Gemini 4 Hydrogen: $1.50 per input token, faster execution
50 per million input tokens. The model outperforms Argon on benchmarks while completing tasks faster across coding and agentic workflows.
Model News
Model News tracks new releases, price changes and policy shifts from the labs that build foundation models, such as OpenAI, Anthropic and Google. Every story starts from the primary announcement and ends with what actually changes for the people who use these models.
Edited by Mohamed Aminar · 97 stories · RSS
Nov 12, 2026
Gemini Decisions API: $0.10 per million input tokens
Google launches Gemini Decisions API for fast multimodal classification at $0.10 per million input tokens with zero output charges. Processes up to 1,000 questions in...
Fish Audio Live 3: Full-duplex voice model with 186ms latency
Fish Audio Live 3 listens and speaks simultaneously for real-time voice agents. Full-duplex capability with 186ms response time; $0.035/min through Dec 31, 2026.
Meta Muse Ember: Haiku-level reasoning at one-third the cost
Meta Muse Ember costs $0.10 per million input tokens. The model matches Claude Haiku 5.5 performance while reducing per-task costs by 66%, reshaping agent economics.
Claude Fable 5.2 and Mythos 5.2: 22% science gains, 20% cheaper
Anthropic's Claude Fable 5.2 and Mythos 5.2 deliver improvements in coding and scientific reasoning. Fable 5.2 is open to all; Mythos 5.2 is restricted to verified...
NVIDIA Cosmos 3 Nano adds four interactive simulation checkpoints
NVIDIA releases four open-weight Cosmos 3 Nano checkpoints for interactive world simulation in robotics and autonomous driving applications.
ERNIE 5.5 Flash Base 280B: Baidu sparse MoE model
Baidu released ERNIE 5.5 Flash Base 280B, a sparse mixture-of-experts model with 14B active parameters per token for agent workflows at lower compute cost.
Ripple-65B reaches 80.52 on TerminalBench 2.1
Ripple-65B Preview achieves 80.52 on TerminalBench 2.1, the top score among US open-weight coding models. Ricursive Intelligence trained it for $32 million under...
Inkling 2 Small 594B MoE: 17B Active Parameters
Thinking Machines releases Inkling 2 Small 594B MoE with 17B active parameters for multimodal agentic work. Scores 31 on Artificial Analysis Intelligence Index.
Gemini Omni 1.2 Flash: 30% price cut, 15-second video
Google cuts Gemini Omni 1.2 Flash pricing 30% to $0.07 per second and extends native video generation to 15 seconds with multi-turn editing.
Phonon TTS: On-device model tops Hugging Face leaderboard
Phonon TTS ranks first in German and Spanish intelligibility on Hugging Face Open TTS Leaderboard. Small on-device speech synthesis model.
StateBench: Language model state tracking benchmark
StateBench benchmark tests how well AI models track state changes across long documents, revealing frontier models struggle on harder problems.
Anthropic Message Threads reduces latency 35% for long conversations
Message Threads beta reduces time-to-first-token by 35% for long agent sessions by transmitting only new messages instead of full chat history, lowering latency...
Prime Mega MoE speeds up mixture-of-experts training 34%
Prime Intellect's trainable MoE kernel boosts full training steps by up to 34% on Qwen3 models, with forward passes 3.11x faster than PyTorch baselines.
Drex Max v1 27B decision model runs locally on Apple Silicon
Nace AI open-sources Drex Max v1, a 27B decision model that runs on Apple Silicon Macs without cloud dependency. Scores 67.52 on Decision Index 0.4.
GPT-5.6 Sol completes Factorio in 13.5 hours
GPT-5.6 Sol running in Codex completed a peaceful Factorio game in 13.5 hours, demonstrating frontier model reasoning on complex strategy tasks.
Audio8-TTS-Realistic: 0.6B voice cloning model, 23 languages
Edge0 releases Audio8-TTS-Realistic, a 601M-parameter text-to-speech model with 23-language support and Apache 2.0 licensing for zero-shot voice cloning.
Sterling finance research agent outperforms Claude Fable 5.1
Samaya AI's Sterling scores 58.9% on FrontierFinance versus Claude Fable 5.1's 55.9% while costing 11x less at $0.61 per question.
GPT-6.1 Terra: $1 input, $5 output for OpenAI's mid-tier model
GPT-6.1 Terra halves OpenAI's mid-tier input price to $1 per million tokens and cuts hallucinated guesses from 88% to 38%. What it costs and where it fits.
Gemini 4 Carbon: Google's top-ranked model, defenders first
Gemini 4 Carbon tops the Artificial Analysis Intelligence Index at 59, but Google is releasing it to cyber defenders first. What changed and what it will cost.
joey-1 decision model: $0.01/M tokens, 20ms latency
Baseten's joey-1 decision model costs $0.01 per million tokens with 20ms latency. The 2B model ranks top under 4B parameters and is 2.5x faster than elephant-1.
AssemblyAI-TTS-1: $15 per 1M characters for voice agents
AssemblyAI launches text-to-speech API at $15 per million characters with unlimited concurrency. The TTS model targets live voice agents with sub-100ms latency.
FlashMap flow map speculative decoding reaches 6.14x speedup
FlashMap speculative decoding achieves 6.14x speedup on Qwen3-8B by drafting 16 tokens jointly in one pass, outperforming diffusion baselines across benchmarks.
Genie 4 adds real-time audio and physics to world models
Google DeepMind's Genie 4 world model generates interactive 3D simulations in real-time with native audio, physics, and natural human motion for research and development.
LFM3: Six specialized 2B models on one backbone
Liquid AI releases LFM3 multimodal model family with text, vision, audio, and decision variants. LFM3-1.7B uses 35% fewer parameters than predecessor while matching...
Mirendil Fern 0.1: 10M-token context coding model
Mirendil releases Fern 0.1 with 10M-token context for code analysis. Hosted early access only; strong retrieval but weak agentic coding versus frontier models.
d3-ultra 27B multimodal decision model scores 87.2
vLLM Semantic Router releases d3-ultra, a 26.09B open-weight multimodal decision model for classification and routing tasks across text, images, and video.
Drex for Search scores 92.05% on BrowseComp-Plus
Nace AI releases Drex for Search, a decision model for agent workflows scoring 92.05% on BrowseComp-Plus with 16.6x fewer search calls than competitors.
GPT-6.1 Sol and Claude Opus 5.5: 75% price cuts reshape LLM market
OpenAI and Anthropic slashed LLM prices dramatically as open-weight models and inference provider fragmentation drive commodity-like competition in 2026.
AKV cache engineering cuts agent costs 27.7%
Tsinghua's AKV system lets LLM agents drop and reposition KV cache instead of summarizing context, cutting costs 27.7% while improving accuracy by 1.7 points.
Nanbeige4.3-33B scores 80.9 on Terminal-Bench
Nanbeige4.3-33B is a 33B open-weight coding model with 512K context and Apache 2.0 license. It scores 80.9 on Terminal-Bench 2.1, offering a local alternative to...
Carbon-X: 14.8B open DNA model scores 66.7% on benchmarks
Hugging Face releases Carbon-X, a 14.8B open DNA language model that reads raw DNA sequences and answers in English, scoring 66.7% on DNA benchmarks.
OpenMobile-2 smartphone agents match proprietary rivals
OpenMobile-2 trains AI smartphone agents on 110+ simulated apps to rival closed models. The 27B model scores 59.6 on real Android devices using SPA-Bench.
Kyutai CADRE video autoencoders cut world model lag to 185ms
Kyutai CADRE video autoencoders enable interactive world models to run at 20 fps with 185ms latency on one H100 GPU. Open-source code and weights released.
Block Dessa speech model processes raw audio to detect emotions
Block's Dessa speech model processes raw audio instead of text, letting it detect emotions, sarcasm, and pauses.
Kimi K3.1 Pro benchmarks show 3D coding strength
Moonshot AI's Kimi K3.1 Pro scored 92.7% on BenchCAD 3D modeling and outperforms Claude Opus 5.5 on financial analysis. Available now via cloud services.
Canva Craft image model ranks 4th, beats Google and OpenAI rivals
Canva launches Canva Craft image model, its first proprietary tool. The model generates 2K images in 55 languages and ranks fourth on DesignArena, challenging OpenAI,...
Google Agent Compute Engine: microVM isolation, zero credentials
Google Agent Compute Engine isolates each Gemini agent in a hardware-isolated microVM with zero embedded credentials. Customer preview targeted for coming months.
Agent Skills standard donated to Agentic AI Foundation
Anthropic donates Agent Skills open standard to the Agentic AI Foundation. Over 70 platforms including ChatGPT and GitHub Copilot have adopted it.
Granite Speech 5.0 2B ranks #2 on Open ASR Leaderboard
68% average WER and ranks second on the Open ASR Leaderboard, offering teams an open-weight alternative for transcription and translation.
Recursive Flint 40B MoE coding model runs 125+ tokens/sec locally
Recursive Flint 40B coding model delivers 2.5B active parameters and 125+ tokens/sec on consumer hardware, competing with closed models for local agent development.
DeepInfer: DeepSeek open-sources inference engine
DeepSeek open-sources DeepInfer inference framework under Apache 2.0. The engine powers V4.1-Flash and cuts compute costs for teams running AI models at scale.
Mercury 3 diffusion LLM reaches 1,068 tokens per second
Inception's Mercury 3 diffusion LLM hits 1,068 tokens per second with 1M context window, leading 8 of 12 benchmarks against GPT-6 Luna and Claude Haiku 5.5. Launch...
Drex Max: $0.06/M tokens, 67.59 on Decision Index
06 per million tokens, optimizing price and speed for computer-use agents that analyze screenshots and images.
MiniMax M3.1 Flash: $0.15/M tokens cuts agent costs 70%
MiniMax M3.1 Flash pricing at $0.15 per million input tokens cuts agentic task costs by 70% versus M3. The 428B MoE model runs three times faster.
Harness2 framework improves agent performance 11.5 points
Google's Harness2 recursive agent framework optimizes LLM agent behavior at test time without retraining. 5 points on WorkBuddy-Bench.
Replit Switches to OpenAI as Default Model Provider
Replit CEO Amjad Masad announced the platform is shifting OpenAI to its default model, replacing Anthropic based on performance, cost, and partnership strength.
DeepSeek V4.1-Flash-1017 outperforms larger models, cuts costs
DeepSeek-V4.1-Flash-1017 outperforms Claude Opus 5 and GPT 5.6-Sol on coding benchmarks while reducing per-task costs to $0.20, offering efficiency gains for...
Step 5 Preview 600B: 27B active MoE, 1M context
StepFun releases Step 5 Preview open-weights model with 27B active parameters and 1M context window. Self-hostable sparse MoE in FP8 and BF16 formats now available.
MAI-Base-2: Microsoft's 2 trillion parameter base model
Microsoft announced MAI-Base-2, a 2 trillion parameter base model claiming up to 12x hardware efficiency gains over DeepSeek on internal benchmarks.
Nemotron-pretraining-Indic-v1: 2.8T tokens, open pipeline
NVIDIA and AI4Bharat release 2.8 trillion tokens of Indic pretraining data across 22 Indian languages with an open GPU-accelerated generation pipeline.
Beam 501B: 23B active parameters, Apache 2.0 licensed
Beam 501B sparse MoE model from Reflection launches open-weight with 1M context, coding focus, and Apache 2.0 license for customizable deployment.
Muse Spark Open 816B: Meta's Open MoE Model
Meta releases Muse Spark Open 816B, an open-weight multimodal MoE model for coding and agentic work with 1M context. Available on Hugging Face and hosted APIs.
d1-ultra decision model scores 65 on AA index
Liquid AI's d1-ultra multimodal decision model ranks first with a 65 score on the Artificial Analysis Decision Index, processing text and images with calibrated...
Nemotron 3.5 Super VL 120B: Open Multimodal Agent Model
5 Super VL 120B, an open-weight multimodal model for enterprise agents with 1M context window and claimed 5x faster throughput than comparable open models.
Granite Switch 4.1: Chat, decisions, speech endpoint
IBM's Granite Switch 4.1 combines language generation, decision-making, and speech recognition capabilities into a single endpoint. Available in 8B and 30B sizes.
Timestral time-series forecasting model ranks first on three benchmarks
Mistral's Timestral probabilistic forecasting model achieves state-of-the-art results on GIFT-Eval, TIME, and fev-bench with 209M parameters.
Apertus 1.5: 262K context, image and audio support
Apertus 1.5 extends 8B and 70B models with image and audio understanding, quadruples context to 262K tokens, and adds optional thinking mode under Apache 2.0.
Grok Voice 3.0 ranks first on benchmarks, adds tool-calling
xAI's Grok Voice 3.0 speech-to-speech model tops Artificial Analysis index at 86.9 and integrates tool-calling for voice agents. Pricing unchanged at $0.08/min.
HiRO-ACEv2 cuts extreme weather simulation costs by 300x
Ai2 releases HiRO-ACEv2 weather model generating 3km resolution extreme weather data at 1/300th the cost of traditional atmospheric models.
Orion-1 cybersecurity model outperforms Claude Opus 5.5, GPT-6 Sol
Artemis Security launches Orion-1, a purpose-built cybersecurity model that outperforms general AI on threat detection and investigation.
LFM2.5-24B-A2B: Fast on-device reasoning agents
Liquid AI releases LFM2.5-24B-A2B, a lightweight on-device reasoning model for running private AI agents on consumer hardware without cloud processing.
Gemini 4 Argon: $2 per million input tokens, 95% cached discount
Google's Gemini 4 Argon frontier model launches globally via API and app with introductory pricing through November. Ranks first on Vals Index benchmark.
RTX Spark PC AI launches Oct 16 with local agents
RTX Spark arrives Friday with 1 petaflop of AI compute, optimized agent apps, and up to 2x faster Adobe creative software. Local AI, gaming, and coding in one device.
QwenGyre elastic RL framework trains long-horizon agents
Alibaba's QwenGyre framework dynamically reallocates GPUs between rollout and training for extreme-long-horizon agent tasks, delivering up to 1.85x speedups.
Meta Muse Wonder frontier model launches today
Meta releases Muse Wonder frontier models with advanced memory capabilities, priced at $4 per million input tokens and $20 per million output tokens.
NVIDIA Vera Rubin Architecture Advances AI Factory Standards
NVIDIA contributes Vera Rubin rack designs to open standards at OCP Summit, enabling enterprises to build interoperable AI infrastructure that scales to gigawatt power levels.
vLLM Decision 3.0: Five Open Decision Models, 0.8B to 27B
vLLM Semantic Router released Decision 3.0, five open-source decision models from 0.8B to 27B parameters for classification and scoring across text and images.
Google DeepMind's AlphaProof Nexus Solves Open Math Problems
AlphaProof Nexus solved nine open Erdős problems and 44 conjectures using formal verification in Lean, with peer-reviewed results published in Science.
Microsoft Maia 300: 40% better efficiency, production in 2027
Microsoft's Maia 300 AI inference chip delivers 40% better energy efficiency and 7x greater throughput than Maia 200, rolling out to production workloads in 2027.
TII Releases Falcon-ASR Arabic Speech Recognition Model
Technology Innovation Institute launches Falcon-ASR, a 1.6B parameter Arabic speech model with Emirati dialect focus, achieving 20.92% word error rate.
Ai2 Replaces Priority Scheduling With GPU Time Budgets
Ai2 replaced priority-based GPU scheduling with hierarchical fair-share allocation and time budgets, improving resource utilization across its research clusters.
Odyssey-3 World Model Public Preview Launches Oct 11
Odyssey AI launched a public research preview of Odyssey-3, a 14B generative world model that creates interactive environments from text prompts in real time.
OrcaCyber Zero 1.5 Cybersecurity Model Launches With 1M Context
OrcaRouter releases OrcaCyber Zero 1.5, a gated cybersecurity model with 1M context window, reporting 100% on Cybench and 95.8% on CVE-Bench.
Claude Dynamic Workflows: 1,000 Parallel Agents Now Available
Anthropic added dynamic workflows to Claude Managed Agents, enabling up to 1,000 AI agents to run in parallel per execution through managed agent infrastructure.
Liquid AI Releases Multimodal Open D1 Decision Models for Edge
Liquid AI released d1-3B and d1-omni-600M, open multimodal decision models designed for edge devices, supporting text, images, and audio inputs.
Microsoft releases Decision-1 routing and classification model
Microsoft releases Decision-1, a Qwen3.5-9B decision-scoring model for routing and classification tasks, available in Foundry and OpenRouter.
Nace AI Open-Sources Drex 1.5 Decision Model for Option Scoring
Nace AI open-sources Drex 1.5, a 9B decision model that scores multiple options in one forward pass, ranking first under 10B parameters on Decision Index.
Qwen-Image-2.1-Turbo: 8-step image generation on 7B model
Alibaba's Qwen-Image-2.1-Turbo cuts denoising from 40 steps to 8 for fast 2K image generation and editing on a 7B model. What changed and what to test.
Cloudflare releases Clef-omni with audio and video support
Cloudflare launches Clef-omni, a decision model processing audio, video, images, and text in one API call, with price cuts and speed gains for existing models.
JetBrains Mellum2.1: 12B MoE Open Model for Coding Agents
JetBrains released Mellum2.1, a 12B mixture-of-experts open model for coding agents trained on real repositories under Apache 2.0 license.
NVIDIA RTX Spark Windows AI agents arrive October 16
NVIDIA and Microsoft co-engineered RTX Spark hardware and Windows infrastructure to run AI agents locally on PCs, with laptop preorders open now.
Perplexity Releases pplx-embed-v2-late for Edge Device Deployment
Perplexity AI released pplx-embed-v2-late embedding models: a 0.6B edge-deployable model and a 9B high-performance variant for on-device and server RAG.
Claude Haiku 5.5 price cuts, benchmarks reshape model selection
Anthropic released Claude Haiku 5.5 with up to 90% lower token costs and 72.4% computer-use benchmark performance, available now on AWS, Google Cloud, and Azure.
EmbeddingGemma 2: Google's 740M On-Device Embedding Model
Google DeepMind launches EmbeddingGemma 2, a 740M open embedding model mapping text, images, video, and audio into a unified space for on-device retrieval.
Mistral Large 4: Europe's 1 Trillion-Parameter Model Launches
Mistral AI releases public preview of Mistral Large 4, a 1 trillion-parameter model trained in Europe with strengths in cybersecurity, coding, and multimodal tasks.
Reka Rho-1: One 19B Model for Video, Images, Text and Robots
Reka released a 19B omni-reasoning model that handles text, images, video, and robot actions in a single neural network, replacing multi-model pipelines.
GPT-Synopsys: OpenAI and Synopsys Build a Chip Design Model
OpenAI and Synopsys are building GPT-Synopsys, a model meant to reason about chip design and operate Synopsys EDA tools. What is known and what to verify.
Five Lessons from Anthropic's September 2026 Threat Report
A measured reading of Anthropic's September 2026 threat report: how AI misuse is changing, where its evidence stops, and what defenders can do now.
GPT-Live-1 and the New Voice API Decision
What GPT-Live-1 changes for voice apps, what OpenAI claims it costs and scores, and how to test latency, interruptions, accessibility, and cost before launch.
GPT-6 Astra for Work: What the Launch Actually Changes
A sourced guide to GPT-6 Astra: what OpenAI claims, what it costs, the new admin controls, and the checks teams should run before adopting it.
Claude Fable 5.1 and Mythos 5.1: How to Read the Release
Claude Fable 5.1 and Mythos 5.1 are one model with two safeguard levels. Here is what Anthropic claims, what it costs, and how to evaluate it.
Anthropic's Model Hardware Standard, Explained
What Anthropic's Model Hardware Standard research preview does, what early lab partners reported, and the safety checks teams need before agents touch hardware.
Claude Text Watermarking: What It Can and Cannot Prove
How Anthropic's Claude text watermark works, why it differs from AI detectors, and why a positive or missing signal is evidence, not proof.
Claude Opus 5 and the Shift to Long-Running Agents
What Anthropic's Claude Opus 5 launch means for coding and professional work, with pricing, safeguards, and the controls longer autonomous tasks need.
Gemini 3.5: What Frontier Intelligence with Action Means
Google launched Gemini 3.5 Flash as an agentic model family. This guide separates the launch claims from the tests teams should run themselves.
Meta Muse Spark: The New Multimodal Model in Context
What Meta announced with Muse Spark, what its multimodal reasoning and multi-agent Contemplating mode imply, and which claims need independent checks.