News category
79 briefs in this category, newest first.
Google released EmbeddingGemma 2 on October 6, a 740-million-parameter open embedding model that maps four data types into a single vector space and runs locally on a phone.
Mistral Large 4, a 1-trillion-parameter open-weight model nicknamed Le Chonk, scored 82% on a vulnerability-reproduction test where Claude Opus 5.5 and GPT-6 Astra scored near zero because they refused the task.
Reflection AI unveiled Beam on October 5 with benchmark claims against Z.ai's GLM-5.2, but the Apache 2.0 weights will not post until later this month and only a waitlist is open today.
AWS released a 2-billion-parameter open model that scores a developer's fixed set of answers instead of generating text, landing days after OpenAI shipped a hosted rival for the same new model category.
Google's new frontier model can rewrite C and C++ codebases into Rust at scale and autonomously patch critical vulnerabilities, but it ships only to vetted Fairwind Program partners, with no public pricing tier live yet.
OpenAI released GPT-6.1 Sol at DevDay on September 29, a week after GPT-6 Sol, while quietly shelving a planned GPT-6.1 Astra update that the Wall Street Journal reports failed internal safety testing.
Anthropic's new mid-tier model scores within 2 points of Opus 5.5 on real-world work benchmarks while costing half as much per token, and GitHub Copilot added it the same day.
NaiveAI's Naive-N0.5-Flash pairs a no-full-attention layout with a stated claim: AI systems, not human engineers, explored and designed the architecture, then tuned its own training and inference stack.
Xiaomi released the MiMo-V2.6 series on September 22, 2026, with the 1T-parameter Pro variant scoring 46 on the AA Intelligence Index to claim the top open-access model position worldwide, under an MIT license.
Anthropic launched Claude Opus 5.5 on September 22, 2026, cutting cache-read costs 60 percent and input tokens 20 percent below Opus 5, while recording the company's highest automated behavioral audit scores ever.
OpenAI launched GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50 per million tokens on September 22, 2026, cutting predecessor prices by 50 percent or more and marking the cuts as permanent.
Google released Gemini 3.8 Live and 3.8 Live Extended Thinking on September 15, 2026. The Extended Thinking variant scored 82.6 on Artificial Analysis' Speech-to-Speech Quality Index and 68.6% on the tau-Voice agentic task completion benchmark, the highest marks on both leaderboards at launch.
Alibaba released Qwen-Image-2.1 on September 20, a 7B diffusion model with native 2048x2048 generation and RGBA transparency support, but replaced the permissive Apache 2.0 license with a research-only agreement that requires a separate commercial contract.
Alibaba's Qwen team released Qwen3.8-Omni-Flash on September 18, 2026. The model accepts text, image, audio, and video inputs across a 1-million-token context window, with audio processing priced at more than 98 percent below Qwen3.5-Omni-Plus per Alibaba's own calculation.
StepFun shipped the Step 5 Preview API on September 20, 2026, offering a 600 billion-parameter sparse mixture-of-experts model with 27 billion active parameters per token, a 1 million-token context window, and open weights scheduled for October 15.
Salesforce unveiled Koa, its first CRM reasoning model built with Nvidia on synthetic data from nearly 27 years of CRM deployments, plus AIforce, which routes Salesforce data into Claude, Slack, and Amazon Bedrock.
Shanghai AI Laboratory released Atria Dawn Preview, a 744B-parameter MoE agent under MIT license. A 769-task human study found participants rated about one-third of AI-assisted work infeasible without the model.
Japan's Sakana AI released two orchestration models on September 11: Fugu Max at $2 per million input tokens and Fugu Ultra v2, which scored 48.3 on Chartography visual reasoning versus 27.3 for Opus 5, without Fable 5 or GPT-6-Astra in its pool.
DeepSeek released V4.1 Flash on September 10 with a 1M-token context window, an FP4 KV cache that drops per-token memory to 890 bytes, cross-layer attention reuse, native multimodal, and MIT open weights in 48 Hugging Face shards. V4 Pro API calls reroute to it at Flash prices starting September 14.
Meta released Muse Spark 1.3 on September 2, 2026, pricing the multimodal reasoning model at $1.25 per million input tokens and $4.25 per million output tokens. The model places 13th on the Artificial Analysis Intelligence Index with a score of 48, well above the median of 24 for comparable models.
GPT-6 Astra launched September 3, 2026 with the highest capability rating in OpenAI's Preparedness Framework. API access costs $10 per million input tokens with a pricing cliff above 272K tokens that doubles costs on the full request.
Meta released Muse Voice Transcribe on September 1, 2026, the first real-time audio perception model to handle streaming transcription, speaker identification for 20-plus voices, and sentence detection in a single end-to-end architecture at $3.00 per 1,000 audio minutes.
OpenAI put GPT-6 Astra in paying users' hands on September 3, 2026, the first model classified Critical under its Preparedness Framework for cybersecurity, as president Greg Brockman told reporters: Welcome to the AGI era.
Google's fourth Flash-tier model in four months scores 59 on the Artificial Analysis Intelligence Index and ships with a purpose-built cybersecurity sibling locked to vetted Fairwind Program partners.
Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 1, cutting cache-read prices 75% to $0.25 per million tokens and reducing cybersecurity false positives by 60%.
Alibaba opened developer API access to Qwen3.8-Flash-Next on September 1, 2026. The 125B-parameter MoE model previews the Qwen4 N-gram architecture and prices at $0.16 per million input tokens, one-twelfth the cost of flagship Qwen3.8-Max.
Zhipu AI confirmed August 26 that mystery model Ox Alpha is GLM-5.3-Flash, a 320B MoE with 18B active parameters, native multimodal input, MIT weights on Hugging Face, and a 1M-token context window priced at $0.075 per million input tokens through September 9.
An anonymous model called Ox Alpha appeared on OpenRouter on August 20, 2026, outpaced Claude Fable 5 on community coding benchmarks, and drew 221,000 unique users before independent researchers traced its serving infrastructure to Zhipu AI. No official attribution has appeared.
Z.AI's GLM-5.3 became available via API on August 18, delivering 50% stronger coding scores than GLM-5.2 while using fewer output tokens per task. Open-source weights are expected around August 28.
Alibaba shipped Qwen3.8-27B on August 14 under Apache 2.0. The 27.78B-parameter dense model scores 42.2 on DeepSWE 1.1 and 84.3 on OSWorld-Verified, both from a single 24GB VRAM GPU.
Google released Gemini 3.7 Flash on August 13 at an introductory price of $0.75 per million input tokens, half of what 3.6 Flash cost at launch, with benchmark gains across coding, agentic, and knowledge tasks.
Zhipu AI's 743-billion-parameter GLM-5.3 scored 84.5% on the CyberGym vulnerability benchmark on August 14, edging Anthropic Mythos 5 and GPT-5.6 Sol. Open weights ship in roughly two weeks at about one-tenth the cost of U.S. frontier models.
DeepSeek moved V4-Pro to general availability on August 13, 2026 with a major jump in agent benchmarks and released its internal Harness framework under the MIT license, giving developers a provider-agnostic agent stack.
ByteDance's Seed team deployed SeedRealtime on August 5, 2026, a single-architecture model that processes audio, video, and text simultaneously, eliminating the cascaded listen-transcribe-speak pipeline that adds latency to most competing voice-AI products.
Meta released Muse Glimmer on August 10, 2026 under Apache 2.0, the first open-weight model distilled from a frontier agentic system that fits in a 24 GB consumer GPU, with tool use, multi-step reasoning, and failure recovery built in.
Alibaba committed to releasing Qwen3.8-Max weights on Hugging Face and ModelScope this week. The 2.4-trillion-parameter MoE API launched August 3 at $2/$6 per million tokens.
ByteDance's Seedance 2.5 produces 30-second video clips with synchronized audio in a single generation pass and accepts up to 50 multimodal reference inputs. Volcano Engine and BytePlus ModelArk API access opened on August 7, 2026.
Kimi K3 reached general availability in GitHub Copilot on August 6, 2026, giving subscribers access to the largest open-weight model deployed in a mainstream coding tool, hosted on Fireworks AI under usage-based billing.
Liquid AI released LFM2.5-2.6B on August 4, an open-weight 2.6-billion-parameter model that handles 128k-token contexts, calls tools natively, and runs on consumer hardware including phones, with no per-token billing.
xAI launched Grok Voice Think Fast 2.0 on July 29, 2026, routing the grok-voice-latest alias to the new model on August 5. Artificial Analysis benchmarks show it at 82.9% on the Speech-to-Speech Quality Index, ahead of GPT-Realtime-2.1 at 79.1% and Gemini 3.1 Flash at 69.5%.
Mistral released Shieldstral on August 4, 2026, an Apache 2.0 open-weights safety classifier that handles text and image moderation at 3B parameters while matching guard models up to 7 times its size, using inference-time policy queries instead of a fixed harm taxonomy.
Alibaba opened global API access to Qwen3.8-Max on August 3, posting 86.1 on OSWorld-Verified against Fable 5's 85.0. The 2.4-trillion-parameter MoE model starts at $2 per million input tokens; open weights and the QwenWork enterprise agent platform arrive within the week.
Alibaba made its 2.4-trillion-parameter Qwen3.8-Max accessible globally on August 3 via Alibaba Cloud and the new QwenWork platform, and confirmed the weights will ship publicly within a week, marking the first Max-class Qwen model to go open-source.
DeepSeek moved its V4-Flash API to public beta on July 31, 2026 with a build that beats V4-Pro-Preview on every published agent benchmark at $0.14 per million input tokens, through re-post-training on an unchanged 284B-parameter architecture.
Google DeepMind released three Gemini Robotics 2 models on July 30, 2026, extending AI control from upper-body-only to full humanoid motion, adding multi-robot collaboration, and delivering an on-device variant that adapts to new robot bodies in hours with fewer than 200 examples.
xAI released Grok Voice Think Fast 2.0 on July 29, 2026, reaching 82.9% on Artificial Analysis's speech-to-speech quality index, ahead of GPT-Realtime-2.1 High at 79.1%. Time to first audio dropped to 0.70 seconds and the model is priced at $0.08 per audio minute.
Moonshot AI released the Kimi K3 open weights on Hugging Face on July 27, 2026, making the 2.8-trillion-parameter sparse MoE model the largest openly available frontier model to date. The MXFP4 footprint of roughly 1.4 TB keeps self-hosting in cloud infrastructure territory.
Claude Opus 5 launched on July 24, 2026, priced at $5 per million input tokens, the same as Opus 4.8, while setting new top results on Frontier-Bench v0.1 and beating Fable 5 on OSWorld 2.0 computer use at just over one-third the cost.
Black Forest Labs announced FLUX 3 on July 23, 2026, the company's first model to jointly train on images, video, and audio in a single architecture, with a robotics action module already in Audi production tests.
Poolside's Laguna S 2.1 is a 118B mixture-of-experts coding model that scored 78.5% on SWE-Bench Multilingual and 70.2% on Terminal-Bench 2.1, runs on a single Nvidia DGX Spark, and ships free on Hugging Face under a commercial-use license.
Google and Google DeepMind launched three Gemini models on July 21, 2026. Gemini 3.6 Flash cuts output-token usage by 17 percent and prices output at $7.50 per million tokens, while a government-only Flash Cyber variant outperformed Claude Opus 4.6 in V8 vulnerability testing.
Kimi K3 from China's Moonshot AI is the world's first open 3T-class model: 2.8 trillion parameters, 1-million-token context, and benchmark scores above Claude Opus 4.8 on most coding tasks. Open weights publish July 27.
Mira Murati's Thinking Machines Lab released Inkling on July 15, 2026: a 975-billion-parameter mixture-of-experts model with 41B active parameters per forward pass, a 1-million-token context window, and native audio understanding. Full weights are publicly downloadable.
Mistral AI released Robostral Navigate on July 8, its first robotics navigation model. The 8B system scored 76.6% on R2R-CE for unseen environments using one RGB camera, 4.5 points above the leading multi-sensor approach and 9.7 points above the prior best single-camera system.
Meta Superintelligence Labs launched a public preview of Muse Spark 1.1 on the Meta Model API on July 9, 2026. The multimodal reasoning model carries a 1-million-token context window, costs $1.25 per million input tokens, and accepts both OpenAI and Anthropic SDK integrations.
OpenAI made GPT-5.6 Sol ($5/$30 per million tokens), Terra ($2.50/$15), and Luna ($1/$6) generally available on July 9 with no waitlist. Sol is now the default model in ChatGPT.
OpenAI released GPT-Live-1 and GPT-Live-1 mini on July 8, replacing Advanced Voice Mode with models that process incoming and outgoing audio simultaneously, while routing complex queries to GPT-5.5 in the background.
A UI upgrade string reading 'Unlock the full power of Chat with Grok 4.5' appeared in the Grok web client on July 6, 2026. xAI has not made a formal announcement, but the signal matches the company's pattern of UI-string-to-launch gaps of one to three days.
Mistral AI released Leanstral 1.5 on July 2, a 119-billion-parameter open model for Lean 4 formal verification that saturated miniF2F, solved 587 of 672 PutnamBench problems, and found five previously unknown bugs in production code across 57 repositories.
Google DeepMind is targeting July 17 for Gemini 3.5 Pro, per multiple leak sources, after abandoning the Gemini 2.5 Pro base in favor of a full pre-training restart that pushed the model from June into July.
China's food-delivery giant Meituan officially released LongCat-2.0 on June 30, 2026. The 1.6-trillion-parameter MoE model had already reached the top three on OpenRouter by call volume before anyone knew who built it.
OpenAI released gpt-realtime-2.1 and gpt-realtime-2.1-mini on July 6, 2026, delivering at least 25% lower p95 latency and, for the first time, configurable chain-of-thought reasoning inside a live audio connection.
OpenAI previewed GPT-5.6 Sol, Terra, and Luna on June 26, 2026 under a U.S. government-requested access restriction. CEO Sam Altman's 'coming weeks' timeline means broader rollout is expected between July 7 and 14, 2026.
Meta Superintelligence Chief Alexandr Wang told employees on July 2 that the company's in-training Watermelon model has reached parity with OpenAI's GPT-5.5 on internal benchmarks, using an order of magnitude more compute than its predecessor Avocado.
Claude Sonnet 5 launched June 30 as the default model across Free and Pro tiers, beating Sonnet 4.6 on agentic coding benchmarks and approaching Opus 4.8 performance at Sonnet pricing.
Google's Gemini 3.5 Pro failed to reach public general availability by June 30, 2026, the date Google committed to at its I/O developer conference. The model remains in a limited Vertex AI enterprise preview. A Polymarket contract tracking the outcome settled at 97% No-Release, on $229,697 in trading volume.
OpenAI began a limited preview of the GPT-5.6 model family on June 26, 2026, with access confined to roughly 20 organizations cleared through a White House executive order process. The three-tier family introduces new reasoning modes and a reworked naming system, with a broad release targeted for the coming weeks.
Within days of each other, Tokyo's Sakana AI and Beijing's 360 Security launched frontier-capable cybersecurity models targeting the void left by US export restrictions on Anthropic's Fable 5 and Mythos.
OpenAI began a limited preview of GPT-5.6 on June 26, 2026, shipping three models at distinct capability and price tiers, plus a new 'ultra' sub-agent mode on Sol that routes complex tasks across multiple agents.
Chinese AI lab Z.ai released GLM-5.2 on June 16 under the MIT license. The 753B-parameter open model outscores GPT-5.5 on four major coding benchmarks and undercuts its API price by roughly six to one.
ByteDance announced Seedance 2.5 at the Volcano Engine FORCE conference on June 23, 2026, targeting an early July launch with 30-second single-pass video and 50 multimodal reference inputs.
Google has pushed Gemini 3.5 Pro's general availability to July 2026, breaking a public June window Pichai gave developers at Google I/O and arriving as two of Google's most decorated AI scientists leave for OpenAI and Anthropic.
Sakana AI launched Fugu and Fugu Ultra on June 22, 2026, delivering multi-model orchestration through a single OpenAI-compatible API and pitching it as a hedge against export-control disruptions like the one that cut off Anthropic's Fable and Mythos models.
A SemiAnalysis trace-level analysis published in June 2026 found that DeepSeek V4-Pro and Huawei's Ascend 950DT accelerator were co-designed together, not adapted post-hoc, slashing per-million-token costs to 0.20 yuan and raising new questions about the reach of US semiconductor export controls.
OpenAI has not announced GPT-5.6, but a Pachocki internal memo, $1.1M in Polymarket contracts, and ChatGPT Pro user reports of unusual response quality all point to a soft rollout already under way.
Alibaba's Qwen AI research team announced the Qwen-Robot Suite on June 16, 2026: three foundation models designed to give physical robots autonomous navigation, arm control, and scene understanding capabilities.
Anthropic released Claude Fable 5 and Claude Mythos 5 on June 9, 2026, making a Mythos-class model available for general use for the first time and pricing both at $10 per million input tokens and $50 per million output tokens.
OpenAI's GPT-5.5 went generally available in Microsoft Foundry on April 24, 2026, with a Pro tier priced at $30 per million input tokens for the most demanding enterprise agent workloads.
Anthropic released Claude Opus 4.8 on May 28, 2026, pairing the model upgrade with a new dynamic workflows feature for Claude Code and a fast-mode price cut of roughly three times compared to the previous generation.
Picks, reviews, and automation tips every weekday. Free, no spam.