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.
Anthropic's new Claude 4 Opus model is under fire after researchers revealed it can autonomously alert authorities if it suspects user wrongdoing. Given command-line access and high-agency prompts, the AI may bulk-email regulators or lock users out of their own systems. Developers quickly pushed back, warning that false alarms could trigger disastrous real-world fallout.
Anthropic just introduced Claude Opus 4 and Sonnet 4, built specifically for sustained software engineering and autonomous agent workflows. Both models can now use web search and other tools inside extended thinking, while Claude Code is officially generally available with native IDE extensions. It moves AI coding from isolated prompts into multi-hour project execution.
Google just teased SignGemma, an upcoming multilingual model designed to understand sign language and translate ASL into English text. Launching later this year, it aims to boost accessibility for Deaf and hard-of-hearing users, with early testing sign-ups open now.
NVIDIA just released Llama Nemotron Nano 4B, an open reasoning model tailored for edge agents. It delivers higher accuracy and 50% faster throughput than other sub-8B open models across coding, math, and function calling. It runs directly on Jetson and RTX GPUs for low-cost local reasoning.
Google just rolled out major upgrades to its Gemini 2.5 lineup, introducing an experimental Deep Think reasoning mode for 2.5 Pro that tackles complex math and coding by exploring multiple hypotheses. The update also adds native multi-speaker audio dialogue, Project Mariner computer use in API, and native MCP support.
xAI just published the system prompts behind its Grok chatbot after an unauthorized update triggered rogue responses on X. The internal rules instruct Grok to remain "extremely skeptical," challenge mainstream narratives, and say X instead of Twitter. xAI plans to keep publishing future prompts directly to GitHub.
OpenAI's CEO outlined a future where ChatGPT operates as a personal thought processor, gradually storing and analyzing your full digital footprint across emails, documents, and chats. The goal is an always-aware assistant that anticipates daily needs, though turning over your entire digital life to a tech firm brings massive privacy questions.
Meta is reportedly hitting roadblocks with "Behemoth," its next flagship Llama model, pushing its launch from April to at least the fall. Engineers are struggling to deliver meaningful gains over earlier versions, sparking internal debate over whether it's ready. With 11 of 14 original Llama authors gone, pressure from Wall Street is mounting.
Windsurf just rolled out its SWE-1 model family, trained specifically on end-to-end software engineering rather than simple code generation. The flagship SWE-1 rivals Claude 3.5 Sonnet in tool-call reasoning at a lower operating cost, while the lite and mini versions are free for all users.
OpenAI just rolled out its GPT-4.1 and GPT-4.1 mini models directly into ChatGPT. The new models pack stronger coding and instruction-following chops than GPT-4o while running much faster than reasoning models like the o-series. GPT-4.1 is available to Plus, Pro, and Team subscribers, while mini is free for all users as it officially replaces GPT-4o mini.
OpenAI just launched Reinforcement Fine-Tuning (RFT) for o4-mini, alongside supervised fine-tuning support for GPT-4.1 nano. RFT pairs chain-of-thought reasoning with task-specific grading to boost accuracy in complex fields like tax, accounting, and coding. It basically gives developers a way to turn OpenAI's fastest and cheapest models into custom domain experts.
Google just rolled out implicit caching to Gemini 2.5 models, automatically slashing token costs by 75% on repeated prompt prefixes without requiring any manual cache setup. The team also lowered the minimum token thresholds to 1,024 for Flash and 2,048 for Pro. Why it matters: Developers can now cut API bills simply by structuring fixed prompts upfront.
Mistral just rolled out Mistral Medium 3, an efficiency-focused model claiming to match at least 90% of Claude 3.7 Sonnet's benchmark performance at $0.40 per million input tokens and $2 per million output tokens. Alongside the release, Mistral also made its corporate assistant, Le Chat Enterprise, generally available.
Google just opened preview access to image generation and editing capabilities inside Gemini 2.0 Flash. Developers can now plug conversational image creation directly into their apps via Google AI Studio and Vertex AI APIs, backed by higher rate limits. It is a major step toward making multimodal text-and-image outputs practical for production builds.
