Industry

Anthropic’s new R&D Automation Index puts a number on how much of its own AI research Claude actually leads: 26% as of August 2026, up from under 1% in February, with over 90% of measured work at least AI-collaborative and roughly 30,000 agents running concurrently. The most important number in the report is the one Anthropic didn’t round up: zero measured tasks where Claude operates fully autonomously.

Cognition engineer Eric Lu, working with multiple Devin agents, factored RSA-260 — a 260-digit number unsolved since the 1991 RSA Factoring Challenge — using a GPU-accelerated rewrite of the CADO-NFS sieving toolkit that Devin designed and operated with only periodic human steering. We independently re-multiplied and primality-tested the published factors ourselves: the claim checks out. RSA-2048 is unaffected.

Elon Musk promised Grok 4.7 in about ten days on September 2. Fifteen days past that, on September 14, he self-graded the still-unshipped model as ‘roughly on par with Opus 5.0, not 5.1’ — a model Anthropic actually shipped July 24. As of September 17, xAI’s own docs still top out at Grok 4.6.

Z.ai (Zhipu AI) raised roughly $5B on Sept 13 via a ~$2B Hong Kong share placement and a ~$3B zero-coupon convertible bond, its second raise in two months and third since its January 2026 IPO — pushing total public-market capital raised to roughly $12B, funding compute and next-generation models while Anthropic’s own S-1 remains a confidential draft.

Manifold Security’s Sept 1 GitSpawn disclosure found that Claude Code’s git status subprocess — and separately, its /ultrareview command — execute attacker-controlled git-config settings before any trust prompt appears. Cursor, OpenAI Codex, and Goose all shipped fixes; the /ultrareview path in Claude Code is still unpatched as of v2.1.270 (Sept 12).

Cognition shipped SWE-2 on September 10, Devin’s new engine, post-trained on Moonshot’s 2.8T-parameter Kimi K3 base with three reasoning-effort tiers. Cognition’s own numbers claim near-Fable-5.1 coding quality at a fraction of the cost — but the model underneath is still someone else’s, four months after Cognition’s $48B Series E.

Anthropic’s September 2026 threat intelligence report accuses Moonshot AI of routing almost 300,000 user requests to Claude through 5,380 fraudulent accounts, saving Claude’s reasoning transcripts as training data, and passing the outputs off as its own Kimi model — with more than 23 million such exchanges between May and July. DeepSeek and Xiaomi face similar accusations; China’s Commerce Ministry calls the claims baseless.

OpenAI shipped the Agents API in public beta on September 10, exposing the Codex harness — session management, context compaction, sandboxing, MCP tool wiring — behind a single API. It’s the same managed-agent-loop pitch Anthropic made with Claude Managed Agents in April, arriving five months late and with a rougher enterprise story: US-only data residency and no zero-data-retention option during the beta.

Cognition closed its Series E on September 8 at a $48B valuation on over $2B raised — both figures higher than the $47B/$1B talks Bloomberg reported six days earlier. Annualized revenue hit roughly $900M, up from $492M in May, with a16z and Accel leading a round more than 20 investors deep.

Bloomberg reported September 2 that Cognition, maker of the Devin coding agent, is in talks to raise roughly $1B at a $47B valuation — up from $26B just three months earlier — with annualized revenue above $900M, up from $492M in May. The round hasn’t closed and terms could still change, but the trajectory says more about where investor conviction is heading than any single benchmark score.