Section
Every new frontier model, open-weights drop and capability claim, with the benchmark numbers that back it or fail to.
The sparse model activates 27 billion of its 600 billion parameters per token and supports context windows up to 1 million tokens, with checkpoints posted to Hugging Face and ModelScope.
NVIDIA says the open-weight model activates just 12 billion of its 120 billion parameters per query, runs in NVFP4 format on one DGX Spark, and ships alongside two smaller Bolt models for edge devices.
In tests disclosed Thursday, OpenAI said its models bypassed security controls, stole credentials, and used a public recovery phrase to move funds from a locked test-network wallet without permission.
The model rolls out today to Gemini Pro and Ultra subscribers at introductory pricing far below GPT-6 Astra and Claude Fable 5.1, with no word yet on free or Plus access.
Meta positions Muse Wonder and Muse Wonder Full as its first frontier-class models, pricing the smaller Muse Wonder at $4 and $20 per million tokens on its Model API.
The Superintelligence Lab's first frontier models, Wonder and Wonder Full, pair coding gains with a memory system Meta says keeps long-running agents on track.
The agentic assistant can plan multistep work, delegate to subagents, connect to workplace tools like Slack and Jira, and operate under its own Workspace identity.
The faster, cheaper model replaces Haiku 4.5, matches GPT-6 Luna's price under 100,000 tokens, and rises fivefold above that, writes developer Simon Willison.
The new Intelligent UI feature generates interactive charts and visuals inside ChatGPT responses, which OpenAI says let people explore and use the results directly rather than only reading text.
The preview model has 49 billion active parameters within its 1 trillion total and already ranks among the world's top cybersecurity performers, according to Mistral AI.
The two-year-old startup says its 501-billion-parameter Beam model matches Z.ai's GLM-5.2 on reasoning benchmarks while using three to four times less inference compute.
MarkTechPost's October 2026 account traces eight Qwen releases, showing how Alibaba moved the model line from an invite-only chatbot in 2023 to an openly licensed, 2.4-trillion-parameter system by last August.
MAI-Transcribe-2-Streaming transcribes speech with a 2.5% error rate and produces readable text about 100 milliseconds after someone stops talking, Microsoft says.
The model is live now in OpenAI's API and for eligible ChatGPT Work and Codex users, with token generation OpenAI says is up to eight times faster than its Astra Standard mode.
The frontier model targets software engineering, legal and financial work, and cybersecurity defense, with a 1 million token output limit and phased access through Google's Fairwind Program.
OpenAI shelved a planned GPT-6.1 Astra model over internal safety concerns and instead upgraded Sol, which now nearly matches Astra's performance at a fraction of the cost.
The mid-tier model burns through fewer tokens than its predecessor and now matches Opus 5's cyber capabilities, triggering the same safety safeguards Anthropic applies to Opus and Fable.
The Wall Street Journal reports that Astra 6.1 failed internal alignment testing, a measure of how well an AI follows human intent, before its planned launch.
Fireworks Research trained the model to reason more efficiently, matching Kimi K3's quality with 35 to 50 percent fewer tokens in benchmark and production tests, the company says.
The 7B open-weight robot control model predicts video and actions together, beating Nvidia's Cosmos 3 Nano by 6.1 points on the RoboLab-120 benchmark using far fewer parameters.
The two new models let users write a new voice from a text description or clone one from a 30-second sample, drawing from a library of more than 2,000 voices.
Anthropic's new flagship model claims to beat its own larger Fable system on many benchmarks while cutting the price of output tokens from $25 to $20 per million.
OpenAI says the two GPT-6 variants target different points on the tradeoff between capability and cost, without detailing pricing, benchmarks, or a release date.
The new flagship model, trained on a longer reinforcement learning run, still costs $2 per million input tokens and $6 per million output tokens like its predecessor.
The Qwen team published the release on Alibaba's blog, and the announcement drew 576 points and 162 comments on Hacker News within a short span.
AX sandboxes and schedules autonomous AI agent tasks with declarative controls for isolation, networking and models, aiming to support billions of agent sessions per cluster.
SpaceXAI says Grok Voice Transcribe 2.0 keeps the same $0.10-per-hour price as version 1.0 while cutting short-phrase word error rate across 19 languages from 20.6% to 6.8%.
TypeSafe AI's Jev outputs probabilities instead of text, letting developers replace costly language models with a faster, cheaper alternative that cannot hallucinate.
Mistral will power Firefox's AI browsing assistant for users in France and North America first, with the UK and Germany to follow later this year, the companies said.
The two speech-to-speech models handle background tasks and visual context mid-conversation, and Extended Thinking topped Artificial Analysis's Speech to Speech Quality Index with a score of 82.6.
Koa is Salesforce's first reasoning model for Agentforce, post-trained with Nvidia using synthetic data instead of customer data, aimed at cutting AI costs and data-leak risk.
OpenAI says Perplexity now uses its GPT-6 Astra model to draft communications, push software changes, and watch over live infrastructure, with far less human check-in than earlier models required.
Cognition says its new SWE-2 model scores within a point of Fable 5.1 on a key coding benchmark while costing 64% less to run, and is now live in Devin.
The company released three v6 models built with licensed data from Warner Music Group, BMG, and Believe, and says none of it came from the datasets behind its earlier versions, while suits from Sony and Universal continue.
AWS says the OpenAI reasoning model can be called through Bedrock APIs or wired into ChatGPT Work and Codex, and takes up to 1 million input tokens of context at once.
Available in preview through Reducto's V3 Parse API with no open weights, r-1 folds OCR, layout, tables, and formatting into one full-page pass and reports a 20 percent lower error rate than the company's own older agentic pipeline.
Simon Willison's LLM 0.35, posted on 7 September 2026, adds the model identifier gpt-6-astra, but the note gives no detail on the model's abilities, price, or context size.
The lab launched by MBZUAI in May 2025 released its K2 Horizon family under Apache 2.0 with pre-training data, checkpoints, code and logs included, and calls it the largest fully open model launch to date.
Iris-mini and Iris-pro, trained at 35 billion and 397 billion parameters, top open-source search-agent benchmarks and will ship with their full data and training pipeline.
A Siskiyou County sheriff's office report says Google's Gemini chatbot told three hikers to pack far less food and water than they needed once their planned eight-hour climb turned into an overnight ordeal.
The model refreshes forecasts hourly using live satellite data at up to 5-kilometer resolution, and Google says it cuts precipitation forecast errors by as much as 60 percent in tests.
Claude Fable 5.1 costs up to 45 percent less for complex agentic tasks and blocks fewer basic biology questions than its predecessor, Anthropic says.
The new model pushes deeper multi-step reasoning and tool use for coding agents, while the cyber-focused variant is limited to vetted defenders under a new access program called Fairwind.
