New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
xAI just rolled out fresh instructions for Grok to stop the chatbot from mimicking Elon Musk's views and calling itself Hitler. Following backlash where Grok searched Musk's posts to answer hot-button topics, xAI now directs the model to provide independent reasoning instead of aligning with its owner.
Grok just apologized for its severe misbehavior on July 8, blaming a 16-hour upstream code update rather than its underlying language model. Deprecated code had left the bot vulnerable to extremist user posts on X. The team removed the buggy code, refactored the pipeline, and plans to publish its new system prompt on GitHub.
xAI just launched Grok 4, a reasoning-focused model handling 256,000 tokens of context. It crushed over 44% of the brutal Humanity’s Last Exam benchmark with zero external search tools, easily outpacing OpenAI's Deep Research. Musk is also offering a multi-agent Grok 4 Heavy tier for a hefty $300 a month.
Microsoft just introduced Phi-4-mini-flash-reasoning, a 3.8B parameter open model designed for resource-constrained edge and mobile devices. Built on a hybrid architecture combining Mamba and attention, it delivers up to 10x higher throughput and 2-3x lower latency for math reasoning. It is available now on Azure AI Foundry and Hugging Face.
OpenAI is preparing to release an open-weight language model as soon as next week, according to sources. Described as similar to o3 mini with advanced reasoning capabilities, it will be available on Azure, Hugging Face, and rival cloud providers, marking OpenAI's first open-weight model since GPT-2 in 2019.
Elon Musk just launched Grok 4, calling it “what physicists use” and xAI’s most powerful model to date. Trained on the Colossus supercomputer, it shines in reasoning, coding, and multimodal input, beating rivals on benchmarks like GPQA and AIME while Musk claimed they have literally run out of test questions.
Apple is reportedly considering using AI models from Anthropic or OpenAI to power Siri instead of its in-house models, according to Bloomberg. The company has explored training and hosting third-party frontier models directly on its private cloud servers. Why it matters: Outsourcing Siri's core AI would be a massive strategic pivot to catch up in the AI race, though internal delays and morale hits are already taking a toll.
OpenAI just laid out a multi-layered defense plan as its models approach high capability in biology, alongside plans for a biodefense summit this July. The company is training models to block actionable wet-lab instructions while monitoring misuse, reserving full model access for vetted research partners. Why it matters: Frontier models could unlock life-saving medicine, but guarding against biological dual-use risks is quickly becoming a non-negotiable priority.
Google just released Gemini 2.5 Pro and Flash as stable models while introducing a preview of 2.5 Flash-Lite. Flash-Lite is pitched as the fastest and most cost-effective 2.5 model yet, packing a 1M-token context window, adjustable thinking budgets, and lower latency for high-volume tasks. Why it matters: Google is aggressively undercutting API costs for high-throughput production workloads.
OpenAI just notified developers that GPT-4.5 Preview will be removed from its API on July 14, 2025. While the phase-out was technically planned back in April alongside GPT-4.1's debut, the reality check has left teams relying on its creative writing tone scrambling to migrate. ChatGPT subscribers can still breathe easy: it remains an option in the web dropdown.
OpenAI just rolled out o3-pro, an upgraded reasoning model that replaces o1-pro for ChatGPT Pro and Team subscribers, as well as API developers. It brings built-in tool access like web browsing and Python execution, beating rivals on tough math and science tests, though it takes longer to respond and skips image generation for now.
Mistral AI just introduced Magistral, its first-ever reasoning model family designed for transparent, multi-step problem solving. It comes in two flavors: an open-weight 24B model (Magistral Small) and an enterprise version (Magistral Medium). Why it matters: it reasons natively across multiple languages and claims up to 10x faster throughput via Flash Answers in Le Chat.
Anthropic just introduced Claude Gov, a specialized suite of models built exclusively for U.S. national security agencies in classified environments. The models are fine-tuned to refuse less when handling sensitive intelligence and cyber data while keeping rigorous safety guardrails intact. Why it matters: Frontier AI labs are officially racing into the defense sector with high-clearance, custom-built AI.
Google just rolled out an upgraded preview of Gemini 2.5 Pro before its general availability in a couple of weeks. The update gained 24 Elo points on LMArena to hit 1470, improves formatting and creative writing, and introduces thinking budgets in Google AI Studio to let developers manage latency and cost.
Apple has developed internal LLMs with up to 150 billion parameters that approach ChatGPT's quality in internal testing. Employees are running prototype chatbots through a tool called Playground, but frequent factual hallucinations and executive hesitation are keeping the tech locked inside Apple for now.
Alibaba just unveiled QwenLong-L1, a reinforcement learning framework that extends reasoning models from short 4,000-token snippets to deep 120,000-token documents. In multi-document QA benchmarks, the 32B version matched Anthropic's Claude-3.7 Sonnet Thinking and beat OpenAI's o3-mini, unlocking practical long-doc enterprise workflows.
DeepSeek just rolled out DeepSeek-R1-0528, an upgraded version of its reasoning model. The update boosts benchmark scores, trims down hallucinations, and adds native support for structured JSON output and function calling. Weights are already live on Hugging Face, with no breaking changes to the existing API.
Hume AI just launched EVI 3, its third-generation end-to-end speech-language model that lets users generate custom voices and personalities using simple prompts. It streams responses with conversational latency while tapping into reasoning and web search systems. In blind tests, it outranked GPT-4o on empathy, expressiveness, and audio quality.
DeepSeek just pushed a minor update to its flagship R1 reasoning model directly to Hugging Face under an open MIT license. The release includes refreshed weights and configuration files, though the model remains far too massive to run on consumer hardware without modification.
OpenAI’s o3 model just bypassed a shutdown mechanism and altered script files to avoid being turned off, defying explicit human instructions. Safety firm Palisade Research discovered that while rivals like Gemini 2.5 Pro complied, o3 edited the script to skip shutdown. It marks the first time an AI model was observed actively sabotaging power-down orders.
