Pulse Beat — Week of Oct 26, 2026
Your weekly summary of top AI insights and roundups.
Was this newsletter forwarded to you? to get it in your inbox.
The week in one line
The week in one line: Anthropic filed to go public, Google took the model crown with Gemini 4 Carbon — and then lost $500B in a day as investors realized AI search eats its own ad revenue. Everything else this week orbits those three facts: who builds the best models, what they cost, and who pays. Five-minute read, drawn from five weekly sources, Oct 26 – Nov 1, 2026.
The week’s top stories
| Date | Story | Why it matters |
|---|---|---|
| Oct 30 | Anthropic drops the 5-hour limit for its $200 "Ultra" plan; $100 tier renamed "Max" | Heavy Claude Code users get longer sessions; weekly caps unchanged |
| Oct 30 | Google stock falls 15% ($500B) after saying AI Mode monetizes worse than classic search | First hard evidence that AI answers cannibalize search ads |
| Oct 30 | Malicious Claude Code copy found in Cursor's extension marketplace, up to 5,000 users hit | Agent tooling is now a supply-chain target |
| Oct 28 | Gemini 4 Carbon launches, #1 on the Artificial Analysis index (59) | Google leads independent rankings, but access is limited to cyber defenders first |
| Oct 28 | ChatGPT "dots" can text you; Pro users get $100–$500/month API credits | Always-on agents move to your phone |
| Oct 26 | Anthropic files its S-1 to list on Nasdaq as "ANT" | Reportedly targeting ~$2T, the defining AI IPO |
| Oct 26 | GPT-6.1 Terra released: fast, cheaper computer-use model | Mid-tier price, mixed reviews on reliability |
Models: Google leads the scoreboard, Claude still leads coding
Gemini 4 Carbon arrived two weeks after Argon and took #1 on the rebuilt Artificial Analysis index, which now runs 60% on private tests. On coding agents, Claude Code with Opus 5.5 still scores higher.
| Model | AA Index v5 | Best for | Watch out for |
|---|---|---|---|
| Gemini 4 Carbon (Google) | 59 | Long tasks, cybersecurity, finance and legal research | Limited access for now; hallucination rate up from 4% to 11% |
| Claude Opus 5.5 (Anthropic) | 58 | Coding agents (62 vs. Carbon's 59 on the coding index) | Most expensive per task (~$10 vs. ~$7.76 for Carbon) |
| GPT-6 Astra (OpenAI) | 57 | Cheapest frontier option (~$6.25 per task) | — |
| GPT-6.1 Sol (OpenAI) | 55 | Value: ~1/8 the cost per task of Opus | — |
| GPT-6.1 Terra (OpenAI) | 51 | Fast, cheap computer use | Users report it claims tasks are done when they aren't |
Also this week: Mistral released Large 4, a 1-trillion-parameter open-weight model, and Alibaba says Qwen4 is coming next week.
So what: there's no single best model anymore. Match the model to the job — Opus for code, Carbon for research once it opens up, Sol when cost matters.
Agents and work: coworkers, factories — and new risks
Always-on agents are becoming teammates. OpenAI's Dots now text your phone, and Slack-based agents are answering "where's that doc?" questions in under a minute, as one first-week diary shows. They also get things wrong with total confidence.
In coding, "software factories" are now normal: plan with one model, have another review, let parallel agents build and test. One engineer's reflection on the cost — losing the trial and error that builds skill — suggests sorting tasks into automate, supervise, or do by hand.
The risks grew too. A fake Claude Code extension slipped into Cursor's marketplace, and Anthropic research shows thousands of agents can craft harmless-looking training data that teaches bad behavior. Heavy Claude Code users also reported quota burning 2–8x faster than normal.
So what: treat every agent tool and extension like software you're installing on your laptop — verify the source, sandbox it, and watch your usage meters.
Research and policy: AI is speeding up AI
The big research theme is AI accelerating its own development. Import AI 478 highlights a single researcher at startup Mirendil who pretrained a decent coding model alone, using AI tools. Google's CEO told investors he expects continued acceleration over the next 12 months as AI speeds up model training.
Math is the new proving ground. OpenAI and Anthropic both claimed AI-produced results against a major complexity-theory conjecture this month; experts are still checking them, so treat them as claims, not settled results.
Security is the policy flashpoint. Zvi Mowshowitz's weekly roundup covers reports that red-teamers using AI agents broke into Anthropic's internal Slack in under 12 hours through third-party weaknesses. Anthropic closed the gaps and briefed federal officials.
So what: if AI can find holes in an AI lab, it can find them in your stack. Budget for AI-assisted security review now.
Business: the AI trade gets its reality check
Anthropic's S-1 shows the fastest growth in software history, and the biggest bet. Every analyst this week landed on the same tension: revenue that can walk away in weeks, against compute bills locked in for a decade.
| Metric | Figure | Source |
|---|---|---|
| Revenue, H1 2026 | $16.3B (vs. $1.2B a year earlier) | S-1 |
| Net revenue retention | ~900% | S-1 analysis |
| Revenue customers have committed | ~12 weeks | Exponential View |
| Compute Anthropic has committed | 12+ years, ~$587B | Exponential View |
| Price at $2T valuation | ~30x annualized Q3 sales | Exponential View |
The wider market wobbled. Google's ~15% drop came as it admitted AI Mode earns less than classic search, and OpenAI reportedly raised at $1.25T rather than the $1.4T it sought. Exponential View's Sunday edition adds a longer view: AI agents could let firms grow far larger than before, entering many markets at once — as Anthropic is doing in coding, bio, legal and finance.
So what: AI demand is real, but so is the risk of overbuilding. Expect volatile AI stocks around the Anthropic listing.
Three things to do this week
- ☐ Audit the AI extensions and agent tools your team has installed; remove anything not from the official publisher.
- ☐ Pick one workflow and test it on two models (e.g. Opus 5.5 vs. GPT-6.1 Sol) — compare quality and cost per task.
- ☐ Check your Claude Code or API usage dashboard for unexplained spikes.
Coming up: Qwen4 launch (expected next week) · Thesis: 2027 livestream, Nov 5 · AI Engineer Code Summit, San Francisco, Nov 10–12 · Anthropic IPO pricing.
Sources
- AI #192: Anthropic Is Going Public — Don't Worry About the Vase, Oct 29
- Import AI 478 — Import AI, Oct 26
- 39 charts on Anthropic #608 — Exponential View, Nov 1
- Anthropic: From a Lab to a Real Business — summary of Exponential View analysis
- Gemini 4 Carbon tops leaderboards — AINews (Latent Space), Oct 29
- Google admits AI Mode hurts Search revenue — AINews (Latent Space), Oct 31
- GPT-6.1 Terra analysis — Artificial Analysis
- Is Anthropic Worth $2 Trillion? — every.to, Oct 26
- My First Week With an AI Coworker — every.to, Oct 29
- What Karl Marx Would Say About My Coding Agents — every.to, Oct 30