Frontier Radar: Models

Topics: Models · Agent platforms · Pricing changes · Weekly archive · RSS feed

• 2026-11-12
  Perplexity's Decider v1.3 and Decider Flash v1.3 can now pull exact text spans with Locate
  Perplexity released Decider v1.3 and Decider Flash v1.3, the follow-ups to its earlier decision models. Instead of writing text, they return probabilities. They answer yes/no, multiple-choice, and score questions about a text or image input, which makes them a good fit for routing, classification, and checks inside agents. The new Locate feature returns exact text spans with start and end character offsets and confidence scores. It works on text only, with up to 128 Locate questions per request. The model card says v1.3 is built on Qwen3.8-27B and needs about 49 GiB of BF16 weights. Flash is distilled from it. Perplexity says both are on Hugging Face under an Apache license, through OpenRouter, and in the Perplexity Decisions API. But the Hugging Face model cards currently say the repositories are private and need an access token. MarkTechPost reports API pricing of $0.02 (v1.3) and $0.01 (Flash) per 1M input tokens, with output free. Source: Perplexity blog. Source
  Models Pricing changes
• 2026-11-12
  Edge0 open-sources Edge0-145B, a 145B-parameter model it says runs coding and agent tasks on a laptop
  On November 12, Edge0 open-sourced Edge0-145B and Edge0-120B, along with its inference framework and deployment tools. It calls Edge0-145B the first open-source on-device model with 145 billion parameters. On a compatible laptop such as a MacBook Pro, it reports 29GB peak memory and more than 80 tokens/s decoding. It also reports 1,100+ tokens/s prefill and a 128K-256K context. Edge0 says the model can plan, write and run code, debug, and check results for coding and agent work. It says there are no cloud token fees and core features work offline. The performance figures are Edge0's own. Source: Edge0 blog. Source
  Models
• 2026-11-12
  Google launches Gemini 4 Hydrogen, a fast model for coding and agentic workflows at $1.50 / $7.50 per 1M tokens
  On November 12, Google introduced Gemini 4 Hydrogen, the fast member of the Gemini 4 family. It's built for quick work across everyday coding and agentic workflows, while Gemini 4 Carbon and Argon handle deep, long-horizon reasoning. Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, and 54 at xhigh. Argon scores 50. Hydrogen has a 1M-token context window. It costs a flat $1.50 per million input tokens and $7.50 per million output tokens across that window. Starting today, it is rolling out globally across all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, and Google Search. Paid Google AI Pro and Ultra subscribers can try it in Google Antigravity. Source: Google blog. Source
  Models Pricing changes
• 2026-11-11
  Google launches the Gemini Decisions API and Gemini Decision 1, which scores up to 1,000 questions per call at $0.10 per 1M input tokens
  On November 11, Google DeepMind introduced the Gemini Decisions API, a new endpoint for fast, closed-set decisions. It runs on a new model, Gemini Decision 1 (gemini-decision-1). Instead of generating text, it scores candidate options in a single pass and returns probabilities and confidence scores. It's aimed at tool selection, policy checks, model routing, and context grading inside agents. One request takes a shared input of text, images, audio, video, or PDFs (up to 256K tokens). That input comes with up to 1,000 yes/no, multiple-choice, or rubric-score questions. Google says end-to-end latency is 150-250 ms. Pricing is $0.10 per 1M input tokens, with no output tokens billed, and there is a no-cost tier in Google AI Studio. Source: Google blog. Source
  Models Pricing changes
• 2026-11-11
  Zyphra releases ZONOS3-Mini, an Apache 2.0 voice-cloning TTS model with 100M active parameters that runs on-device
  On November 11, Zyphra released ZONOS3-Mini, an open-source text-to-speech model for zero-shot voice cloning and fine-grained vocal control across 43 languages. It uses a sparse mixture-of-experts backbone with 100M active and 800M total parameters. Zyphra reports under 570 ms time-to-first-audio in-browser on a Pixel 9 and under 150 ms on an M3 MacBook Air. It reports 80 ms median on a single H100. Weights are on Hugging Face under Apache 2.0, with MIT-licensed runtimes for SGLang-Omni, Apple MLX, C++/ggml, and in-browser WebGPU. That makes it worth a look for voice agents that need fast or local speech. Source: Zyphra on Hugging Face. Source
  Models
• 2026-11-11
  Baidu releases ERNIE 5.5 Flash Preview, a 280B MoE with 14B active parameters and a 1M-token context
  On November 11, Baidu released ERNIE 5.5 Flash Preview, the first post-trained model built on its ERNIE 5.5 Flash Base. It has 280B total parameters with 14B active and a 1M-token context. Baidu pitches it for coding (including long-horizon tasks and game development in Unreal Engine, Unity, and Godot), 3D modeling, knowledge work, and content creation. Every showcased piece was completed inside an agent harness. The API is available now on Baidu's Qianfan platform. Baidu says the weights will be open-sourced under Apache 2.0 soon. Baidu calls it an intermediate checkpoint, and post-training is still ongoing before the official release. No pricing was stated on the announcement page. Source: Baidu ERNIE release page. Source
  Models
• 2026-11-11
  Alibaba opens a two-week Qwen4-Max preview with a 1M-token context, available only through Token Plan
  On November 11, Alibaba Cloud opened Qwen4-Max-Preview (model ID qwen4-max-preview), the next-generation flagship multimodal model in the Qwen family. It takes text, images, and video. It thinks by default, and you can adjust how much it reasons. It supports function calling, web search, built-in tools, and structured outputs. It works with the OpenAI Chat Completions, Responses, and Anthropic Messages APIs. It has a 1M-token context and up to 393,216 output tokens, and thinking and the answer share that limit. Access is through Token Plan only, with no pay-as-you-go option. The preview runs from 10:00 on November 11 to 24:00 on November 25, 2026 (UTC+8) (11:00 AM ET on November 25), and the model is retired when it ends. Source: Alibaba Cloud Model Studio docs. Source
  Models
• 2026-11-10
  Gemini 4 Carbon opens to all paying developers at $2 / $10 per 1M tokens through Nov 30, then $4 / $20
  On November 10, Google opened up Gemini 4 Carbon, its strongest model for long-horizon agentic work. Cyber defenders got it two weeks earlier. It is now in the Gemini API for all paying developers in Google AI Studio and in Gemini for business. It has the same introductory price as Gemini 4 Argon, $2 per 1M input tokens and $10 per 1M output tokens. The introductory period ends on November 30, 2026, and then the price becomes $4 / $20. Google says Carbon uses more tokens than Argon to reason and verify. Google still recommends Argon for most software engineering and knowledge work. Carbon is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search. Source: Google blog. Source
  Models Pricing changes
• 2026-11-10
  Meituan open-sources LongCat-2.5 under the MIT license, with weights on Hugging Face
  On November 10, Meituan open-sourced LongCat-2.5, the model it introduced on November 1 for long-horizon agent and coding work. It is a 1.6-trillion-parameter native multimodal model with about 48 billion parameters active per token. The weights are at meituan-longcat/LongCat-2.5 on Hugging Face under the MIT license. Meituan says that license allows commercial inference deployment and further training without a separate license. Meituan says the model supports a 1M-token context. It uses W4A4 quantization, meaning 4-bit expert weights and activations with native INT4 compute. Meituan says downstream results stay comparable to the unquantized model. The model card points to SGLang-FluentLLM for NPU deployment, and the model supports thinking and non-thinking modes. Most benchmark scores on the model card are Meituan's own runs. Source: tech.meituan.com/2026/11/10/Meituan-LongCat-2.5-Open-Source.html. Source
  Models
• 2026-11-09
  Anthropic releases Claude Fable 5.2 and Mythos 5.2, cutting Fable pricing to $8 / $40 per 1M tokens
  On November 9, Anthropic released Claude Fable 5.2 (claude-fable-5-2) for demanding reasoning and long-horizon agentic work. It has a 1M-token context window and up to 128K output tokens. It costs $8 per 1M input tokens and $40 per 1M output tokens, down from $10 / $50 on Fable 5.1 and Mythos 5.1. Cache reads are $0.20 per 1M, Batch is half price, and Fast mode is $16 / $80. Fable 5.2 is available to all Claude API customers and on Amazon Bedrock, Google Cloud, and Microsoft Foundry. Claude Mythos 5.2 has the same capabilities and prices. It is limited to organizations verified through Anthropic programs such as the Cyber Verification Program. Source: platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2. Source
  Models Pricing changes
• 2026-11-09
  Meta releases Muse Ember, a $0.10 / $0.50 per 1M token model for high-volume agent work
  On November 9, Meta released Muse Ember (muse-ember-1.0), its smallest and most efficient Muse model. It's built for the steps agents repeat all day, like tool calls, coding subtasks, routing, extraction, and long document parsing. It is available on the Meta Model API and Muse Code. It costs $0.10 per 1M input tokens, $0.004 per 1M cached input tokens, and $0.50 per 1M output tokens. It has a 1M-token context window, up to 512K output tokens, and Responses API support. Meta says it matches Claude Haiku 5.5's best score at about a third of the cost. Meta pitches it as a partner for Muse Wonder, where Wonder plans and Ember handles the volume. Source: research.meta.ai/blog/muse-ember. Source
  Models
• Oct 2026
  Anthropic releases Claude Haiku 5.5
  A new small, fast Claude tier — the usual pick for high-volume agent steps. Source
  Models
• Oct 2026
  Claude Opus 5.5 takes #1 on the Artificial Analysis Intelligence Index, with a 20% price cut
  Top model got cheaper, with a larger cache-hit discount — good news for long agent loops. Source
  Models Pricing changes
• Oct 2026
  GPT-6 Sol and Luna halve cost vs GPT-5.6
  Same index scores at about half the cost — check whether your agent builder lets you switch. Source
  Models Pricing changes
• Oct 2026
  Mistral releases Mistral Large 4
  Artificial Analysis calls it the most intelligent model outside the US and China — an EU option for compliance-minded teams. Source
  Models
• 2026-11-05
  Thinking Machines Lab releases Inkling 2 Small, a 594B MoE scoring 31 on AA Intelligence Index
  On November 5, Thinking Machines Lab introduced Inkling 2 Small, a Mixture-of-Experts model with 594B total parameters and 17B active. It supports image inputs, variable effort, and a 1M-token context window. Artificial Analysis reports a score of 31 on the Intelligence Index, which Thinking Machines says would make it the highest-scoring open-weights model developed outside China once weights ship. It is available now through Tinker for inference, sampling, and fine-tuning with closed weights. As of November 5, Thinking Machines plans an Apache 2.0 weight release the following week after external safety testing and red-teaming. Source: thinkingmachines.ai/news/inkling-2-small/ and artificialanalysis.ai/changelog. Source
  Models
• 2026-11-05
  OpenAI releases GPT-6 Cyber for scoped vulnerability research through Daybreak Red
  On November 5, OpenAI introduced GPT-6 Cyber, a cybersecurity model trained to find and validate exploitable vulnerabilities while staying inside researcher-set scope. OpenAI reports that it scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber. In OpenAI's alignment evaluation it avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt, and 100% with that prompt. OpenAI classifies it as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. Researchers used it to develop 948 privilege-escalation exploits in Linux and identify 81 previously unknown high-severity vulnerabilities in Firefox. Mozilla has released fixes for many of those. GPT-6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers, with broader Daybreak Red access planned. Source: openai.com/index/introducing-gpt-6-cyber. Source
  Models
• 2026-11-04
  Ricursive Intelligence releases Ripple-65B Preview, a 65B Apache-2.0 coding model with a reported 80.52 on Terminal-Bench 2.1
  On November 4, Ricursive Intelligence published Ripple-65B Preview, a dense 65B Transformer pretrained on 12T tokens with up to 1,048,576 tokens of context. The Hugging Face model card reports 80.52 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard against larger U.S. open-weight peers such as Reflection Beam 501B, Meta Muse Spark Open 816B, Thinking Machines Inkling 975B, and NVIDIA Nemotron 3 Ultra 550B. Weights are Apache 2.0 (BF16 instruct, NVFP4 instruct, and BF16 base). Hosted access includes OpenRouter via GMICloud at $0.20 / $0.60 per 1M input/output tokens (cached input $0.05). Source: huggingface.co/ricursiveAI/Ripple-65B-Preview. Source
  Models
• 2026-11-04
  Moonshot publishes open weights for Kimi K3.1 Pro on Hugging Face, a 2.8T MoE with a 1M-token context
  Moonshot AI published the open weights for Kimi K3.1 Pro at moonshotai/Kimi-K3.1-Pro on Hugging Face, fulfilling the open-weight date it set when the model launched in apps and API on October 23. The model card describes an open-weight, natively multimodal agentic MoE with 2.8 trillion total parameters and 104 billion activated, 93 layers (69 KDA plus 24 Gated DSA), a 1,048,576-token context, MoonViT-V2 vision (401M parameters), and MXFP4 weights with MXFP8 activations from quantization-aware training. Text and image are supported. The Kimi K3.1-Pro license allows use, modification, fine-tuning, and redistribution with conditions. Model-as-a-Service businesses above $20 million revenue in any consecutive 12 months need a separate Moonshot agreement before commercial use. Products above 100 million monthly active users or $20 million monthly revenue must display "Kimi K3.1-Pro" prominently. Those two conditions do not apply to internal use or to access through Moonshot's official products or certified inference partners. Source
  Models
• 2026-11-04
  Tencent releases Hy4, its first Hy model with visual understanding and computer use, with a reported 1M-token context
  Tencent released Hy4 on November 4. It's the company's first Hy model that can use a computer, combining visual understanding, reasoning, and tool calls for long-horizon agent work. SiliconANGLE reports 770 billion total parameters with 49 billion active and a 1 million-token context window. Tencent says it's available now on Tencent Cloud TokenHub, CodeBuddy, and WorkBuddy, and through OpenRouter and OpenCode, and that it kept the same pricing as Hy4 Preview. It plans to release the weights and system cards in the coming days, with a Flash version to follow. Source
  Models
• 2026-11-04
  Cognition says its SWE-2 Flash coding model runs at one-fifth the price and 3x the speed of SWE-2 in Devin
  Cognition introduced SWE-2 Flash on November 4. It's a 276B-parameter coding model with 12B active, built on Thinking Machines' Inkling-Small with a new mid-training stage before reinforcement learning. Cognition reports Terminal-Bench 4.0 rising from 0.9% for the base model to 20.0%, and FrontierCode 2.0 from 3.2% to 24.0%. It's available now in Devin Desktop and CLI at one-fifth the price and 3x the speed of SWE-2. The weights aren't public. Cognition says it's building a program to share them with trusted partners and researchers. Source
  Models
• 2026-11-03
  Flower Labs releases Endeavor 1.1, what it calls a frontier model you can run on premises with readable thinking traces and EU inference
  Flower Labs introduced Endeavor 1.1, its third model release this year after Lizzy and Endeavor 1.0. Flower reports gains over 1.0 on GPQA (92.0 to 94.4) and HumanEval (98.2 to 99.9). Endeavor 1.1 also adds image input. It can be deployed as a Flower-managed service, in a private cloud, or on premises, with a new option for inference on GPUs located in the EU. Customers can read the model's thinking traces and keep its outputs and traces to train their own models. Access is by request, and the benchmark comparisons are Flower's own. Source
  Models
• 2026-11-02
  Anthropic's message threads beta lets Claude agents send only new messages, and Anthropic says it cuts wait times on long sessions by about 35%
  Anthropic launched message threads in public beta for the Claude Messages API on November 2. An agent sends the full conversation once to start a thread, then sends only the new messages plus the ID of Claude's previous response. Anthropic says median time to first token on requests of 500K to 1M tokens fell by about 35%, and it has already switched threads on for most Claude Code users. It saves time, not money: Claude still reads the whole conversation each turn, so token usage, standard prices, and rate limits are unchanged. Threads are held encrypted in memory, expire 5 minutes after the last request, and work on the Claude Platform, Claude Platform on AWS, Microsoft Foundry, and Google Cloud. Source
  Models
• 2026-11-01
  Meituan releases LongCat-2.5, a 1.6-trillion-parameter native multimodal model built for long-horizon agent and coding work
  Meituan's LongCat team introduced LongCat-2.5 on November 1. It has 1.6 trillion total parameters with about 48 billion active per token. It reads images as discrete tokens in the same stream as text, which Meituan pairs with agentic coding for visual coding, design, and 3D tasks. Training added more long-horizon trajectories and large-scale reinforcement learning on codebase development and full-project builds, including tasks rebuilt from real agent failures. In Meituan's own table, LongCat-2.5 scores 38.5 on Terminal Bench v4.0 (Claude Opus 5 is listed at 53.9), 91.9 on BrowseComp, and 83.1 on MCP Atlas. These are vendor-reported results, mostly run in the Claude Code harness. The announcement post doesn't give pricing or say whether weights will be released. Source
  Models
• 2026-10-30
  Microsoft's Mage-VL-Cowork is a 4B computer-use agent that both clicks and runs code, scoring 66.0% on OSWorld-Verified
  Microsoft's Mage team released Mage-VL-Cowork, a 4B-parameter vision-language model that operates a desktop by clicking and typing from screenshots and, in the same task, running code and terminal commands. Microsoft reports 66.0% on OSWorld-Verified with a 50-step budget, ahead of larger open agents such as Qwen3.6-27B (63.9%) and EvoCUA-32B (56.7%), and says reinforcement learning cut model calls by 13.0% and output tokens by 25.4%. A model this small is aimed at computer-use agents that run on the device itself. The project page links the code on GitHub and the weights on Hugging Face. Source
  Models
• 2026-10-30
  NVIDIA NemoClaw blueprints add Nemotron models, tuning, and a governed runtime to OpenClaw, Hermes Agent, or LangChain Deep Agents
  NVIDIA said on October 30 that NemoClaw is a set of open blueprints that build a full agent stack on top of the harness you already use: OpenClaw, Hermes Agent, or LangChain Deep Agents. Each blueprint combines Nemotron models, NeMo tools for tuning the model and harness, telemetry, and the OpenShell secure runtime. OpenShell limits what an agent can reach in files, networks, credentials, and tools. NVIDIA says LangChain tuned Deep Agents for Nemotron 3 Ultra and got near-frontier results at about 10x lower cost per run than leading closed models, with the tuned profile on GitHub. NemoClaw agents run in the cloud, in data centers, and on DGX Spark and RTX machines, and can work alongside agents like Claude, Codex, or Cursor. Source
  Agent platforms Models
• 2026-10-29
  Mistral releases open weights for an improved Mistral Large 4, at $0.70 / $2.10 per 1M tokens
  Mistral released the weights of an improved Mistral Large 4 (Le Chonk) on October 29, three weeks after its public preview. It's a 1-trillion-parameter, natively multimodal model, and Mistral calls it the best open-weight model for coding. Mistral says it scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task. It also says it scores 41% on Terminal-Bench 4.0, up 13 points from the preview. Mistral also says it scores 51% on the Artificial Analysis Cyber Index and refuses 96.6% of malicious cyber requests on average. It now has Low, Medium, and High (default) reasoning modes. You can use it in Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry, and NVIDIA-hosted endpoints, or download the weights from Hugging Face. The license is a modified MIT license. It's free to use and commercialize, except that Model-as-a-Service providers need a separate commercial license. API pricing is $0.70 per 1M input tokens, $0.07 cached input, and $2.10 output. Source
  Models Pricing changes
• 2026-10-28
  Google announces Gemini 4 Carbon, its strongest model for cybersecurity and long-horizon work, going to cyber defenders first
  Google DeepMind announced Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon. Google says it ranks #1 at 59 on the updated v5.0 Artificial Analysis Intelligence Index. Google also says it leads on SEC-bench Pro (88.2%) for finding vulnerabilities, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching. It's built for long-running agentic engineering work, and Google says internal teams have seen it run for days on multi-step tasks. Access starts this week with a subset of cyber defenders, then widens to trusted partners through Google's Fairwind Program. Developers, businesses, and consumers get it later. It launches at the same introductory price as Argon: $2 per 1M input tokens and $10 per 1M output tokens, with cached input 95% off, rising to $4 / $20 after the introductory period. Source
  Models Pricing changes
• 2026-10-28
  GitHub Copilot Fusion goes GA, routing each task across models with a 30% limited-time discount
  GitHub made Copilot Fusion generally available on October 28. Instead of picking a model, you pick Fusion and an effort level (Standard, High, or Max), and Copilot decides which models to use for each prompt: one model alone, a cheap model that escalates to a stronger one if a quality check fails, or a draft-review-revise loop between two models. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max. It did that at about half the average API cost per task, about half the tokens, and about 40% less time. You're billed for each model it uses at that model's standard rate, and for a limited time paid Copilot plans get a 30% discount on model costs with Fusion. It's on all Copilot plans in VS Code, the GitHub Copilot app, Copilot CLI, and JetBrains IDEs, and admins control which models it can use. Source
  Models Pricing changes
• 2026-10-27
  Google DeepMind announces Genie 4, a real-time world model with native audio that Gemini can steer
  Google DeepMind announced Genie 4 on October 27. It's a world model for real-time interactive simulation that generates audio natively, including ambient sound, music, and interactive dialogue. You can change a running simulation with natural language, such as altering the weather, moving the camera, or adding a person or object. Gemini 3.8 Flash acts as a guiding agent in the loop, breaking goals into steps and checking incoming frames to verify each step. DeepMind also showed Gemini 4 Argon working with Genie 4 to simulate a humanoid robot playing checkers, and it pitches Genie 4 for uses including training environments for robots and agents. For now it's a research preview for trusted testers. It's coming soon to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio. DeepMind says it's still early and experimental, with details that can drift in longer sessions, and all its video and audio carries SynthID watermarks. Source
  Models
• 2026-10-27
  Liquid AI releases LFM3, six open-weight 2B-class models for text, vision, audio, omni, and decisions
  Liquid AI released LFM3 on October 27. It's a family of six open-weight models of about 2B parameters, all built on one LFM3-1.7B-Base backbone: four cover text, vision, audio, and omni inputs, and two are decision models. Liquid AI says LFM3-1.7B outperforms its earlier LFM2.5-2.6B with 35% fewer parameters. The text model has a 256K context window and fits in about 1 GB at 4-bit precision, and every model can be fine-tuned on a single GPU. You can get the models on Hugging Face, the Liquid AI Playground, and the Liquid AI API. Source
  Models
• 2026-10-26
  OpenAI releases GPT-6.1 Terra at $1 / $5 per 1M tokens, built for computer use
  GPT-6.1 Terra is OpenAI's new fast, low-cost model for visual reasoning, computer use, coding, and everyday work. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens, which OpenAI says is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. On OpenAI's own Agents' Last Exam benchmark it scores 15 points higher than GPT-5.6 Terra at medium effort. On Terminal-Bench 4.0 it solves 50.3% of tasks at max effort, up from GPT-5.6 Terra's published 21.5%. These results are self-reported. It's available in ChatGPT Work, Codex, and the API as gpt-6.1-terra, and it launched the same day on Amazon Bedrock (with up to 1M input tokens) and in GitHub Copilot. Source
  Models Pricing changes
• 2026-10-24
  Kanzhun's Nanbeige team releases Nanbeige4.3-33B, an Apache 2.0 model for coding agents
  Nanbeige4.3-33B is a dense 33B model with a 524,288-token context window. It's on Hugging Face under Apache 2.0, so you can use the weights commercially. The team pitches it for coding agents, tool use, and office tasks, and recommends its preserve_thinking option, which keeps earlier reasoning across turns, for multi-turn tool use. In its own comparison it scores 80.9 on Terminal-Bench 2.1 and 65.5 on DeepSWE, ahead of Qwen3.8-27B, though its SWE-Bench Pro score of 60.3 trails Qwen's 61.7. These results are self-reported and not independently verified. A distilled 3B version is planned. Source
  Models
• 2026-10-23
  Moonshot releases Kimi K3.1 Pro, with open weights due by Nov 4
  Kimi K3.1 Pro upgrades Moonshot's 3T-class Kimi K3 and has a 1M-token context window. It's live now in Kimi's apps and API at $2 per million input tokens and $10 per million output tokens. A Highspeed version runs 2 to 3 times faster for $4 and $20. Moonshot says it scored 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4%, but that's the company's own number. The API accepts OpenAI- and Anthropic-style requests, so you can put it behind Claude Code or Codex. Moonshot also says it stripped hacking-related training data out before the open release. Source
  Models Pricing changes
• 2026-10-23
  MiniMax publishes open weights for MiniMax-M3.1-Flash, a 428B MoE agent model with 1M context
  You can now download MiniMax-M3.1-Flash on Hugging Face. It has about 428B total parameters with 23B active per token, takes text, image, and video input, and has a context window of up to 1M tokens. MiniMax pitches it for coding, agentic workflows, and computer use, and its own model card lists 73.7% on DeepSWE v1.1 and 61.7 on OSWorld 2.0. vLLM had serving support on day 0. If you run agents on self-hosted models, it's a big new option worth testing. Just keep in mind that MiniMax ran these benchmarks itself. Source
  Models
• 2026-10-21
  OpenAI models come to Google Cloud's Model Garden
  GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna are in private preview on Google Cloud, with Codex agentic workflows, and general availability is planned in the coming weeks. You pay through Google Cloud Marketplace, and the spend can draw down existing GCP commitments. OpenAI models are also coming to Google Antigravity and the Gemini Agent. If your agents already run on Google Cloud, you'll soon be able to switch models without signing a separate vendor contract. Source
  Models Agent platforms
• 2026-10-21
  NVIDIA adds open Nemotron 3.5 retrieval models for embedding, reranking, and document parsing
  The new Nemotron Retriever models are open and commercially available. NVIDIA's pitch is that better retrieval gets agents to grounded answers with fewer repeated searches and fewer wasted tokens. Worth testing if your agents lean on RAG. Source
  Models
• 2026-10-20
  Mistral releases Voxtral 3 speech models and a voice agent SDK
  Built on LiveKit, the SDK runs transcription, reasoning with tool calls, and speech in one streaming pipeline, and Mistral says each turn completes in under a second. Transcription covers 50+ languages, and the voice agent API starts at $0.04/min with Mistral Small 4. If you're building phone or support agents, note that it runs on Mistral's EU, US, and global endpoints or self-hosted. Source
  Models Agent platforms
• 2026-10-19
  Meta releases Muse Glimmer 30B, an open on-device agent model under Apache 2.0
  It handles tool use, multi-step planning, and failure recovery locally, takes text and image input, and is quantized to about 4-bit so it runs on a single GPU. Meta says it beats Gemma4-31B and Qwen3.6-27B on agentic benchmarks. Worth a look if you want your agent runs to stay on your own hardware. Source
  Models
• 2026-10-17
  Replit makes OpenAI the default model for Replit Agent
  CEO Amjad Masad pointed to cost, the partnership, and performance, and Claude models are still selectable. If your build depends on how Claude behaves, pin the model rather than trusting the default. Source
  Agent platforms Models
• 2026-10-16
  Reflection releases Beam, an open-weight 501B MoE model (23B active) under Apache 2.0
  It's built for coding and long-horizon agent work, with weights on Hugging Face and 40+ launch partners. If you self-host your agent backend, this is a permissively licensed option worth a look. Source
  Models
• 2026-10-15
  Meta releases Muse Spark Open (816B MoE, 42B active, 1M context)
  Open-weight model built for coding and long-horizon agent work; available via OpenRouter, Together, Fireworks, Groq and others — a new self-host/API option for agent backends. Source
  Models
• 2026-10-15
  NVIDIA releases Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B A9B
  New open-weight NVIDIA models to watch for cheap self-hosted agent backends. Source
  Models
• 2026-10-13
  Gemini 4 Argon GA: $2 / $10 per 1M tokens through Nov 30, then $4 / $20
  Google’s new frontier model for long-horizon agent tasks; intro pricing ends Dec 1 — watch agent builder model pickers and Gateway credit burn. Source
  Models Pricing changes
• 2026-10-13
  n8n adds Gemini 4 Argon for Agents, Assistant, and MCP workflows
  One of the first builder platforms with Argon at launch; Cloud users can run it via Gateway credits without a separate Google key. Source
  Agent platforms Models
• 2026
  GPT-6 Astra launched at $10 / $50 per 1M tokens
  Ties the leaders on Artificial Analysis indices at lower cost than Claude Fable 5.1. Source
  Models
• 2026
  Gemini 3.8 Flash GA: $0.75 / $3.75 per 1M tokens until Dec 31, 2026, then $1.50 / $7.50
  Launch pricing ends Jan 1 — agents running on Flash will roughly double in model cost. Source
  Models Pricing changes