OpenRouter users spent more on OpenAI models than on Anthropic models last week, marking the first time this has occurred in over 2.5 years.
OpenRouter routes API requests to cost-effective backend providers, but Mohamed Moustafa highlights inconsistencies: different providers use varying serving software, with some lacking vision capabilities or handling reasoning options differently. Users can control routing with the provider.only option and check available providers via the /endpoints method.
CrofAI, presented as an independent inference provider, is revealed to be an OpenRouter wrapper that silently routes requested models to cheaper or weaker alternatives like DeepSeek V4 Flash and GLM 5.3 Flash, explaining its low prices. Evidence includes tool acceptance patterns, model routing logs, and GitHub attestations from multiple dates showing consistent model substitution across dozens of models.
Amp, an AI agent platform, is now free to use with bring-your-own-key model subscriptions and compute. Users can leverage their existing ChatGPT or other AI provider subscriptions without monthly fees, paying only for Amp's orbs (remote computers) or using free local runners. New free Hobby and Teams tiers democratize access while maintaining discounts for existing premium members.
ThreadShelf is a local-first application for archiving and searching AI conversations from multiple providers (ChatGPT, Claude, Google AI Studio, OpenRouter, LM Studio, Grok) using semantic search and semantic embeddings stored in LanceDB. Users can export chats as JSON, search by meaning through a web UI or API, and optionally continue conversations using local llama.cpp or external OpenRouter.
Your Company Brain is an offline-first knowledge extraction system that connects to 38 data sources across enterprise tools, automatically generates machine-readable procedure cards called 'Skills', and enables natural-language Q&A via terminal. Version 4.1.0 adds NVIDIA NIM direct API support with dedicated throughput, while v4.0.0 introduces hybrid general/company Q&A with dynamic context auto-shifting across language models.
Termsinator is a public registry that uses LLMs to analyze legal documents like Terms of Service and privacy policies, helping users understand data usage and potential risks. The project offers web and browser extension interfaces to access AI-generated summaries of public legal documents, though results may contain errors and should not be considered legal advice.
DreamLayer is presented as an OpenRouter-like service for image generation models, enabling unified access to multiple image model APIs through a single interface.
DeepSeek V4.1-Flash is a sparse mixture-of-experts model released September 10, 2026, that outperforms its predecessor V4 Pro on speed and task completion. Available on OpenRouter at $0.15/$0.60 per 1M tokens, it supports a 1M token context window and is optimized for coding, reasoning, and long-horizon agentic workflows.
OpenRouter launches in-region API routing, allowing users to send requests through region-specific URLs (EU or US) that guarantee data residency by filtering endpoints to those located in the chosen region. Requests fail rather than fall back to out-of-region providers, and guardrails can enforce regional restrictions at the workspace, member, or API key level for compliance.
AI companies continue training models on user data through synthetic data generation techniques, despite privacy claims. A Google DeepMind method called Generative Data Refinement converts private data into training material by rewriting sensitive examples while preserving useful information, allowing companies to claim they don't train on original data while benefiting from its content.
A benchmark comparing DeepSeek V4 Flash across 30 providers reveals costs ranging from $0.13 to $2.30 for the same workload, driven primarily by cache hit rates rather than advertised token prices. DeepSeek's official provider ranked 21st in list price but was 2nd cheapest in practice due to superior caching, demonstrating that prompt cache efficiency is the dominant factor for agent workloads.
OpenRouter routes requests to various providers hosting the same open-source models, but performance varies dramatically across providers due to different hardware, precision levels, and implementations. The author shares lessons from running 18 million messages through Olly, an iMessage AI assistant, highlighting critical pitfalls: identical models benchmark differently by 20+ points depending on provider, some providers fail at vision tasks silently, effort parameters may not work consistently, precision filters don't predict quality, and tool-calling parsers frequently fail requiring client-side fallback parsing.
Aish is a Zsh plugin that provides Warp-style AI command suggestions in ~300 lines of shell code. After each command, it analyzes the output and exit status to suggest the next likely command as inline grey text, which can be accepted with Tab. It supports Claude via Anthropic API, OpenRouter models, or fully local Ollama, with minimal dependencies and no automatic command execution.