New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Anthropic is prepping a two-way voice mode and memory features for Claude to remember past conversations, CEO Dario Amodei just revealed at Davos. Smarter models are also slated for the coming months, though Amodei admits a recent demand surge has completely overwhelmed their compute capacity as the company seeks another $2B in funding.
DeepSeek just launched DeepSeek-R1, an open-weight reasoning model that matches OpenAI-o1's performance on math, code, and reasoning. It arrives under a permissive MIT license alongside six distilled smaller models, with chat and API access live immediately. Basically, developers can now freely commercialize and distill frontier-level reasoning outputs without restrictions.
Sam Altman just announced that OpenAI has finalized o3-mini and started its rollout process, targeting a launch in about two weeks. Addressing developer demands, the team plans to release the model on both the API and ChatGPT at the exact same time.
NVIDIA just introduced the Llama Nemotron and Cosmos Nemotron model families, custom-built to power autonomous AI agents across enterprises. Available in Nano, Super, and Ultra sizes, the open models use distillation and pruning to run everywhere from edge PCs to multi-GPU clusters. Why it matters: It gives developers purpose-built open building blocks to deploy multi-agent workflows without bleeding compute budgets.
NVIDIA just unveiled Cosmos at CES, a platform of open generative world foundation models designed for physical AI like autonomous vehicles and robots. Built into its three-computer AV pipeline, Cosmos turns thousands of human-driven miles into billions of synthetic driving miles to stress-test edge cases. The models are already available under an open license.
Elon Musk just revealed that xAI's upcoming Grok 3 has finished pretraining, using 10 times more compute than Grok 2. He added that the new flagship model is coming soon. Why it matters: xAI is rapidly scaling its infrastructure to challenge OpenAI and Google at the very top of frontier AI.
Sonus AI just launched its Sonus-1 family of language models, offering Mini, Air, and reasoning-boosted Pro tiers. The flagship Sonus-1 Pro with reasoning claims a 91.8% score on MATH-500 and 90.0% on HumanEval, taking aim at frontier proprietary models across math and code.
OpenAI just capped off its 12 Days of OpenAI event by revealing its next-generation reasoning model, o3, alongside o3 mini. The new models comfortably beat o1 on math, coding, and the ARC-AGI benchmark, featuring adjustable reasoning effort. General users will have to wait, though, as rollout begins in late January.
Microsoft just launched Phi-4, its latest 14-billion-parameter small model focused heavily on math problem-solving. Trained on high-quality synthetic and human datasets with fresh post-training tweaks, it takes aim at lightweight rivals like GPT-4o mini and Gemini 2.0 Flash. For now, it is only available for limited research use on Azure AI Foundry.
Anthropic just rolled out its Claude 3.5 Haiku model to Claude web and mobile users. The lightweight model matches or beats the previous flagship 3 Opus on certain benchmarks, excelling at coding, data extraction, and longer outputs. The only real trade-off is the lack of image analysis.
Grok just rolled out a faster, sharper version of Grok-2 to everyone on X for free. The upgraded model runs three times faster with better accuracy and multilingual support, while bundling web search, citations, and the Aurora image model. Developers also get updated API endpoints at lower token pricing.
Google just introduced its next-gen Gemini 2.0 model family, rolling out Gemini 2.0 Flash Experimental to developers and Gemini web users today. The new model beats 1.5 Pro at twice the speed while bringing native multimodal outputs, tool calling, and a new Deep Research mode. Why it matters: Google is moving fast from basic chatbots to proactive browser-controlling and code-executing agents.
OpenAI just expanded its Reinforcement Fine-Tuning Research Program, offering select developers alpha API access to build domain-specific expert models. The technique reinforces reasoning using verified task-answer pairs, excelling in areas like law, healthcare, and finance where objective right answers exist. A public release is slated for early 2025.
OpenAI just published the full System Card for its o1 and o1-mini reasoning models. By using deliberative alignment during its chain-of-thought, o1 drastically outperforms GPT-4o on jailbreak resistance and reduces hallucinations. The trade-off: smarter models also introduce new risks, meaning OpenAI now has to monitor chain-of-thought deception.
OpenAI’s ChatGPT search tool is unpredictable and frequently inaccurate when identifying news sources, according to researchers at Columbia’s Tow Center. The bot gave partially or entirely incorrect responses to most requests while rarely admitting uncertainty. OpenAI pushed back, calling the study an atypical test and promising ongoing improvements.
Amazon just unveiled Amazon Nova, its new lineup of foundation models spanning text, multimodal reasoning, and image/video generation. Available in Amazon Bedrock, the Micro, Lite, and Pro models claim to be the fastest in their classes while slashing costs by at least 75%.
OpenAI just got caught using a hard-coded filter that halts ChatGPT whenever users mention specific names like Brian Hood or Jonathan Turley. The kill switch was implemented to dodge defamation lawsuits after past hallucinations, though the names still process normally through OpenAI's API.
Cohere just launched Rerank 3.5, a cross-encoder model designed to boost enterprise search and cut hallucinations in RAG pipelines. It tackles complex reasoning across over 100 languages and sets new high marks in finance and ecommerce benchmarks. AWS customers can already deploy it via a single Amazon Bedrock API call.
Amazon just developed a new multimodal AI model code-named Olympus that can process images and videos alongside text, according to The Information. The model aims to let users search video scenes with simple text prompts and lessen AWS's heavy reliance on Anthropic's Claude, despite Amazon recently pumping an additional $4 billion into the startup. An official reveal could drop as early as next week at AWS re:Invent.
Alibaba's Qwen team just released QwQ-32B-Preview, a 32.5B reasoning model licensed under Apache 2.0. Alibaba says it outperforms OpenAI's o1-preview on math benchmarks like AIME and MATH by using inference-time compute, though it can still get caught in loops or stumble on common-sense reasoning.
