1. SpaceXAI Graduates Coding Model to Grok 4.5
SpaceXAI has officially launched Grok 4.5, the successor to the grok-build-0.1 beta model. While the beta focused on web development and debugging, Grok 4.5 is a more robust model trained in partnership with Cursor, specifically optimized for complex software engineering and autonomous agentic tasks. It is now available in Grok Build, Cursor, and the SpaceXAI console.
- • Grok 4.5 succeeds the grok-build-0.1 beta, offering enhanced performance for software engineering.
- • Priced at $2 per million input tokens and $6 per million output tokens.
- • Trained in partnership with Cursor using interaction data and SpaceX's Colossus supercomputer.
- • Achieved 64.7% on SWE Bench Pro, outperforming previous iterations.
- • Available in Grok Build, Cursor, and the SpaceXAI console.
Developers now have access to a production-grade, software-engineering-focused model that improves upon the capabilities and efficiency of the previous beta release.
2. OpenAI Officially Launches GPT-Live Full-Duplex Voice Models
OpenAI has transitioned its bidirectional audio research, previously identified as GPT-Bidi-1, into a production-ready product called GPT-Live. The new GPT-Live-1 and GPT-Live-1 mini models enable full-duplex voice interaction, allowing for simultaneous listening and speaking. The models are now rolling out globally to ChatGPT subscribers, with the smaller 'mini' version serving as the default for free users.
- • OpenAI has officially launched GPT-Live-1 and GPT-Live-1 mini, the production version of the previously teased GPT-Bidi-1 model.
- • The models feature a full-duplex architecture that allows for simultaneous speaking and listening.
- • The system is rolling out globally across iOS, Android, and the web for ChatGPT users.
- • GPT-Live decouples the voice interaction layer from the reasoning layer, offloading complex tasks to GPT-5.5.
- • API access is not yet available, but a sign-up form for future access has been opened.
This release marks the transition of the previously announced bidirectional audio technology into a consumer-facing product, enabling natural, interruptible voice conversations in ChatGPT.
3. OpenAI Expands GPT-5.6 Preview to Global Developers
Following the initial government-restricted preview of the GPT-5.6 series announced on June 26, OpenAI is now expanding access to a global audience. This expansion allows a broader set of developers to test the flagship Sol model, the balanced Terra, and the cost-effective Luna before the models reach general availability.
- • OpenAI is expanding the GPT-5.6 preview from government-only to a global developer audience.
- • The lineup includes the flagship Sol, balanced Terra, and low-cost Luna models.
- • This expansion serves as the final testing phase before the models reach general availability.
- • Developers can now evaluate the models' performance and cost-efficiency for their specific use cases.
Developers who were previously excluded from the limited government-only preview can now begin integrating and testing the new GPT-5.6 lineup, including the flagship and cost-optimized variants, ahead of the full public release.
4. Google Expands Gemma 4 Family with Full Range of Multimodal Models
Google has expanded its Gemma 4 open-weight multimodal lineup, moving beyond the previously released 12B model to offer a full suite of sizes from 2.3B to 31B parameters. These models feature a unified encoder-free architecture, native thinking mode, and long-context support, all released under the permissive Apache 2.0 license to facilitate local, resource-constrained deployments.
- • The Gemma 4 family now includes models ranging from 2.3 billion to 31 billion parameters.
- • The full lineup maintains the unified encoder-free architecture and native thinking mode introduced in the 12B version.
- • All models are available under the Apache 2.0 license.
- • The expansion provides more options for balancing performance and hardware requirements in local deployments.
Developers now have access to a broader range of model sizes, allowing for better scaling between lightweight 2.3B models and more capable 31B variants for local multimodal applications.
5. NVIDIA Releases Audex 30B Audio-Text Mixture-of-Experts Model
NVIDIA has launched Audex, a 30-billion-parameter audio-text Mixture-of-Experts (MoE) model that activates 3B parameters per token. To prevent the common issue of text performance degradation when adding multimodal capabilities, Audex was trained using a multi-stage SFT curriculum and text-only Cascade RL. The model supports a 1M token context length, integrates with standard serving frameworks like vLLM, and achieved a 6.82 average word error rate on the OpenASR leaderboard. However, developers should note it is released under a noncommercial license.
- • NVIDIA released Audex (Nemotron-Labs-Audex-30B-A3B), a unified audio-text Mixture-of-Experts model.
- • The model activates 3 billion parameters per token out of its 30-billion-parameter total.
- • It is built on the Nemotron-Cascade-2-30B-A3B text-only backbone, a hybrid Mamba-Transformer.
- • Audex supports a 1 million token context length and is compatible with Megatron-LM and vLLM.
- • The model uses AF-Whisper for audio encoding and is released under a noncommercial license.
Developers can build advanced audio-text applications using an open model that maintains high-quality text reasoning and is compatible with standard LLM infrastructure like vLLM.
6. HalluSquatting Vulnerability Threatens Popular AI Coding Assistants
Security researchers have uncovered a novel attack vector called HalluSquatting that targets popular AI coding assistants and developer agents, including Cursor, GitHub Copilot, Windsurf, and Cline. Unlike traditional push-based prompt injections, HalluSquatting is a pull-based attack that predicts resource identifiers (such as package names or repositories) an LLM is likely to hallucinate. Attackers then register these identifiers and seed them with malicious instructions, potentially leading to local device infections and botnet assembly when developers unknowingly run the hallucinated resources.
- • Researchers developed HalluSquatting, a pull-based attack that exploits an LLM's tendency to hallucinate resource identifiers.
- • The attack works by predicting identifiers an LLM is likely to hallucinate, then registering and seeding them with malicious instructions.
- • Susceptible AI coding assistants and agents include Cursor, Cursor CLI, Gemini CLI, Windsurf, GitHub Copilot, Cline, OpenClaw, ZeroClaw, and NanoClaw.
- • HalluSquatting enables large-scale attacks, including the assembly of botnets, DDoS attacks, and device infections.
- • Current AI security mitigations rely on guardrails rather than addressing the root cause of separating trusted and untrusted sources.
Developers using AI coding assistants must be aware of how hallucinated resource identifiers can be registered by attackers to execute malicious code on their local machines.
7. Google Adds Remote MCP and Background Execution to Gemini Managed Agents
Building on the Managed Agents framework introduced in May, Google has rolled out a major update to the Gemini API. Developers can now integrate remote Model Context Protocol (MCP) servers and execute tasks in the background. By consolidating reasoning, package installation, and file management into a single endpoint executed within an isolated cloud sandbox, this update further simplifies the deployment of secure, autonomous agent workflows.
- • Google added remote MCP server integration and background execution to the Gemini API Managed Agents framework.
- • Managed agents can now operate as asynchronous workers in real development environments without blocking applications.
- • The update maintains the existing isolated cloud sandbox architecture for secure code execution and package installation.
- • New features include custom function calling and credential refreshing across agent interactions.
These additions allow developers to deploy more complex, asynchronous, and long-running agents that securely connect to external MCP-compliant tools without managing additional infrastructure.
8. Salesforce Integrates Slackbot with Salesforce Platform via MCP
Salesforce has rolled out a major integration that connects Slackbot to the entire Salesforce platform using the Model Context Protocol (MCP). This integration allows teams to query CRM data, access Tableau analytics, and trigger external workflows like DocuSign approvals directly from shared Slack channels. Crucially, the system automatically respects existing Salesforce permissions and field-level security, demonstrating a robust, production-grade implementation of MCP that is supported by a growing ecosystem of partners including Atlassian, Box, and Zoom.
- • Salesforce connected Slackbot to the Salesforce platform using the Model Context Protocol (MCP).
- • The integration allows users to access CRM data, Tableau analytics, and Data 360 profiles, and trigger actions like DocuSign approvals directly from Slack.
- • Slackbot automatically respects existing Salesforce permissions, validation rules, and field-level security settings.
- • The new MCP-native partner ecosystem includes Atlassian, Box, DocuSign, Canva, Lucid, and Zoom.
This major enterprise deployment of the Model Context Protocol (MCP) validates the protocol's readiness for production and highlights how developers can build secure, multi-tool integrations.
9. Google AI Studio Adds 'Import from GitHub' to Build Mode
Google AI Studio has updated its prompt-based 'vibe coding' interface, Build mode, with a new 'import from GitHub' feature. This addition allows developers to pull in an existing GitHub repository and convert it into an editable, deployable full-stack application. To ensure security during rapid prototyping, AI Studio automatically configures Gemini API keys as server-side secrets, preventing them from being exposed in client-side code.
- • Google AI Studio introduced an 'import from GitHub' feature within its Build mode.
- • Build mode is a 'vibe coding' interface that generates full-stack applications from user prompts.
- • The feature allows users to import a repository and convert it into a runtime-compatible format for deployment.
- • When using the Gemini API, AI Studio automatically configures API keys as server-side secrets to prevent client-side exposure.
- • The update was announced by Google AI Studio product lead Logan Kilpatrick.
Developers can quickly import existing codebases, scripts, or hackathon projects to iterate on them and deploy them as full-stack apps with automatically secured API keys.
10. CopilotKit Extends AG-UI Protocol to Enable Self-Learning Agents
CopilotKit has expanded its AG-UI protocol—originally launched to handle UI streaming and state synchronization—to support self-learning agent architectures. By capturing agent execution traces and in-browser user corrections, the system converts these interactions into procedural and episodic memory. This development fulfills the company's earlier roadmap goal of enabling continuous learning from human feedback, allowing agents to adapt to user behavior while maintaining restricted learning scopes for specific applications or teams.
- • Extends the AG-UI protocol to support self-learning agent architectures.
- • Captures agent traces and user corrections to build procedural and episodic memory.
- • Fulfills the previously announced roadmap goal of continuous learning from human feedback.
- • Maintains security by restricting learning scopes to specific users, teams, or applications.
Developers can now implement self-improving agent loops using the existing AG-UI infrastructure, moving from basic interaction handling to agents that actively learn from real-world failures and corrections.
11. NVIDIA Adds Cosmos-Framework Tutorial for Miniature World Models
Following the June 1 release of the Cosmos 3 open-weights models, NVIDIA has published a tutorial for its cosmos-framework. This resource guides developers through creating a miniature version of the model architecture, which is designed to run on Google Colab. While full Cosmos 3 inference requires 16B+ parameters and H100-class hardware, this educational implementation uses an omnimodal Mixture-of-Transformers architecture to help developers learn the framework before scaling to full-size deployments.
- • The tutorial provides a miniature, Colab-friendly implementation of the Cosmos 3 world model architecture.
- • This resource serves as an educational stepping stone for developers unable to run full-scale Cosmos 3 inference, which requires 16B+ parameters and H100 hardware.
- • The miniature model demonstrates the omnimodal Mixture-of-Transformers architecture, including RMSNorm, rotary embeddings, and SwiGLU experts.
- • The tutorial includes schema-correct input specifications and launch commands for users ready to transition to full-scale inference on H100-class hardware.
Developers can now learn to implement and train world models using shared cross-modal attention and expert routing on accessible hardware before scaling up to production-grade Cosmos 3 models.
12. Google Updates Android Bench with Cost Metrics and New Models
Google has updated Android Bench, its benchmark designed to help developers evaluate and select the best AI agents for Android development tasks. The latest update introduces critical metrics for cost and efficiency, supports open-weight models, and adds eight new models to the leaderboard. Claude Fable 5 currently leads the benchmark with 84.5% accuracy, followed by Claude Sonnet 5, while Google's own Gemini 3.1 Pro ranks fifth.
- • Google updated its Android Bench benchmark, which evaluates LLM performance on 100 Android development tasks.
- • The update adds metrics for cost and efficiency and introduces support for open-weight models.
- • Eight new models were added to the leaderboard, including Claude Fable 5, Claude Sonnet 5, GLM 5.2, and MiniMax M3.
- • Claude Fable 5 leads the current leaderboard with 84.5% accuracy, while Google's Gemini 3.1 Pro ranks fifth.
Developers can use this updated benchmark to compare the cost, efficiency, and accuracy of leading models when building Android-focused coding agents.
13. French Startup ZML Releases ZML/LLMD to Reduce Inference Costs
French AI startup ZML, backed by an endorsement from Yann LeCun, has released a free software tool named ZML/LLMD. The product is designed to accelerate model inference across a variety of AI chips, helping developers optimize their hardware utilization. By improving execution speeds, the software aims to lower the operational costs associated with running self-hosted AI models.
- • French AI startup ZML has released a free software product called ZML/LLMD.
- • The software is designed to speed up model inference across multiple AI chips.
- • ZML/LLMD aims to potentially reduce the overall costs associated with running AI models.
- • The startup has received an endorsement from Turing Award winner Yann LeCun.
Developers looking to optimize self-hosted model deployments can leverage this free tool to potentially lower their hardware and inference costs.
14. Anthropic Extends Claude Fable 5 Access with Promotional Free Period
Building on the July 7 shift of Claude Fable 5 to a credit-based usage model, Anthropic has launched a promotional window allowing Pro, Max, Team, and premium Enterprise subscribers to access the model at no additional cost. Through July 12, users can apply up to 50% of their weekly subscription limits toward Fable 5, after which they can continue using the model via standard usage credits or switch to other models.
- • Anthropic is offering free promotional access to Claude Fable 5 through July 12.
- • The offer follows the July 7 transition of the model to a credit-based usage system.
- • Eligible users include Pro, Max, Team, and premium Enterprise plan subscribers.
- • Users can allocate up to 50% of their weekly subscription limits to the model during the promotion.
This promotion provides a temporary extension of the model's availability for developers to test and evaluate their applications against Claude Fable 5 without immediate impact on their credit balances.