AI Daily Digest November 7, 2026: Thinking Machines Unveils 594B Inkling 2 Small, METR Warns Anthropic Accelerated by AI

Welcome to the AI Daily Digest for November 7, 2026! Today’s briefing presents a dramatic and multi-dimensional snapshot of the artificial intelligence frontier: from the fierce race in colossal open-weight foundation models and the emergence of persistent autonomous agents, to an unexpected economic plateau in corporate software budgets. Leading today’s headlines is Thinking Machines - the stealth startup spearheaded by former OpenAI Chief Technology Officer Mira Murati - which officially launched Inkling 2 Small, a massive 594-billion parameter Mixture-of-Experts (MoE) system claiming the crown as the most potent open-weight model from the West. On the agentic horizon, Perplexity has unveiled “Teammates,” introducing always-on AI colleagues operating around the clock inside dedicated cloud virtual machines for $200 per month. Meanwhile, an exhaustive eleven-week audit by non-profit research institute METR revealed that Anthropic has effectively doubled its model development speed, with 80% of that acceleration driven directly by AI systems conducting AI research - potentially crossing internal safety warning thresholds set back in July. Rounding out our digest, Ricursive Intelligence launched Ripple-65B Preview under a permissive Apache 2.0 license to lead terminal coding benchmarks, while enterprise card platform Ramp released spend analytics indicating corporate AI outlays are leveling off amid intense vendor price cuts and rigorous prompt optimization. Let us dive into the five biggest stories shaping tech today!

🧠 Thinking Machines unveils Inkling 2 Small: 594B MoE model claims top spot among western open-weight AI

Thinking Machines, the high-profile AI startup founded by former OpenAI Chief Technology Officer Mira Murati, has formally published Inkling 2 Small - a 594-billion parameter Mixture-of-Experts (MoE) architecture designed to compete directly against Meta’s Llama family and prominent open models originating from Asia. Despite carrying the “Small” moniker, this formidable model activates only a fractional subset of its expert sub-networks during any single token inference pass, maintaining manageable memory bandwidth overhead while delivering frontier-tier reasoning capabilities.

Benchmark results released alongside the launch demonstrate that Inkling 2 Small outperforms existing Western open-weight alternatives across advanced mathematical proofs, complex systems programming, and granular multi-step instruction-following evaluations. The model exhibits deep conversational context comprehension and clean problem decomposition - attributes traditionally monopolized by closed, proprietary commercial APIs. Thinking Machines announced that model weights are publicly available for academic and enterprise researchers, along with native integration toolkits supporting distributed inference frameworks.

The debut of Inkling 2 Small represents both a milestone for Thinking Machines and a critical validation for the broader open-source ecosystem. An independent startup training and deploying a model nearing the 600-billion parameter threshold underscores that the technological moat once monopolized by trillion-dollar hyperscalers is progressively narrowing, empowering researchers worldwide to self-host world-class foundational intelligence.

Source: The Decoder

🤖 Perplexity launches Teammates: Always-on 24/7 autonomous agents on dedicated VMs for $200/month

Perplexity is transforming its conversational answer engine into an autonomous digital workplace with the debut of “Teammates.” Unlike traditional chatbot interfaces that terminate after a single query-response round-trip, Teammates are always-on autonomous agents engineered to tackle complex, multi-day project roadmaps without requiring users to maintain active browser sessions. At the architectural core of Teammates is dedicated virtual machine isolation: each agent is provisioned with its own sandboxed cloud VM, persistent file storage, and shell terminal execution rights.

With Teammates, knowledge workers can delegate exhaustive background workflows, such as cross-referencing thousands of regulatory filings, assembling real-time market competitive analyses, or prototyping and stress-testing automated data processing pipelines. When encountering environment runtime errors or conflicting source data, the agent autonomously iterates, debugs command outputs, and updates a centralized Markdown briefing document. Perplexity has positioned the feature exclusively within its new Max subscription tier, priced at $200 per month - a tenfold step up from the standard $20 Pro plan.

This steep price point underscores Perplexity’s strategic positioning: Teammates is marketed as an investment in digital human capital rather than an ordinary software add-on. While the $200 monthly commitment may deter hobbyists, venture capital firms, management consultants, and specialized engineering squads will find the economic proposition of an indefatigable, 24/7 autonomous teammate on dedicated compute extraordinarily compelling.

Source: The Decoder

⚡ METR audit reveals Anthropic doubled model research velocity using AI, potentially crossing warning threshold

Independent nonprofit safety auditor METR (Model Evaluation and Threat Research) has published the findings of an exhaustive eleven-week evaluation of Anthropic’s internal engineering workflows and model research records. The audit confirmed a stunning acceleration: Anthropic’s operational cadence for developing and iterating frontier language models has roughly doubled over the past twelve months. Most significantly, an estimated 80% of this development speedup is attributed not to expanding human headcount or raw compute clusters, but to autonomous AI systems writing training scripts, diagnosing model failures, and designing automated evaluation harnesses.

The most sensitive takeaway from METR’s assessment is that Anthropic may have inadvertently reached or breached its own internal AI acceleration warning threshold, established under its Responsible Scaling Policy (RSP) framework back in July. This specific tripwire was formulated to alert executives and safety oversight boards when autonomous systems begin driving recursive self-improvement at rates that risk outstripping human interpretability and empirical validation.

METR’s findings have reignited intense industry debate regarding recursive capability compounding and the timeline to an intelligence explosion. As autonomous agents progressively assume responsibility for data filtering, hyperparameter search, and novel architecture generation, model lifecycle iterations are compressing from calendar years into single quarters. This exponential feedback loop places unprecedented scrutiny on regulatory authorities and safety verification frameworks attempting to keep pace with algorithmic speed.

Source: The Decoder

💻 Ricursive Intelligence releases Ripple-65B Preview: Apache 2.0 open model tops terminal coding benchmark

San Francisco startup Ricursive Intelligence has released Ripple-65B Preview, a 65-billion parameter open-weight language model licensed under the fully unencumbered Apache 2.0 license for unrestricted commercial deployment and academic research. Rather than pursuing generalized multi-modal chat features, Ricursive purpose-built Ripple-65B as an agentic terminal model, specializing in bash command synthesis, Unix system orchestration, and software automation pipelines.

On the rigorous TerminalB benchmark - an industry-standard evaluation measuring an AI agent’s ability to navigate realistic command-line operating environments, compile system packages, debug Linux kernel modules, and configure container clusters - Ripple-65B Preview secured the top score among all United States open-weight language models. The architecture demonstrates an exceptional grasp of system return codes, directory hierarchy traversal, and proactive error remediation without falling into repetitive loop failures.

Releasing Ripple-65B under genuine Apache 2.0 licensing provides a massive advantage for enterprise DevOps teams, platform engineers, and security-sensitive organizations subject to strict data governance mandates. Rather than piping proprietary source code and infrastructure configurations to closed third-party cloud APIs from OpenAI or Anthropic, organizations can now reliably self-host an elite coding agent entirely on private GPU hardware.

Source: The Decoder

📊 Ramp financial data shows enterprise AI spending stalling just as Anthropic nears IPO

Corporate card and financial spend management platform Ramp has issued a detailed corporate spending retrospective uncovering a striking anomaly in enterprise technology budgets: even as employee AI usage and token consumption rates continue to set historic highs, aggregate enterprise spending on commercial AI software has begun to plateau. Analyzing transactional data across tens of thousands of corporate clients, the report highlights a fundamental behavioral pivot among corporate technology buyers.

The divergence between soaring usage and flattening expenditure is primarily propelled by two market dynamics. First, aggressive price reductions and cutthroat API discounting among frontier model providers - including Google, OpenAI, and Anthropic - have introduced lightweight “Flash” and “Mini” variants delivering high capability at nominal per-token costs. Second, enterprise engineering teams have matured rapidly in production efficiency: aggressively deploying prompt caching, smart multi-model routing, and migrating repetitive background workloads to self-hosted open-source models.

Ramp’s spend data lands at a particularly delicate juncture as Anthropic prepares its initial public offering (IPO) documentation. A deceleration in top-line corporate spending growth could challenge prevailing Wall Street projections of uninhibited AI enterprise hypergrowth, pushing institutional investors to scrutinize recurring software margins and customer retention economics rather than relying on gross adoption metrics.

Source: TechCrunch