New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Anthropic just launched Claude 3.7 Sonnet, introducing an extended thinking mode that users can toggle on demand and developers can control via a custom thinking budget. The model now exposes its raw reasoning process in real time so users can audit how it arrives at answers. On agentic tasks, Claude even managed to navigate Pokémon Red and beat three gym leaders completely autonomously.
Microsoft is reportedly prepping server capacity to host OpenAI’s next frontier models, expecting GPT-4.5 as early as next week and GPT-5 in late May. GPT-5 is set to integrate OpenAI's o3 reasoning model, basically unifying the confusing GPT and o-series pickers into a single system just in time for Build.
Elon Musk just made xAI's flagship Grok 3 model temporarily free for everyone, no subscription required. Users can test drive the new "Think" reasoning and "DeepSearch" modes across the web and app, though paying tiers still keep server priority, voice mode, and the extra-heavy Big Brain features.
Elon Musk's new Grok 3 model might top benchmark leaderboards, but security firm Adversa AI just found it failed three out of four basic jailbreak tests and leaked its entire system prompt. Why it matters: as vulnerable models power autonomous AI agents, simple prompt injections could turn into real-world corporate security breaches.
Microsoft Research and Xbox studio Ninja Theory just introduced Muse, a generative "world and human action model" trained on Bleeding Edge and published in Nature. Muse understands 3D physics and controller inputs to render playable gameplay directly. Microsoft hopes to use it to revive nostalgic games trapped on legacy hardware and speed up prototyping. Weights and an interactive demo are live on Azure AI Foundry.
Google just released PaliGemma 2 mix checkpoints, tuning its vision-language model across a mixture of tasks for instant out-of-the-box use. It handles OCR, captioning, and segmentation based on prompt syntax, and anyone already using original PaliGemma mix checkpoints can upgrade directly without changing their setup.
Elon Musk’s xAI just released its latest flagship Grok 3 family, introducing specialized reasoning models to compete with o3-mini alongside a new DeepSearch tool. Trained on roughly 200,000 GPUs, it claims top benchmark scores and offers a high-compute Big Brain mode. Access is currently rolling out to X Premium+ subscribers.
Mistral just introduced Mistral Saba, a 24-billion-parameter model tailored for Arabic-speaking countries. According to internal tests, it handles Arabic far better than Mistral Small 3 and also performs well with Indian languages like Tamil. It is accessible through Mistral's API and supports on-premise deployment.
Elon Musk's xAI just rolled out Grok 3, trained on roughly 200,000 GPUs with 10x more compute than its predecessor. The lineup includes new reasoning models designed to rival OpenAI's o3-mini on tough math problems, alongside a web research tool called DeepSearch. It is rolling out first to X's Premium+ subscribers.
OpenAI just removed the nagging orange content warnings from ChatGPT's interface to cut down on gratuitous refusals. While hard guardrails against harm and illegal acts remain, the company also updated its Model Spec so models won't shy away from sensitive viewpoints. Why it matters: It directly addresses long-standing complaints about AI over-censorship.
Anthropic is reportedly gearing up to launch its next flagship AI model within weeks. Described as a hybrid system, it can alternate between deep reasoning and instant replies, featuring a sliding scale for developers to balance compute costs. The new model allegedly beats OpenAI's o3-mini-high on several coding benchmarks.
Sam Altman just outlined OpenAI’s plan to ditch the messy model picker once and for all. The company will next launch GPT-4.5 as its final non-chain-of-thought model, before rolling out a unified GPT-5 system that bakes in o3 and all flagship tools. Most notably, free ChatGPT users will get unlimited access at standard intelligence.
Perplexity just rolled out a new version of its in-house Sonar model, powered by Cerebras chips to generate answers at a blazing 1,200 tokens per second. Built on Llama 3.3 70B, it beats GPT-4o mini and Claude 3.5 Sonnet in user satisfaction testing while answering almost instantly. It is now live for Pro subscribers.
OpenAI just updated the chain of thought for o3-mini, revealing far more reasoning steps to both free and paid ChatGPT users. While it still holds back raw internal monologues to prevent distillation, a new post-processing layer cleans up unsafe output and translates the thought process into non-English native languages.
Kyutai just introduced Hibiki, a simultaneous speech-to-speech translation model currently handling French to English. It delivers real-time voice and text translations while keeping the speaker's original timbre and pacing naturally. Evaluations show it rivals professional interpreters in naturalness and accuracy.
Andrej Karpathy just released a massive 3.5-hour video breaking down the entire LLM training pipeline for everyday users. The tutorial walks through pretraining, supervised finetuning, and reinforcement learning, covering everything from tokenization to model hallucinations. It is basically the ultimate free, intuitive masterclass on modern AI.
Google just made Gemini 2.0 Flash generally available to everyone and introduced an experimental Gemini 2.0 Pro packing a massive 2M-token context window for heavy coding and reasoning tasks. The lineup also adds Flash-Lite, offering 1.5 Flash performance at cut-rate pricing. Gemini Advanced users can already pick the Pro model from the app switcher.
Anthropic just rolled out Citations for the Claude API, allowing Claude 3.5 Sonnet and Haiku to trace claims back to exact sentences in source documents. It boosts recall accuracy by up to 15%, removes the need for brittle prompt engineering, and developers aren't charged for the quoted output tokens.
OpenAI CEO Sam Altman just confirmed that the company's upcoming lightweight reasoning model, o3-mini, will be available to all ChatGPT users, including the free tier. Paid Plus subscribers will receive generous usage allocations. The model focuses on faster and cheaper math and coding reasoning, though a specific launch date has not been set.
Google just updated its Gemini 2.0 Flash Thinking model, hitting 73.3% on math benchmark AIME and 74.2% on GPQA Diamond. The latest release introduces code execution, a 1M token context window, and reduces contradictions between reasoning steps and final answers. Why it matters: Google is moving fast on reasoning models, combining long-time planning research with frontier LLMs.
