto Networks, CrowdStrike, SentinelOne and Wiz are embedding Opus 4.7 into their own platforms. Team and Max access is coming soon.
• Claude Agent Memory launches for Managed Agents. Agents now remember information across sessions. Memory is stored as files, exportable and scoped with permissions.
• Cloudflare lets agents create accounts and deploy. Agents can now provision a Cloudflare account, register a domain and get an API token without any dashboard work. Humans still approve terms and permissions.
• Stripe ships the Agentic Commerce Suite at Sessions. Link wallet for agents, streaming payments and one-time-use card credentials so agents can pay without exposing real card details. 288 launches in total.
• Meta opens its ads platform to MCP agents. Run Facebook and Instagram ad campaigns through Claude, Cursor or any connected agent, with direct access to ad data via the new MCP server.
• Claude connects to Blender, Adobe and seven more creative tools. Nine new connectors let Claude control Photoshop, Premiere, Ableton, SketchUp and others directly. Anthropic also made a one-time donation to support the Blender project.
• Warp terminal goes open source. The full Rust codebase is on GitHub under AGPL v3, with over 40,000 stars within days. An agent-first contribution model where AI handles the coding. OpenAI is the founding sponsor.
• Google Gemini Enterprise Agent Platform launches, with new tools for building and managing AI agents securely in Google Cloud.
• Microsoft Agent 365 hits general availability. It extends Microsoft's enterprise identity, security and governance stack to every AI agent and the systems they touch. A direct play to own the agent control plane inside large organisations.
• Cursor's margins are negative 23 percent. $2.7 billion in annualised revenue, but power users cost more to serve. SpaceX has secured a $60 billion acquisition option, with Colossus compute access to cut Cursor's dependence on Anthropic and OpenAI.
• DeepSeek V4 is an incredibly good open-source model. Free to download and run yourself, it nearly matches Claude Opus 4.6 on coding tasks. A US government review puts it just eight months behind the top closed models.
• DeepSeek is reportedly in talks to raise its first outside funding round at a valuation above $20 billion. Tencent and Alibaba are said to be discussing investments.

Claude Code updates this week

Here are the highlights from the changelog since our last issue.

• /usage is now the single command for spend, plan limits and session stats. It folds in what used to live across /cost and /stats.
• Paste a PR URL into /resume and Claude finds the session that built that pull request. Works with GitHub, GitHub Enterprise, GitLab and Bitbucket. Useful when you forget which conversation produced what.
• Custom themes. Build your own colour theme interactively from /theme, or hand-edit the JSON file in ~/.claude/themes/.
• /resume is up to 67 percent faster on large sessions. Sessions over 40 MB that used to crawl now load quickly, and dead-fork entries are handled more efficiently.
• See what Claude is thinking. The thinking spinner now shows inline progress ("still thinking", "thinking more", "almost done thinking") instead of a silent wait.

Coffee reads

Two takes for the coffee break.

How to make Claude last longer. When you hit the limit and Claude stops mid-flow, the instinct is to upgrade. The better instinct is to use what you have more carefully.

• Use Sonnet for most things. Reserve Opus for the heavy lifts and file-heavy work.
• Start fresh conversations often. Long threads bloat the context with information you no longer need.
• Feed only the relevant material. Markdown beats PDF, and CSV beats spreadsheet.
• Plan before you produce. The "redo everything" prompt is the most expensive one you can write.
• Break work into separate threads instead of one mega-conversation.
• Use Claude inside Word, Excel and PowerPoint for targeted edits, not full rewrites.
• Check /usage or the settings panel to see where your tokens actually go.

Honestly, the Max subscription is the best money we spend nowadays. If you are tired of overthinking token usage, you have our full support in upgrading. Read the full take on LinkedIn (in Norwegian).

Why most AI ideas should be skills, not apps. A skill packages how the work should be done end to end. Not a prompt, not a template, not an app. When the task involves assessment, structure, analysis or text, you often do not need a new system. You need a skill.

A few months ago we taught a healthcare team to build skills. They shipped a working solution to a concrete problem in three hours. The same scope as a traditional IT project would have cost millions.

Start with the problem. If the work lives in judgement and language, build a skill first. The distinction between skill and app may be short-lived anyway. Live Artifacts, Canvas and MCP are already blurring the line. Read the full take on LinkedIn (in Norwegian).

Using ChatGPT? Try Codex

Claude Code is Anthropic's everything agent. It reads files, runs commands and ships apps while you watch. We use it every day and have built our skills and workflows around it.

OpenAI ships a similar product called Codex. If you pay for ChatGPT Pro or Plus, you already have access. The new Codex desktop app for Mac is the closest thing OpenAI has to Claude Code, with a clean interface, cloud-to-local sync and parallel agents.

Many people miss this. Like Claude Code, Codex can do much more than write code. People use it for research, analysis, product marketing, content drafting and decision support. Anything that benefits from an agent that can read your files and run on your data.

If you have ChatGPT Pro, give Codex a real workflow this week. The two products are converging fast, and the right tool depends on what you have already built around.

Join the GritAI community

2025 was about learning to work with ChatGPT, Claude and Gemini. 2026 is about something else entirely. Agents. Practical AI skills and workflows, new content every week, and a live hangout with the community to learn and share what you are building. Join the GritAI community.

Best, Aleksander and Simon