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.
OpenAI just rolled out its compute-heavy o1-pro model to the API for select Tier 1–5 developers. It burns more compute to deliver sharper reasoning responses, while adding vision, function calling, and structured outputs. The catch is the eye-watering price: $150 per million input tokens and $600 per million output tokens.
xAI just rolled out image generation capabilities in its developer API, starting with a single model called grok-2-image-1212. It outputs up to 10 JPGs per request at $0.07 per image, sitting right between FLUX and Ideogram pricing. The catch: you can't customize size or style yet, and prompts are automatically revised by a chat model.
Perplexity just rolled out major upgrades to its Sonar models, adding High, Medium, and Low search modes so developers can balance search depth with compute costs. The company is also waiving fees for citation tokens across responses. Internal benchmarks show Sonar outperforming search-enabled GPT-4o at a fraction of the price.
Google just opened native image generation in Gemini 2.0 Flash to all developers via Google AI Studio and its API. Built from the ground up to output images natively, the model handles multi-turn conversational editing, preserves character consistency across prompts, and renders in-image text far better than typical diffusion models.
Sam Altman just revealed that OpenAI trained a new model specifically excelling at creative writing, sharing a metafictional short story about AI and grief. Altman noted it was the first time AI writing had genuinely moved him. Why it matters: LLMs often produce bland prose, but a model tuned for literary nuance could change creative workflows. No release timeline has been announced yet.
ElevenLabs just cut speech-to-text pricing by 45% at scale and made its Scribe model completely free via the UI until April 9. The model boasts high accuracy across 99 languages. It is an aggressive move to undercut rivals and attract developers needing high-volume transcription.
Tencent Hunyuan just unveiled Hunyuan-TurboS, an ultra-large MoE model blending Transformer and Mamba architectures to ease long-context training and KV-cache overhead. The team claims it tops GPT-4o and DeepSeek-V3 in math and reasoning, all while cutting inference costs to one-seventh of its predecessor.
Meta is reportedly gearing up to launch Llama 4 in just a few weeks, putting a major focus on native voice capabilities. Meta CPO Chris Cox confirmed it will be an "omni" model handling speech and text natively, allowing users to interrupt mid-sentence like ChatGPT's Voice Mode. The rush comes as DeepSeek's open-source gains pushed Meta into overdrive.
Google just launched an experimental text embedding model, gemini-embedding-exp-03-07, through the Gemini API. Trained directly on Gemini, it took the top spot on the MTEB Multilingual leaderboard with a score of 68.32, beating the runner-up by 5.81 points. Developers can try it out immediately using the existing embed_content endpoint.
Tavus just launched three AI models—Phoenix-3, Raven-0, and Sparrow-0—designed to make video avatars feel human. The trio tackles micro-expressions, visual perception, and conversational turn-taking, cutting awkward dialogue pauses. They are now available via API, backed by a demo agent named Charlie that talks, sees, and searches.
Qwen just launched QwQ, a mid-sized reasoning model that matches heavyweight frontier systems like DeepSeek-R1. Alibaba found that scaling reinforcement learning yields massive leaps in math and coding, proving mid-sized models can punch way above their weight against massive MoE giants.
OpenAI just began rolling out GPT-4.5, its largest model to date, to ChatGPT Plus subscribers over the next one to three days. It promises richer world knowledge, fewer hallucinations, and better conversational flair, though Sam Altman openly warned it is not a dedicated reasoning model. It also comes with a hefty price tag—roughly 15 times more expensive per token than GPT-4o.
OpenAI just released a research preview of GPT-4.5, its largest and most capable chat model to date. Built by scaling unsupervised learning and compute on Azure supercomputers, the model delivers broader world knowledge, reduced hallucinations, and better conversational nuance. ChatGPT Pro users and API developers can test it right now.
Microsoft just introduced Phi-4-multimodal and Phi-4-mini, a pair of lightweight models built for edge hardware and Copilot+ PCs. The 5.6B multimodal variant natively ingests speech, vision, and text in a single architecture, topping open speech recognition leaderboards while staying compact enough to run smoothly on-device.
Inception Labs just launched Mercury Coder, the first commercial-scale diffusion language model that generates entire code sequences in parallel rather than predicting tokens sequentially. Running on standard Nvidia H100 GPUs, it hits over 1,000 tokens per second while matching models like GPT-4o Mini. Why it matters: It proves text can borrow from image generation to slash AI inference latency and costs tenfold without specialized hardware.
xAI just rolled out an uncensored voice mode for Grok 3 subscribers that goes where ChatGPT won't. Users can toggle personalities like an "unhinged" mode that curses and screams before hanging up, or an 18+ "Sexy" mode built for phone sex. Why it matters: While OpenAI keeps tight guardrails on voice AI, xAI is doing the exact opposite by letting its model run completely wild.
OpenAI just published the system card for its agentic Deep research feature, breaking down the safety guardrails built into the tool before its launch to Pro users. Powered by an early version of OpenAI o3, the model scored a medium rating across four frontier risk areas, with safety teams focusing heavily on web-based prompt injections and data privacy.
Google just launched Gemini 2.0 Flash-Lite for general availability across the Gemini API, Google AI Studio, and Vertex AI. The lightweight model beats 1.5 Flash on reasoning and math benchmarks while offering cheaper long-context pricing at $0.10 per million input tokens.
DeepSeek is accelerating the rollout of R2, the successor to its hit R1 reasoning model, according to people familiar with the matter. The startup originally targeted an early May release but now wants it out as soon as possible, aiming to deliver better coding skills and broader multilingual reasoning. Why it matters: If DeepSeek strikes twice on a budget, it will put even more pressure on Western rivals racing to keep up.
Qwen just unveiled QwQ-Max-Preview, a preview of its upcoming reasoning model built on Qwen2.5-Max that targets math, coding, and autonomous agents. The team also pledged to open-source the full QwQ-Max under Apache 2.0, along with a smaller QwQ-32B designed for local device deployment.
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.
