- OpenAIGPT-6 Astra$10.00 in
- AnthropicClaude Opus 5.5$4.00 in
- GoogleGemini 4 Argon$2.00 in
- DeepSeekDeepSeek V4.1 Flash$0.15 in
- xAIGrok 4.7$2.00 in
- MistralMistral Large 4$1.36 in
Home/Models
The model, and the rate beside it.
Releases, list prices, and the stories we wrote. A vendor score stays a vendor score.
Cache, output, and retries change it.
Published list rates, not a measured invoice.
/ Models
AI model news, releases, cost, and comparisons. What shipped, who measured it, and which figures are still vendor claims.

Banned Nvidia Blackwell chips power China's most advanced AI models
A former OpenAI researcher found that the leading open-weight models from DeepSeek, Moonshot AI, MiniMax, and Ant Group were trained on Nvidia Blackwell GPUs that U.S. export controls ban from China. The finding raises questions about whether the chip restrictions are working.

ChatGPT Wallet lets AI agents shop and check out for you
OpenAI launched ChatGPT Wallet, a payment feature that lets its dots agents make purchases on behalf of users. The wallet works with Visa, Mastercard and Amex, and each transaction requires biometric approval.

Google launches Gemini 4 Hydrogen for fast, everyday agentic work
Google today introduced Gemini 4 Hydrogen, a new model designed for speed and efficiency across everyday coding, agentic workflows, and search. At $1.50 per million input tokens, it replaces the Flash tier with Gemini 4-class intelligence.

OpenAI hits $10B quarterly revenue for the first time
OpenAI hit $10 billion in Q3 revenue, up 50% quarter over quarter, as losses narrowed on a margin basis. CFO Sarah Friar called it a blockbuster quarter.

Google opens Gemini 4 Carbon to everyone with agent swarm preview
Google DeepMind has released Gemini 4 Carbon to all developers and consumers two weeks after previewing it for cybersecurity defenders, alongside a preview of its /swarm multi-agent coordination feature in Antigravity.

Meta releases Muse Ember, a $0.10 per million token lightweight model for AI agents
Meta released Muse Ember, a lightweight model at $0.10 per million input tokens designed for the repeatable work AI agents do all day - tool calls, coding subtasks, routing, and document parsing. It rivals Claude Haiku 5.5 and GPT-6 Luna on several benchmarks at about a third of the cost.

Anthropic launches Claude Fable 5.2 and Mythos 5.2 with lower prices and 30 percent faster performance
Anthropic released Claude Fable 5.2 and Mythos 5.2, its most capable models, with reduced pricing at $8 per million input tokens, response speeds up to 30 percent faster, and significant benchmark gains in coding, cybersecurity, and scientific research.

Underdog's Woof packs a 32MB speech model into a browser game
Underdog, the startup behind the open Saluki models, showed off Woof, a speech generation model weighing just 32MB with 31 million parameters, in a browser game that runs fully offline on most devices.

Tencent launches Hy4, a 770-billion-parameter open-source model with vision and computer-use skills
Tencent released Hy4, a 770B-parameter open-weights MoE model with 1M-token context, native visual understanding, and computer-operation capabilities, positioning it among the world's leading open-source AI systems.

Google launches Gemini Omni 1.2 Flash with 15-second video generation and 30 percent lower pricing
Google's Gemini Omni 1.2 Flash brings native 15-second video generation, multi-turn conversational editing, stronger consistency and a 30 percent price cut. The model is now generally available across the Gemini API, Google Cloud and apps like Flow and Google Vids.

Mistral releases Le Chonk open weights as the best open-weight model for coding
Mistral released the open weights of Mistral Large 4 (Le Chonk), its 1-trillion-parameter multimodal flagship, on 30 October, following a public preview earlier this month — with three additional weeks of reinforcement learning making it the best open-weight model for coding and improving reasoning efficiency by 30% fewer output tokens.

Hugging Face releases open Swift packages for on-device speech and decisions on Apple platforms
Hugging Face's mattt released two open-source Swift packages — AnySpeechRecognizer and AnyDecisionModel — that let Apple developers run speech recognition and decision models entirely on device, with a choice of open models including NVIDIA's Parakeet, Alibaba's Qwen3-ASR and decision models from Jev or Ollama.

SpAItial launches Echo-3, a world model that turns iPhone video into metric 3D reconstructions
SpAItial AI released Echo-3 and a companion iPhone app on 27 October 2026, turning a short phone video into a photorealistic, metric-scale 3D reconstruction of a real room, building or outdoor space — a world model that treats reconstruction as conditional generation rather than freeform invention.

Google launches Gemini 4 Carbon, a frontier model honed for cybersecurity
Google announced Gemini 4 Carbon on 28 October 2026, its strongest model with state-of-the-art cybersecurity and long-horizon capabilities, ranking #1 on the Artificial Analysis Intelligence Index v5 at 59 and leading vulnerability discovery, exploitation and patching benchmarks.

NVIDIA releases Nemotron 3.5 Transcribe, an open-weight speech model that transcribes 40 languages
NVIDIA released Nemotron 3.5 Transcribe on 28 October 2026, an open-weight model combining transcription, speaker attribution, translation and audio question answering in one checkpoint: 40 transcription languages, 31 translation languages and up to eight speaker labels per recording.

Swiss lab Giotto launches Alto and Solo, sovereign reasoning models it says are Europe's smartest
Swiss lab Giotto launched Giotto Alto and Giotto Solo on 27 October 2026, billing them as the smartest European sovereign AI models on the market, with Alto free in the browser and mobile apps and Solo designed to run on a single GPU.

Best open-source LLM for coding: what to test before you pick one
No single open-source coding model is best for every repository. A HumanEval score measures functional correctness on 164 handwritten problems, not your codebase.

Liquid AI releases LFM3, a family of open-weight models built for phones and laptops
Liquid AI released the LFM3 family, six open-weight models in the 2-billion-parameter class covering text, vision, audio, an omni model and a decision model, trained from one 1.7B base.

OpenAI introduces GPT-6.1 Terra, a faster and cheaper computer-use model
OpenAI released GPT-6.1 Terra on 26 October, a fast, efficient model for computer use and visual work, priced 50% lower than GPT-5.6 Terra and GPT-6.1 Sol.

iFlytek launches Spark X2.6, a 293B model it says was trained only on Chinese chips
iFlytek's Spark X2.6 is a 293B-A30B MoE with 1M context for coding and agents, which it says was trained on domestic compute. A small Spark-X2.6-4B ships open under Apache 2.0.

Hi3D open-sources Hi3D v1.0, the image-to-3D refiner formerly called Sparc3D
Hi3D has released code and weights for Hi3D v1.0, a model that turns a coarse 3D mesh and a photo into detailed geometry at up to 1536³. It reports a 90% 3D-printing slice rate, under a custom licence with branding rules.

Nanbeige releases Nanbeige4.3-33B, an open 33B agent model that tops Qwen3.8-27B on its own tests
Nanbeige's Apache 2.0 Nanbeige4.3-33B is a dense 512K-context model for coding agents and office tasks; its own benchmarks beat Qwen3.8-27B on most agent tests, not all.

fal launches FLUX 3 Max, a video model that makes 10-second clips in about 12 seconds
fal post-trained Black Forest Labs' FLUX 3 into FLUX 3 Max, a hosted video model it says matches Seedance 2.5's preliminary arena rating while rendering about 22 times faster.

Grok Imagine releases episode 1 of its AI-made Odyssey with Wonder Studios
xAI's Grok Imagine posted a seven-minute first episode of a Homer's Odyssey series made with London studio Wonder Studios and director Dev Walker, the first step toward Elon Musk's promised AI film.

Shanghai AI Lab releases Atria Dawn open weights under MIT with 512K context
Shanghai AI Laboratory published the full Atria Dawn weights, a text-only agent model built on the 744B GLM-5.2, under the MIT License with a 512K context window.

Ant Group's Ling-3.0-flash-Sante is an open health model that only GPT-5.6 beats on HealthBench
inclusionAI, Ant Group's AI arm, released Ling-3.0-flash-Sante, a 124B mixture-of-experts medical model under MIT. Ant's table puts it second only to GPT-5.6 Sol on HealthBench Professional, but behind on many exam tests.

Tencent open-sources Hy-Nomas, a world model that keeps building a scene from new prompts
Tencent Hunyuan released Hy-Nomas under Apache 2.0: an interactive video world model that extends a scene segment by segment from text and camera paths, with a single RTX 4090 mode.

NVIDIA releases Nemotron 3.5 Bolt E4B, a 4GB model for local voice agents
NVIDIA's Nemotron 3.5 Bolt E4B 4GB is a Gemma 4-based small model trained with reinforcement learning only. It leads 4B peers on instruction following but loses some vision skill.

KRAFTON releases Raon-1 mini, a 30B open model the PUBG maker trained from scratch
KRAFTON's Raon-1 mini preview is a 30B-A4B mixture-of-experts model trained from scratch, released under Apache 2.0 with strong Korean and agent scores and an RTX 5090-sized build.

Block unveils Dessa, a voice model trained only on human speech
Block's small Dessa lab says its full-duplex model learned to talk from recorded conversation alone, with no text, runs on a MacBook Pro and was trained on one GB200 NVL72.

Moonshot releases Kimi K3.1 Pro, a 3T open-weight model that nears Claude Opus 5.5 on coding
Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0 and ships with cyber data removed. After launch rate limits, Moonshot reset a quarter of every subscriber's monthly quota.

Canva releases Craft, its first in-house image generation model
Canva Research's Craft plans a layout before rendering it, was trained on millions of Canva templates, and is live for all Canva users and through the public API, with faster variants to come.

Gradium open-sources Phonon, an on-device TTS that starts speaking in milliseconds
The Kyutai spin-out released 31m and 40m Phonon text-to-speech models with Apache weights for the English 31m checkpoint, five-language 40m variants, and reported single-digit-millisecond first audio on laptop CPUs.

Ai2 releases OlmoEarth v1.3 with free satellite embeddings for the whole planet
Ai2's OlmoEarth v1.3 turns a year of Sentinel and Landsat data into one vector per 10-metre pixel, with global embeddings and a browser explorer that need no GPU or sign-in.

HeyGen launches HeyGen Video, a model that makes picture and sound in one pass
HeyGen's first general-purpose video model generates 5-15 second clips up to 2K with dialogue and effects in one pass, and its top-ranked Voice Turbo is free in the API until 30 November.

Cohere launches Rerank 5 with calibrated scores and passage-level Context Focus
Cohere's Rerank 5 Pro and Fast rerankers are generally available at $0.05 and $0.02 per million tokens, with Context Focus passage extraction and relevance scores that Cohere says read as probabilities.

