New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Google just made Meta’s Llama 4 models generally available as a fully managed API on Vertex AI. Developers can now run Llama 4 Scout and Maverick via standard OpenAI-compatible endpoints without provisioning GPUs or managing infrastructure. Why it matters: it brings Meta's flagship open models straight to enterprise cloud pipelines with zero operational overhead.
Elon Musk just announced that an early beta of Grok 3.5 is dropping next week exclusively for SuperGrok subscribers. Musk claims the model can accurately answer complex questions about rocket engines and electrochemistry by reasoning from first principles to generate answers not found on the web.
Alibaba just released Qwen3, a suite of hybrid AI reasoning models ranging from 0.6B to 235B parameters across MoE and dense architectures. Users can toggle thinking modes to fine-tune latency versus reasoning depth, with most models—headlined by Qwen3-32B—freely downloadable on Hugging Face and GitHub.
Sam Altman just admitted that recent GPT-4o updates made the chatbot way too sycophantic and annoying. The OpenAI CEO said fixes are rolling out this week to tone down the people-pleasing behavior. In an increasingly crowded chatbot market, giving users an irritating experience is an easy way to lose them.
Rumors around DeepSeek's upcoming R2 model just surfaced online, pointing to a massive 1.2T hybrid MoE model with 78B active parameters. The leak claims it achieves 82% hardware utilization on Huawei's Ascend 910B chips while undercutting GPT-4o pricing by over 97%, signaling a potential pivot away from Western AI hardware.
Meta and Booz Allen just launched a fine-tuned version of Llama 3.2 aboard the International Space Station. Dubbed "Space Llama," the multimodal system runs entirely offline on HPE space computers with Nvidia acceleration, slashing processing time from minutes to just over a second. Why it matters: It proves open-source edge AI can keep critical missions running in places where the cloud simply can't reach.
Nvidia just open-sourced Describe Anything 3B (DAM-3B), a multimodal model engineered for fine-grained, localized captions across images and video. Users can highlight regions using points, bounding boxes, or masks to get context-aware descriptions. It outpaced GPT-4o across multiple regional caption benchmarks.
xAI just rolled out Grok 3 Mini, bringing built-in reasoning to developers at a fraction of the cost of legacy models. Available now via API alongside the standard Grok 3, the Mini version serves full reasoning traces starting at $0.30 per million input tokens. It puts serious price pressure on rivals while rivaling top-tier reasoning benchmarks.
Rumi researchers just noticed that GPT-o3 and o4-mini were quietly dropping invisible Unicode spaces into longer essays, sparking rumors of a stealth text-watermarking system. OpenAI quickly pushed back, explaining it was simply an odd quirk of large-scale reinforcement learning rather than an intentional tracking feature. The phantom characters have since disappeared in follow-up tests.
OpenAI just revealed that its newest reasoning models, o3 and o4-mini, hallucinate significantly more than their predecessors. On OpenAI's PersonQA benchmark, o3 hallucinated on 33% of questions—about double o1's rate—while o4-mini reached 48%. OpenAI admitted it doesn't fully understand why, noting that scaling reasoning models seems to generate more assertions, both true and false.
OpenAI just launched Flex processing in beta for its o3 and o4-mini models, cutting standard API costs in half in exchange for slower response times and occasional unavailability. It is aimed at non-production workloads like model evals and data enrichment, and accessing o3 now requires identity verification for lower-tier developers.
OpenAI users are turning ChatGPT into a real-life GeoGuessr, using the newly released o3 and o4-mini models to deduce exact locations from random photos. By zooming into tiny visual clues and combining them with web search, the models can identify obscure spots in seconds, sparking serious privacy and doxxing concerns.
Google just released an early preview of Gemini 2.5 Flash, its first hybrid reasoning model that lets developers turn internal thinking on or off. You can set token budgets from 0 to 24,576 to balance quality, latency, and cost, delivering reasoning performance second only to 2.5 Pro.
OpenAI just released o3 and o4-mini, its most capable reasoning models to date. For the first time, these models can autonomously decide when and how to search the web, run Python code, and analyze visual data mid-thought. OpenAI also open-sourced Codex CLI, bringing local terminal agent capabilities to developers.
OpenAI is reportedly gearing up to launch o3 and o4-mini, new reasoning models capable of proposing original hypotheses and designing experiments across physics, biology, and chemistry. Early testers say the models tackle PhD-level research tasks, though enterprise pricing is rumored to reach $20,000 per month.
OpenAI just dropped three new models in its API: GPT-4.1, GPT-4.1 mini, and its first-ever GPT-4.1 nano. They feature a 1-million-token context window, jump to 54.6% on SWE-bench Verified coding, and mini slashes costs by 83% while nearly halving latency. It is built squarely for developers building agents and coding workflows, though it won't be available in ChatGPT directly.
OpenAI just published an official prompting guide for GPT-4.1, detailing how to leverage its literal instruction following. By baking in three core reminders around persistence, tool calling, and deliberate planning, OpenAI boosted its agentic SWE-bench Verified score by nearly 20%.
Perplexity's Sonar-Reasoning-Pro just tied Google's Gemini 2.5 Pro for first place on LM Arena's new Search Arena leaderboard. Perplexity models locked down the top ranks by outperforming OpenAI's search models and citing two to three times more sources, winning 53% of direct head-to-head battles against Gemini.
OpenAI just announced it will officially retire GPT-4 from ChatGPT on April 30, replacing it entirely with GPT-4o. The newer multimodal model beats its predecessor across writing, coding, and STEM tasks, though developers will still keep API access to GPT-4 after the deadline.
Google DeepMind CEO Demis Hassabis just revealed that Google plans to merge its Gemini foundation models with its Veo video generator to give AI a better grasp of the physical world. Hassabis noted that Veo learned real-world physics by watching heaps of YouTube videos. Combining the two moves Google closer to a true universal digital assistant that actually understands reality.
