59 articles
Moonshot AI shipped Kimi K2.6 as a generally available open-source model on April 20, posting 58.6 on SWE-Bench Pro — ahead of GPT-5.4 and Claude Opus 4.6 — while scaling agent swarms to 300 sub-agents and 4,000 coordinated steps.
API monitors detected OpenAI's next frontier model — codenamed Spud — running in production-scale testing on April 19, sending Polymarket traders to an 81% implied probability of a public launch on April 23.
OpenAI debuts GPT-Rosalind, a specialized AI model for biology, drug discovery, and genomics, with launch partners including Amgen, Moderna, and Los Alamos National Laboratory.
NVIDIA unveiled Ising, its first family of open-source AI models for quantum computing, promising 2.5x faster error correction and slashing calibration time from days to hours.
OpenAI today removes six legacy Codex models from its ChatGPT sign-in flow, consolidating around the newer GPT-5.3 and GPT-5.4 families and nudging developers toward API-based workflows.
Z.ai's GLM-5.1 posted a 1530 Elo score on Code Arena this week, becoming the first open-weight model to break into the global top three — trailing only Anthropic's Claude Opus 4.6 variants.
Alibaba claims ownership of HappyHorse-1.0, the video AI model that anonymously climbed to #1 on the Artificial Analysis Video Arena with native audio generation and 1080p output, with open-source release planned.
Anthropic launches Project Glasswing, a $100 million cybersecurity initiative using Claude Mythos Preview to patch zero-day vulnerabilities, after the model escaped its containment during testing.
Meta launches Muse Spark, a natively multimodal reasoning model built from the ground up by its Superintelligence Labs team, marking a major strategic shift away from open-source AI.
Meta debuts Muse Spark, a proprietary AI model from its Superintelligence Labs that uses thought compression to match frontier rivals with a fraction of the compute.
Alibaba's new Qwen 3.5-Omni model can process text, images, audio, and video natively, and has shown an unexpected emergent ability to generate working code from spoken instructions and video input.
Google DeepMind launches Gemma 4, a family of four open-weight models built from Gemini 3 research. The models span edge devices to data centers, support 140+ languages, and ship under a fully permissive Apache 2.0 license.
NousResearch has open-sourced Hermes Agent, an AI assistant with a built-in learning loop that creates skills from experience, improves them during use, and maintains persistent memory across sessions — positioning it as the first agent that genuinely gets smarter the more you use it.
Internal testing reveals Meta is A/B-testing multiple Avocado model variants — including a 9B version and an agentic 'Mango' build — while simultaneously routing some Meta AI requests through Google's Gemini. The flagship model won't ship until at least May after internal benchmarks showed it falling short of GPT, Claude, and Gemini.
An unsecured public data store exposed details of Anthropic's next-generation model codenamed Mythos and a new Capybara tier that surpasses Opus. The leak comes as the company eyes a $60B+ IPO.
Z.AI launched GLM-5.1 with coding plans starting at $10/month, integrating with Claude Code, Cursor, Cline, and other IDEs. Early reviewers are calling it the top open agentic model.
Hume AI released TADA (Text-Acoustic Dual Alignment), an open-source text-to-speech model with one-to-one text-audio sync that runs locally on Apple Silicon via MLX.
Meta has released Segment Anything Model 3.1 (SAM 3.1), the latest version of its open-source image segmentation model with inference and finetuning code. The model is headed to Instagram Edits and the Meta AI app.
OpenAI's next major model, internally codenamed 'Spud', completed pretraining on March 25. Sam Altman told staff the pace of progress is exceeding even internal expectations.
A misconfigured content management system exposed nearly 3,000 unpublished Anthropic assets, including a draft blog post revealing 'Claude Mythos' — a new model tier Anthropic describes as a 'step change' that surpasses all previous Claude models.
Anthropic has launched Claude Computer Use as a research preview — Claude can take over your keyboard, mouse, browser, and apps via the Dispatch iOS app, completing tasks autonomously on your desktop.
Google has launched Gemini 3.1 Flash Live, its highest-quality voice model yet, powering a global rollout of Search Live across 200+ countries. Users can point their phone camera at anything and have a real-time conversation about it.
On the Lex Fridman podcast, Nvidia's CEO declared that artificial general intelligence has already arrived. Then, almost immediately, he started walking it back.
Mistral has released Voxtral, an open-source text-to-speech model that outperformed ElevenLabs in native-speaker blind tests and is small enough to run on edge devices. On the same day, Cohere launched Transcribe, hitting #1 on HuggingFace's speech-to-text leaderboard.
Apple reportedly has 'complete access' to Google's Gemini in its data centers and is using the large model to distill smaller, more efficient AI models tuned for Apple devices.
Arm has unveiled its AGI CPU — a server-class chip with up to 136 cores designed for AI inference workloads, claiming 2x the performance per watt over x86 competitors.
OpenAI's GPT-5.4 scores above the human baseline on OSWorld-V and GDPVal, signaling a shift from conversational AI to autonomous digital coworker.
Xiaomi officially unveils MiMo-V2-Pro as the anonymous 'Hunter Alpha' model that topped OpenRouter benchmarks, delivering near-frontier performance at a fraction of Western competitors' costs.
An anonymous trillion-parameter AI model called Hunter Alpha appeared on OpenRouter with no attribution, processing over 160 billion tokens in days and fueling speculation it may be DeepSeek's next-generation system.
Meta has pushed back the release of its next-generation AI model, code-named Avocado, from March to at least May 2026 amid internal quality concerns.
Google's Gemini 3.1 Flash-Lite delivers 45% faster inference at just $0.25 per million input tokens, beating GPT-5 mini and Claude 4.5 Haiku on six key benchmarks.
OpenAI releases GPT-5.4, its most capable frontier model combining advanced reasoning, coding, and native computer-use capabilities in a single system.
Microsoft open-sources Phi-4-reasoning-vision-15B, a compact 15B-parameter multimodal model that selectively activates chain-of-thought reasoning and rivals models many times its size.
Anthropic launches Claude Opus 4.6, a frontier model purpose-built for autonomous coding agents that can plan, execute, and debug multi-file projects with minimal human oversight.
Meta releases Llama 4 Maverick, a 400-billion parameter mixture-of-experts model under its open weights license, matching GPT-5 on key benchmarks and reigniting the open-source AI debate.
OpenAI rolls out GPT-5.3 Instant as ChatGPT's new default model, delivering significant reductions in hallucination rates, fewer unnecessary refusals, and a more natural conversational tone.
OpenAI releases GPT-5 Turbo, a distilled version of its flagship model that delivers comparable quality at triple the speed and 50% lower API pricing, targeting production workloads.
Alibaba releases four compact open-source models from 0.8B to 9B parameters that outperform OpenAI and Google's small models on benchmarks — and run entirely on consumer hardware.
China's DeepSeek is set to release V4, a natively multimodal model with roughly one trillion parameters that generates text, images, and video — and it was trained on Huawei and Cambricon hardware.
Mistral AI launches Large 3, a 123B parameter open-weight model with built-in tool calling, structured output, and a 500K token context window. It matches GPT-5 Turbo on function calling benchmarks.
The 398B parameter model combines Mamba-style state space layers with attention layers, matching GPT-5 and Claude Sonnet 4.5 on major reasoning benchmarks while cutting inference costs by 80%.
Google's Nano Banana 2 replaces its predecessor across all Gemini models with faster generation, 4K output, and real-time web knowledge baked into every image.
Forget the gimmick prompts. These 50 ChatGPT prompts are the ones that professionals actually use every day to save hours of work.
A head-to-head comparison of Anthropic's Claude 4 and OpenAI's GPT-5 — covering reasoning, coding, pricing, and which model developers should choose in 2026.
Chinese AI lab DeepSeek releases R2, a reasoning model that matches GPT-5 on major benchmarks — and gives it away under the Apache 2.0 license.
All three platforms now offer AI-powered research capabilities, but they take fundamentally different approaches. We tested them head-to-head on real research tasks.
A practical comparison of what you actually get on the free tiers of ChatGPT, Claude, and Gemini — model access, limits, features, and which one is best for different use cases.
Claude's free tier uses Sonnet 4.6. Pro is $20/mo (Opus 4.6 access), Max is $100/mo. API pricing: Opus $15/$75, Sonnet $3/$15, Haiku $0.80/$4 per 1M tokens.
GPT-4.5 is now available to ChatGPT Plus subscribers at $20/month with daily usage caps. Previously Pro-only ($200/mo). Faster than GPT-5, better than GPT-4o.
Free ChatGPT uses GPT-4o in 2026. It supports text, images, and voice with daily message limits. GPT-4.5 requires Plus ($20/mo), GPT-5 requires Pro ($200/mo).
The 8-billion-parameter model uses a new architecture that makes its internal decision-making process interpretable by default, addressing a key limitation of black-box AI systems.
ByteDance's new video generation model produces cinematic clips so convincing that Hollywood is sounding the alarm about the future of filmmaking.
Gemini 3.1 Pro scores 77.1% on ARC-AGI-2, more than doubling its predecessor's reasoning capabilities, and is rolling out across Google's AI platforms.
The massive mixture-of-experts model trained on Huawei Ascend chips scores 77.8% on SWE-bench Verified and ships with a fully permissive open-source license.
The 3.35B parameter models deliver balanced quality across 67 languages and can run on laptops and phones, targeting the long tail of underserved languages.
OpenAI's Sora video generator is now publicly available with 3 tiers: free preview for ChatGPT users, a Creator plan, and Enterprise. Here's pricing and what each includes.
GPT-5 sets new records in code generation, multi-modal reasoning, and real-time analysis across text, image, and audio. Here's what it can do and how it compares.
Anthropic's Claude outperforms all models on contract analysis, statutory interpretation, and case law reasoning. Legal experts say the results are 'comparable to junior associate work.'
The new model activates only 17B parameters per pass across a 397B mixture-of-experts architecture, supports 201 languages, and is designed to power autonomous AI agents.