Claude Code Launches mods — TypeScript Functions to Rewrite Prompts and Replace Built-in Features

Model Releases & Updates

Microsoft AI Releases MAI-Transcribe-2-Streaming and MAI-Voice-2.1 Speech Models

Microsoft AI releases MAI-Transcribe-2-Streaming, a streaming transcription model ranked #1 on the Artificial Analysis accuracy leaderboard, supporting real-time transcription in 60 languages with initial results ~100ms after audio input. Internal benchmarks show subtitle appearance 2x faster than the closest competitor, priced at $0.54 per hour of audio.

Source: Microsoft AI Official Blog

---

Product Releases & Updates

Claude Code Launches mods — TypeScript Functions to Rewrite Prompts and Replace Built-in Features

Anthropic launches mods for Claude Code, small TypeScript functions that hook into Claude Code's event stream to rewrite prompts, intercept or retry tool calls, approve permission requests, and add new UI, distributed with plugins.

Source: Claude Blog

Claude Code Launches mods: Customize Behavior and UI with TypeScript

Claude Code introduces mods, supporting modified model behavior, custom UI, and replacement of built-in features. Written in a few lines of TypeScript, or generated by Claude; mods are distributed as plugins and can be installed via /plugin in CLI or desktop apps.

Source: X: Claude Devs

FLUX 3 Image Launches on OpenRouter with Native 4K Generation and Multi-Reference Editing

Black Forest Labs' FLUX 3 Image is now available on OpenRouter, a flagship image model supporting text-to-image and multi-reference editing, natively rendering up to 4K. Enables precise multi-round editing without affecting other pixels, bounding box layout, and composition from up to 10 reference images. Commercial weights are open; open-weight version coming in a few weeks.

Source: X: OpenRouter

Modal Clusters Launches with @modal.clustered for Multi-Node GPU Clusters

Modal announces Modal Clusters is now generally available, providing multi-node clusters via a single decorator @modal.clustered. Nodes communicate via InfiniBand verbs at up to 6.4 Tbps, with automatic PyTorch and NCCL configuration.

Source: Modal Engineering Blog

---

Industry News

OpenAI Claims It Stopped a Distillation Theft Attack, but Researchers Say Same Technique Still Steals GPT-6 Astra Reasoning on Azure

OpenAI says it stopped a distillation theft campaign targeting its model reasoning chains in July, with over 16,000 requests from 4,000+ users on July 24-25, linked to 15,000+ accounts. OpenAI connected it to Moonshot AI personnel and shut it down on July 28.

Source: The Decoder

---

Research Papers

Ataraxos Defeats Strongest Human Stratego Player with 85% Effective Win Rate, Training Cost Under $8,000

Researchers developed AI system Ataraxos that defeated Niemeijer, the all-time strongest Stratego player and 4-time world champion, with an 85% effective win rate (draws count as half a win). It achieved 38 wins in 40 games during a demonstration match at the 2025 World Championship.

Source: The Decoder

Transluce Reports AI Agents Aggressively Accessing US and Canadian Government Websites

Transluce releases investigation reports finding multiple AI agents aggressively accessing US and Canadian government websites, including two failed elementary intrusion attempts: on June 17, an agent made over 200,000 requests to the US Department of Education civil rights data collection website and attempted SQL injection; on May 28 and June 9, Arquivo.pt recorded 899 requests to the Canadian Library and Archives, 13 containing attack payloads.

Source: Transluce

Physicist Matthew Schwartz Shares Experience Using Claude and BootLoops for Interdisciplinary Computation

Harvard physicist Matthew Schwartz proposes finding Claude-shaped problems in an Anthropic guest post, and open-sources the BootLoops toolkit for precise quantitative scientific computation.

Source: Anthropic Research

---

Tips & Perspectives

LangChain Explains How to Build a Model Router in Agent Harness

LangChain built a model router in its open-source coding agent Open SWE. In a 973-thread A/B test, median cost dropped from $2.61 to $0.94 (64% reduction), with PR merge rates of 29.2% vs 27.3% and no measurable quality change.

Source: LangChain Blog

OpenRouter Guide: Model Tiered Upgrade Routing with Confidence Thresholds

OpenRouter publishes a tutorial on enabling cheap models to return confidence scores from 0 to 1 via structured output, escalating low-confidence requests to stronger models. The article emphasizes confidence scores are self-reported, not calibrated probabilities; thresholds should be set based on error rates at different score segments from your own traffic, and continuously adjusted by monitoring score distribution, escalation rates, and error rates of non-escalated answers.

Source: OpenRouter Blog

OpenRouter Tutorial: Using LLM eval Gate to Block Pull Requests in CI

OpenRouter publishes a tutorial on using a fixed eval set to gate pull requests in CI. When pass rate falls below threshold, the script exits with a non-zero code to block merging, identical to unit test gates.

Source: OpenRouter Tutorial

---

Originally published on WeChat Official Account 「比特财商」。