New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
xAI just released the API for its flagship Grok 3 and Grok 3 Mini reasoning models. Standard Grok 3 matches Claude 3.7 Sonnet pricing at $3 per million input tokens and $15 per million output tokens, with faster tiers available. Notably, the API maxes out at a 131,072 tokens context window, falling short of the 1 million tokens xAI promised in February.
Google DeepMind just announced it will support the MCP protocol across its Gemini models and SDK. CEO Demis Hassabis called MCP an emerging open standard for the AI agent era, noting Google looks forward to developing it further with the MCP team and the broader industry.
Google just made its Veo 2 video generation model production-ready in the Gemini API, alongside launching the low-latency Live API into public preview. Following the debut of Gemini 2.5 Pro, the team also teased an upcoming 2.5 Flash model that brings adaptive reasoning to its low-cost, fast workhorse. It gives developers an end-to-end multimodal stack from real-time voice agents to video synthesis.
Google Cloud just integrated its Lyria text-to-music model into Vertex AI, making it the first platform covering video, image, speech, and music under one roof. The update also adds camera controls and inpainting to Veo 2, plus 10-second voice cloning in Chirp 3. Enterprise teams can now theoretically generate an entire ad spot from prompts, though most tools remain behind an allowlist.
Nvidia just released Llama-3.1-Nemotron-Ultra-253B-v1, a fully open-source dense model built on Meta's 405B base. By overhauling the architecture to fit on a single 8x H100 node, it edges out the massive 671B DeepSeek R1 on benchmarks like GPQA and LiveCodeBench, though DeepSeek still leads in heavy math.
Meta’s head of GenAI just denied rumors that Llama 4 Maverick and Scout were trained on test sets to game benchmark scores. The speculation flared over the weekend after users noticed performance gaps across platforms. Al-Dahle blamed the mixed quality on early rollout issues across cloud providers, promising bug fixes over the coming days.
Anthropic's Chief Scientist Jared Kaplan just revealed that Claude 4 will arrive "in the next six months or so." Kaplan noted that post-training and RL are compressing AI cycles even faster than hardware upgrades, showing zero signs of a slowdown.
Sam Altman just announced a change of plans at OpenAI: the company will release o3 and o4-mini in a couple of weeks, pushing GPT-5 back by a few months. Altman explained that smoothly integrating everything proved harder than expected, but the delay will make GPT-5 significantly better while securing enough compute capacity for unprecedented demand.
Tencent just rolled out a major upgrade to its Hunyuan T1 deep reasoning model, boosting its project-level code generation and text writing capabilities. The update also sharpens multi-turn comprehension, instruction following, and the handling of mixed Chinese and English text. You can test it out right away on Tencent's platform or Hugging Face.
The Arc Prize Foundation just revised its compute cost estimate for OpenAI's upcoming o3 model, finding that o3 high could cost roughly $30,000 to solve a single ARC-AGI problem—ten times its initial $3,000 guess. The spike reflects intense test-time compute: the model needed 1,024 attempts per task to hit peak accuracy, using 172 times more compute than o3 low. Why it matters: Frontier reasoning gains are real, but brute-forcing complex logic might be economically unfeasible for most early applications.
Google just brought its experimental Gemini 2.5 Pro model to free web users, rolling it out uncharacteristically fast after its paid debut. The model features visible step-by-step reasoning and tops the LMSYS Chatbot Arena leaderboard, though free users miss out on file uploads and get a reduced context window.
Amazon just launched Nova Act, a general-purpose AI agent that takes the wheel in your browser to book reservations, order food, and fill out forms. Developed by its new AGI lab to power the upcoming Alexa+, it takes direct aim at OpenAI's Operator and Anthropic's Computer Use. It's available now as a research preview.
ERNIE 4.5 just swept GPT-4.5 in a three-game chess showdown, winning all three matches decisively. The model demonstrated tighter instruction-following and sharper positional decision-making, even sealing the final match with a 10-move checkmate. It is a reminder that frontier logic tests do not always belong exclusively to OpenAI.
OpenAI just rolled back several image moderation guardrails in ChatGPT, allowing users to generate pictures of public figures like Donald Trump and sensitive symbols in neutral contexts. The company is shifting from blanket refusals toward preventing real-world harm, though strict safeguards around children remain firmly in place.
OpenAI just rolled out an update to GPT-4o in ChatGPT, sharpening its ability to follow complex prompts with multiple requests and handle tricky coding tasks. The model also gets a bump in intuition and creativity—plus OpenAI deliberately tuned it to spit out fewer emojis.
QwenLM just officially launched the first version of QVQ-Max, a dedicated visual reasoning model designed to analyze images and videos. The model goes beyond basic perception to solve complex math and logic tasks, with MathVision benchmark accuracy scaling up as its thinking time increases. Why it matters: QwenLM plans to evolve this foundation into a visual agent that can operate smartphones and computers.
Anthropic is reportedly preparing to boost Claude 3.7 Sonnet's context window from 200K to 500K tokens, according to newly surfaced feature flags. The jump would let developers drop massive codebases into a single session without leaning on RAG. It remains unclear whether the tier will stay exclusive to enterprise users.
Google just released Gemini 2.5 Pro Experimental, its newest flagship model that reasons through thoughts before answering. The model immediately took the #1 spot on LMArena, boasts a 1-million-token context window, and hits new highs on math and coding benchmarks like SWE-bench. It is available right now in Google AI Studio and for Gemini Advanced subscribers.
OpenAI just rolled out native image generation inside GPT-4o, making it the new default engine in ChatGPT. It nails in-image text rendering, handles up to 20 distinct objects at once, and maintains character consistency across ongoing chats. Renders can take up to a minute, but the jump in control is huge.
Alibaba just released Qwen2.5-VL-32B-Instruct, a visual language model fine-tuned with reinforcement learning. It delivers major boosts in math reasoning and human preference alignment, designed as the sweet spot for users who find 72B too heavy and 7B not capable enough for local rigs.
