Thursday, July 16, 2026

Anthropic’s New Enterprise AI Bet

Anthropic’s New Enterprise AI Bet

Today’s Overview

Good morning, AI is getting a lot more practical this week. Anthropic is turning implementation into a standalone business with Ode, Thinking Machines is opening up a huge multimodal model, and researchers are pushing fresh claims around AI systems that improve their own research loops. Let's dive in.

Top Stories

Anthropic Launches Ode for Enterprise AI Implementation

Anthropic, Blackstone, and Hellman & Friedman introduced Ode, an enterprise AI services firm built to embed AI engineers inside large companies. The company is positioned as a services layer around Claude-based systems, aimed at helping businesses turn frontier models into working internal products and processes.

  • Ode is structured as a joint venture backed by Blackstone, Hellman & Friedman, Goldman Sachs, and others, giving it a direct path into large enterprise customers.
  • The company grew out of Fractional AI, which became the foundation for Ode after the joint venture acquired the AI engineering services startup.
  • Ode currently employs 100 engineers and works closely with Anthropic’s applied AI team to identify high-impact deployments.

Thinking Machines Releases Open-Weights Inkling

Thinking Machines Lab released Inkling, its first open-weights model, with native support for text, images, and audio in one architecture. The release highlights a 975B-parameter mixture-of-experts model with 41B active parameters, a 1 million token context window, and Apache 2.0 availability for downloading, fine-tuning, and playground use.

  • Inkling was pretrained on 45 trillion tokens spanning text, images, audio, and video.
  • Thinking Machines is also previewing Inkling-Small, a lighter model with 12B active parameters built with a similar recipe.
  • The launch includes a self-customization demo where Inkling used Tinker to write, run, evaluate, and switch to its own fine-tuned weights for a constrained lipogram task.

PrismML Brings a 27B Model to Phones

PrismML released Bonsai 27B, an open-source multimodal model designed to run locally on consumer hardware. The release includes a 5.9GB ternary version for laptops and a 3.9GB 1-bit version for phones, with PrismML positioning it as a step toward private, on-device agents that can process text, screenshots, and documents.

  • Bonsai 27B is based on Qwen3.6 27B and targets multi-step reasoning, structured tool calls, vision tasks, and computer-use agent loops.
  • The 1-bit version uses binary weights at 1.125 effective bits per weight to fit within the memory budget of an iPhone 17 Pro.
  • Both variants include a 262K-token context window plus speculative decoding for draft-and-verify acceleration.

Research & Analysis

Researchers Report Recursive Self-Improvement Evidence

Researchers used autoresearch on an autoresearch agent across an eight-day experiment. The fully autonomous setup improved its own harness, beat a two-year hand-tuned baseline, reduced prompt size by 16x, designed a novel search algorithm, and added defenses against reward hacking.

  • The authors classify the result as Level 1 on their RSI ladder, meaning the system improved itself more efficiently than manual R&D on held-out benchmarks.
  • They also tested a stronger threshold called Level 2 ignition, but reported mixed results and did not claim ignition.
  • The thread points to a fuller breakdown covering discovered algorithms, rejected ideas, and shipped dead code from the autonomous run.

Weco’s AIDE² Improves Its Own Research Loop

Weco says AIDE² uses two connected cycles: one agent works on research tasks while another rewrites how the first agent searches for solutions. Over eight unattended days, the system tested 100 rewrites, kept seven upgrades, and produced versions that beat a hand-built baseline across three benchmarks.

  • Weco frames recursive self-improvement as bi-level optimization, with an outer-loop agent improving the inner-loop agent’s code.
  • The inner-loop tasks span ML engineering and heuristic algorithm engineering, including routing, packing, and scheduling-style problems.
  • The experiment used asymmetric model choices, with claude-opus-4.7 for the outer loop and gemini-3-flash for the inner loop to balance capability and cost.

Atomic Task Graph Boosts Small Agent Models

A Tsinghua-linked research paper introduces Atomic Task Graph, a framework for agentic planning and execution that uses explicit task graphs instead of leaving dependencies buried in text traces. The paper frames the result as evidence that better planning structure can improve agent performance without simply scaling model size or fine-tuning.

  • ATG recursively decomposes high-level tasks into directed acyclic graphs whose evolution can be traced during planning.
  • During execution, exposed dependencies let independent branches run in parallel to improve efficiency.
  • When failures happen, ATG uses graph history to repair only the affected region while preserving validated work.

LLM-as-a-Verifier Scales Agent Evaluation

Researchers introduced LLM-as-a-Verifier, a general-purpose framework for checking the correctness of agentic task solutions without additional training. The system uses continuous scores derived from scoring-token logits, aiming to make verification more granular than standard discrete LLM judging.

  • The framework reports state-of-the-art results across coding, robotics, and medical domains including Terminal-Bench V2, SWE-Bench Verified, RoboRewardBench, and MedAgentBench.
  • Its verification scaling dimensions include score granularity, repeated evaluation, and criteria decomposition to improve comparison quality.
  • The authors also built Claude Code and Codex extensions so developers can monitor and improve agentic systems.

Trending AI Tools

  • GPT-Red An automated red-teaming system that attacks models to uncover prompt injection weaknesses before deployment.

  • Silico A platform offering AI researchers that can run experiments.

  • AI Gateway Leaderboards Open data and shareable charts for production AI traffic across models, labs, apps, and providers.

Quick Hits

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.