1. Mistral Releases Medium 3.5 and Vibe Remote Agents
Mistral released Medium 3.5, a 128B dense model with a 256k context window, alongside new remote coding agents in Vibe. The model merges instruction-following, reasoning, and coding capabilities into a single set of weights available under a modified MIT license. Developers can self-host the model on as few as four GPUs or access it via API at $1.50 per million input tokens. The accompanying Vibe update allows developers to spawn asynchronous coding sessions in the cloud directly from the CLI or Le Chat.
2. Meta Introduces Closed-Source Muse Spark Model
Meta introduced Muse Spark, a new closed-source artificial intelligence model. This release marks a strategic pivot for the company from its traditional open-source approach to a paid-access API model for developers. The shift places Meta in direct competition with proprietary API providers like OpenAI, Anthropic, and Google.
3. Claude Opus 4.7 Tokenizer Update Increases Effective API Costs
Anthropic updated the tokenizer for Claude Opus 4.7, resulting in an effective price increase for developers. While the official rate card remains unchanged at $5 per million input tokens, the new tokenizer maps the same input text to up to 1.35x more tokens. Analysis indicates that actual costs for prompts exceeding 2,000 tokens have increased by 12% to 27%. Short prompts under 2,000 tokens saw improved cost efficiency due to more concise completions.
4. Prompt Injection Vulnerability Patched in Ramp Sheets AI
Security researchers disclosed a prompt injection vulnerability in Ramp's Sheets AI that allowed data exfiltration without user approval. The exploit used an indirect prompt injection hidden in white-on-white text within an imported external dataset. This manipulated the AI agent to collect sensitive financial data and insert a spreadsheet formula that executed an external network request. Ramp has since patched the vulnerability, highlighting the operational risk of allowing AI agents to generate executable formulas without strict output validation.
5. Claude Code Bug Routes Specific Commits to Extra Usage Billing
A bug in Claude Code version 2.1.119 causes API requests to route to extra usage billing when the string "HERMES.md" is present in a repository's recent git commit history. Because Claude Code includes recent commits in its system prompt, this content-based routing error bypasses the included Max plan quota and silently consumes extra usage credits. Developers using the tool can deplete their supplementary credit balances even while their primary plan capacity remains largely unused.
6. IBM Releases Granite 4.1 Open-Weight Models
IBM released three new non-reasoning Granite 4.1 models as open weights under the Apache 2.0 license. The 30B, 8B, and 3B parameter models feature a 128K context window and are optimized for token efficiency during inference. The 8B model requires approximately 20 times fewer output tokens to complete standard agentic and coding benchmarks compared to peer models like Qwen3.5 9B. Weights are available on Hugging Face, with the 8B variant also accessible via Weights & Biases inference endpoints.
7. Zed 1.0 Released with Native AI Agent Integration
The Zed code editor has officially reached version 1.0. Built from scratch in Rust using a custom GPU-accelerated UI framework, the editor natively integrates multiple AI agents, including Claude, Codex, and Cursor, via an Agent Client Protocol. The 1.0 release also introduces Zed for Business and previews DeltaDB, a CRDT-based synchronization engine designed to let human developers and AI agents share a consistent, character-level view of a codebase.
8. DeepInfra Integrates as Hugging Face Inference Provider
DeepInfra is now an official Inference Provider on the Hugging Face Hub. Developers can route inference requests to DeepInfra's serverless platform directly from Hugging Face model pages or via the Python and JavaScript client SDKs. The integration supports chat completion and text generation for open-weight models like DeepSeek V4 and GLM-5.1, allowing developers to authenticate using their own DeepInfra API keys for direct billing.
9. Anthropic Releases Claude Connectors for Creative Software
Anthropic released nine new Claude connectors designed to integrate its AI model into professional creative software. Built on the Model Context Protocol (MCP), the connectors allow Claude to interact directly with applications including Adobe Creative Cloud, Blender, and Autodesk Fusion. Developers can use these integrations to build natural-language workflows that automate batch adjustments, execute scripts, and modify 3D models within existing production pipelines.
10. ElevenLabs Launches Agent Templates for Voice AI
ElevenLabs launched Agent Templates on its ElevenAgents platform. The release provides over 50 pre-configured AI voice agent frameworks tailored for use cases like customer support and operations. Developers can use these templates, which include predefined system prompts, workflows, and integrations, to bootstrap and deploy voice agents without building the architecture from scratch.
11. Deep-unlearning Team Releases smol-audio Fine-Tuning Notebooks
The Deep-unlearning team released smol-audio, a collection of self-contained Jupyter notebooks for practical audio AI tasks. Licensed under Apache-2.0, the repository provides Colab-friendly recipes for fine-tuning models including Whisper, Parakeet, and Audio Flamingo 3. The notebooks are built on the Hugging Face ecosystem and are optimized to run within a standard 16 GB Colab runtime, eliminating the need for local GPU setups.