Anthropic manager axes coding tests 💻, GPUs need $6.50/hr to pay off 💵, AI benchmarks go private 🔒

Google has patched 62 Chrome vulnerabilities from Anthropic's reports as of Chrome 156, all rated critical or high, including 13 sandbox escapes. A one-click Chrome chain ran attacker code outside the browser's sandbox, and a Linux kernel bug let remote attackers run commands as root on servers using multipath TCP. Anthropic says exploit-building was once the domain of a few specialist teams, and urges defenders to scan their own code, including at pull-request time.

Over 24 days, Axiom staff directed swarms of dozens of agents to rewrite the Classification of Finite Simple Groups, symmetry's basic building blocks, in Lean, a language where a computer checks every step. The work exposed 34 noteworthy deviations from the published record, including a false equality in one lemma, and produced a new group theory result. Axiom is publishing the errata, and says results resting on the classification can now be traced to checked steps.

The hypothesis held that deciding whether a 3-SAT logic formula can be satisfied always takes exponential time as variables grow, and Claude's algorithm runs faster than that. Anthropic's paper says the model found the core algorithm on 16 October with no human input, while OpenAI's model claims a similar result, and the Lean 4 proof is re-checked by a second kernel. The paper stresses the result is purely asymptotic and says nothing about solving formulas in practice.

Claude Opus 5.5 leads the revamped index at 58, ahead of GPT-6 Astra at 57, while the best open models you can download today, DeepSeek V4.1 Flash and GLM-5.3, score 32. Cognition's new FrontierCode 2.0 benchmark, 15% of the score, grades whether a repository maintainer would merge an agent's change, and Opus leads it at only 46%. GPT-6.1 Sol scores within 4 points of Opus at about an eighth of its cost per task.

Users steer each world with keyboard controls and text prompts, while Gemini 3.8 Flash watches the frames, breaks goals into steps and checks that each one succeeded. DeepMind is pairing it with robotics research, such as simulating Apptronik's Apollo humanoid playing checkers, and testing it with filmmakers and game developers. It is a research preview for trusted testers, with a waitlist for Google AI Ultra subscribers, and DeepMind says details still drift in longer sessions.

Oracle's cloud will offer Google's TPU chips, the first time they run on another hyperscaler's cloud, alongside OpenAI's Jalapeño chip and Cerebras hardware. Mahesh Thiagarajan, Oracle's cloud infrastructure president, says chips remain scarce but denied a pivot away from Nvidia, calling its chips "the best, most fungible". Oracle shares are down more than 30% this year, partly on concerns about how it will finance its data centre build-out.

Tinfoil runs customer workloads in hardware-encrypted enclaves it says even it cannot access, and dropped OVMF, a widely used firmware, since it needed only a tiny set of its features. Everything Linux needs now sits in launch memory, measured by the hardware as one hash, so the firmware checks just two host inputs, CPU count and memory map. Tinfoil says LLMs make small, purpose-built code easier to write and test than red-teaming huge legacy codebases.

Claygent, Clay's research agent for sales and marketing teams, has handled over 5 billion queries, from email lookups to deep company research. Applied Compute post-trained Qwen 3.6 35B A3B to predict how well each candidate model would answer, then pick the best trade of score against cost, with per-category floors so no query type pays for the savings. In a live A/B test against sending everything to GPT-5.6 Sol, it held Clay's primary guardrail metrics.

Together AI repairs a node only after a failed health check, watching inside the customer's virtual machine for GPU errors and on the host for network faults, since a node merely dropping offline does not say why. Customers choose, down to individual NVIDIA Xid error codes, which repairs run unattended and how long to wait for jobs to finish. A circuit breaker trips when one cluster bursts with correlated repairs, which points to a cluster-wide cause.

Rieseberg, a hiring manager at Anthropic, says Claude is now better than candidates at the small coding challenges that once hinted at who would succeed on his teams. He says comparatively new engineers are flying past better coders with far longer tenure, and advises engineers without status to pick one radical idea tied to business goals and build a working prototype over a weekend. Claude Code's graphical interface started that way.

Raphaëlle d'Ornano notes the filing leaves blank Anthropic's total committed compute spend and what falls due within 12 months, while disclosing $587 billion of purchase commitments and unstarted leases at 30 June, 5.6 times December's. Broadcom's CEO held last-minute talks to keep wavering banks in the largest AI debt deal ever, which lets Anthropic lease Broadcom-built chips, the Wall Street Journal reported. She would judge Anthropic on whether cash, cash flow and drawable financing cover next year's compute bill.

Across 50,000 runs at today's build costs and GPU rental rates, the median gigawatt lost $12.1 billion, and breaking even needed about $6.50 an hour per GPU against roughly $4.50 today. Open-weight models roughly cancel out for the owner, because tenants pay the same hourly rate whichever model they run. If rates hold and a build costs $35 billion, it makes money in 86% of runs, but at $50-60 billion it still loses nearly every time.

Ioannou, who leads engineering at Bold.org, made 2,102 GitHub contributions by 27 October, more than any full year from 2021 to 2025, despite doing far less feature work. Cheap implementation removed the friction that once stopped him chasing every idea, so each new model sets off a burst of parallel agents until reviewing and context switching wear him down. Even between bursts he takes on more than he used to, and says he is tired.

Peking University's Harness-Zero fine-tuned Qwen3.5-9B on runs where GPT-5.6 Sol made the smallest fix to each step it disagreed with, then removed the specialised harness, the tools, rules and memory wrapped around the model. Task success rose 21 points, more than the full harness had added, though chemistry trailed it by eight. Voss, writing for Arize, says to log every harness correction and switch off components one at a time when a new model arrives.

After SonarQube flagged a file where Codex had packed 15 whole components onto single lines, the author had to inspect code they had never seen, and it was all correct. They argue few companies will pay juniors to learn by years of hand-coding or line-by-line review. Owners instead need fundamentals to spot problems like quadratic slowdowns, broad reading to know where to look, and willingness to dive deep when something breaks.

The Chinese and English parser covers text, tables, formulas, charts and seals, and ships under Apache 2.0, weights included. OpenBMB reports a top score of 93.68 on Real5-OmniDocBench and throughput about 1.49 times PaddleOCR-VL-1.6's. A companion 2B model, Jev, scores each region of a result from 0 to 1 to show what needs human review, and the parser runs on vLLM, SGLang, llama.cpp and Ollama.

The Apache 2.0 extension, pg_pinecone, lets each table pick one of three engines: pgvector's own index, a native engine that applies filters such as a tenant ID during search with no network calls, or a remote Pinecone serverless index while Postgres keeps every row. Every switch can be reversed. It can also sync a table to a Pinecone index for BM25 and hybrid search from SQL, and needs PostgreSQL 17 or 18 with superuser access.