New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Anthropic just released Claude 3.5 Sonnet, leapfrogging its own flagship Opus and rival models on coding and reasoning benchmarks at twice the speed. It also introduced Artifacts, a new side-by-side workspace on Claude.ai that lets you view, edit, and run code or designs in real time. It is available for free on Claude.ai and via API.
Anthropic just made tool use generally available across the entire Claude 3 model family via its API, Amazon Bedrock, and Google Cloud's Vertex AI. Developers can now connect Claude to external APIs to extract structured data, automate multi-step workflows, and stream tool interactions. Why it matters: It gives Claude the hands and feet needed to act as a true agent, though parallel tool calls aren't supported yet.
Mistral AI just dropped Codestral, its first-ever generative model fine-tuned specifically for coding, sporting 22B parameters and a 32k context window. Fluent in over 80 programming languages, it shines at fill-in-the-middle completions and repo-level tasks. Why it matters: Developers get an open-weight, low-latency engine directly inside their IDEs, though commercial deployments require a custom license.
Meta just unveiled Chameleon, a natively multimodal model family that processes and generates interleaved images and text within a single unified token space. The 34B parameter version matches or beats rivals on visual QA while staying competitive on text, though weights aren't publicly released yet.
Microsoft just made its lightweight Phi-3 model family available to developers, bringing Phi-3-mini to Azure AI alongside small and medium versions. The company also previewed Phi-3-vision, a 4.2-billion-parameter multimodal model built for chart and image reasoning. Why it matters: Powerful compact models like this pave the way for real-world AI PC deployments.
Google just introduced LearnLM, a new family of models fine-tuned from Gemini designed to turn AI from an answer generator into an active tutor. Grounded in learning science, it powers features like interactive "raise your hand" prompts on YouTube, step-by-step guidance in Gemini, and simplified Search overviews. Why it matters: generative AI is shifting from an easy cheating shortcut to structured pedagogy.
Google just dropped a teaser for an unreleased multimodal AI interface running on Android, right after OpenAI launched GPT-4o. The video shows the assistant using a phone's camera to understand its surroundings in real time and answer questions about the Google I/O stage setup. OpenAI may have had the spotlight for only 24 hours, though real-world performance remains to be seen at the keynote.
OpenAI just launched GPT-4o, an end-to-end multimodal flagship model that reasons across text, audio, and vision in real time. It cuts speech latency down to human-like speeds, slashes API costs by 50%, and is rolling out to all free ChatGPT users.
Microsoft is reportedly developing an in-house AI language model called MAI-1, packing roughly 500 billion parameters. Headed by recently hired Inflection co-founder Mustafa Suleyman, it marks Microsoft's first massive internal frontier model since investing over $10 billion in OpenAI. Owning its own heavyweight model gives the tech giant crucial leverage and a strong hedge against OpenAI.
SenseTime just launched SenseNova 5.0, a major model trained on over 10TB of tokens with a 200,000 context window. Alongside the cloud model, it rolled out an edge-side matrix that hits 78.3 words per second on flagship chips and cuts enterprise inference costs by 80%. Why it matters: SenseTime is pushing frontier multimodal AI directly into phones, cars, and edge devices.
Microsoft just launched Phi-3 Mini, a lightweight 3.8B-parameter AI model trained on synthetic bedtime stories. Microsoft claims it rivals models 10 times its size, like GPT-3.5, while running on phones and laptops. It is available now on Azure, Hugging Face, and Ollama, with 7B and 14B versions coming soon.
Meta just launched Llama 3, delivering 8B and 70B open-access models trained on over 15 trillion tokens. The smaller 8B variant can easily run and fine-tune on consumer GPUs down to 4-bit precision, backed by an expanded 128k vocabulary. Why it matters: Meta is resetting the benchmark for open AI, giving developers frontier-level capabilities without proprietary API lock-in.
Grok-1.5V just dropped as a first-generation multimodal model, expanding beyond text to handle documents, diagrams, and real-world photos. The team also released RealWorldQA, a new spatial reasoning benchmark where Grok scored 68.7%, edging out rival frontier models. It will roll out soon to early testers and existing users.
OpenAI just dropped a majorly improved GPT-4 Turbo model in its API and began rolling it out to ChatGPT. GPT-4 Turbo with Vision is now generally available, finally letting developers use JSON mode and function calling directly on visual inputs. Why it matters: Multimodal workflows can now return reliable, structured data straight from images.
OpenAI just dropped the sign-up requirement for ChatGPT, allowing anyone to start chatting instantly without creating an account. Conversations from logged-out sessions can still be used for model training unless opted out in settings. It's the ultimate friction-killer for bringing AI to the next hundred million people.
OpenAI just shared early results from Voice Engine, a model that can clone anyone's voice using just a single 15-second audio sample. While trusted partners are testing it for translation and speech therapy, OpenAI is holding back a wider public release over deepfake risks during an election year.
xAI just introduced Grok-1.5, boasting major upgrades in reasoning and math alongside a 128,000-token context window—a 16x jump over its predecessor. It scored 50.6% on the MATH benchmark and 74.1% on HumanEval. Early testers on 𝕏 will get access in the coming days.
Anthropic’s Claude 3 Opus just beat OpenAI’s GPT-4 on the crowdsourced Chatbot Arena leaderboard for the first time. Variations of GPT-4 have dominated the top spot since May 2023, but Claude 3 has finally broken that year-long reign alongside strong showings from its budget Haiku model. Why it matters: This marks the first time the best available LLM isn't from OpenAI.
OpenAI is testing major upgrades in ChatGPT preview builds, according to developer Tibor Blaho. Upcoming features include ditching the rigid 40-message limit on GPT-4, auto-routing between models based on complexity, anonymous access without an account, and an inpainting editor for DALL-E 3.
Stability AI just rolled out Stable Code Instruct 3B, an instruction-tuned coding LLM that beats CodeLlama 7B despite having less than half the parameters. It handles natural language queries, math, and code generation across popular languages. Why it matters: Developers can run powerful coding assistants on consumer-grade hardware, with weights ready to download on Hugging Face.
