AI models roundup, November 2026
Kimi’s weights arrived early and 720p video fell to 2 cents a second. One new open image model excludes the US and the EU.
On this page
On November 2, two days before the date it had promised, Moonshot put the full weights of Kimi K3.1 Pro on Hugging Face: 2.8 trillion parameters in about 1.56 TB of files. Three days later OpenAI launched GPT-6 Cyber, a model most readers cannot call without an application. Those two releases sum up the first week of November. Plenty arrived, and the most useful question for each one is not how good it is but who is allowed to use it, where, and on what terms.
This is edition five of the monthly roundup, and the second live one. It opened on November 6, covering everything since October’s final sweep, and it is updated every Friday and summarised once the month closes. The method has not changed: every date, price and spec below was checked against a primary source this week and linked in place, vendor benchmarks are labeled as such, and the non-events are named at the end. Twenty-one entries landed between October 31 and November 6, and four October releases we missed open the ledger.
The TL;DR so far: three things to act on
If you make short video clips, price them again this week, because the cheap end moved. Google’s Gemini Omni 1.2 Flash costs $0.07 a second at 720p, 30% below the version it replaces, and MiniMax’s H3 Lite costs $0.02 a second at 768P through Replicate. If you make logos, icons or simple animations, GPT Image 2.6 now returns SVG files and animated GIFs at GPT Image 2.5’s prices. That turns an image model into something that hands you an editable vector instead of a picture of one. If you plan to download any of this week’s open models, read the license before the model card. Of the week’s seven open releases, three are plain Apache 2.0, and one image model excludes the US, the EU, the UK and Korea unless you sign a separate deal.
Text: the week’s biggest model is a download, and the newest frontier one needs an application
Kimi K3.1 Pro is the week’s most capable model anyone can use. It scores 46.5 on Artificial Analysis’s Intelligence Index, and Runware began serving it on November 4 at $2 per million input tokens and $10 per million output, with a 1M context. For scale, the same index puts GPT-6.1 Terra at 51 for half that price, so this is not a new mid-tier pick. What is new is that the weights are public, which the open-weights section covers.
OpenAI’s GPT-6 Cyber arrived on November 5. The model page describes a model for investigating vulnerabilities and reproducing them “in an authorized test environment,” and the pricing page lists $20 in and $100 out per million tokens, with a 1,050,000-token context. An organization needs approval through OpenAI’s Daybreak program before it can call it. It is the third model in nine days released only to approved groups, after Gemini 4 Carbon and GPT-6 Rosalind. On the same changelog, OpenAI set February 1, 2027 for the removal of gpt-5.5-cyber-preview.
Tencent’s Hy4 reached OpenRouter on November 4 at $0.417/$1.25 with a 1M context and image input. The listing carries a 50% discount, so the full price would be about $0.83/$2.50, and it scores 34.6 on the index. Tencent’s Hugging Face repository for it is not public yet. Z.ai’s GLM-5.3-FlashX-Legal (November 2) is a legal-domain variant at $0.37/$1.25, confirmed on Z.ai’s pricing page, and Ricursive’s Ripple-65B Preview, covered below, is the week’s other new name at $0.20/$0.60 hosted.
Decision models, which answer with one of the options you give them and a probability instead of prose, kept arriving. Morph’s Reflex (November 3) costs $0.042 per million input tokens with free output and a 65K context. Fastino added four smaller sizes of GLiDE between October 31 and November 4, from 0.4B to 9B, after a 27B on October 19 that our October edition missed. A note on the numbers: on Artificial Analysis’s Decision Index, GLiDE v1.2 scores 65.0, 0.2 behind d1-ultra’s 65.2 for first place among dedicated decision models. The higher figures on Fastino’s model cards, up to 72.49, come from a different benchmark that shares the name, so the two cannot be compared. Reflex is not on the board yet. The benchmark guide explains why the source of a score matters more than its size.
Image: GPT Image 2.6 draws vectors and GIFs at the old price
OpenAI’s GPT Image 2.6 Sunburst and its faster sibling Flare landed on November 4. The changelog lists three additions: adjustable complexity, SVG generation and animated GIF generation. Complexity is low, medium or high with no separate fee. GIFs run from 2 to 16 frames at 20 to 2,000 ms each, and SVG works on new generations only, one at a time. Token prices match 2.5: $8 per million image input tokens, $30 per million output.
Per image, the clearest rates are on Replicate: $0.05 at low quality, $0.09 at medium, $0.128 at high and $0.50 at the maximum, with GIF and SVG billed by output tokens. Runware added both models a day later. The often-quoted “about 50% faster” for Flare comes from OpenRouter’s description, not from OpenAI, so treat it as a gateway’s claim until OpenAI repeats it. The comparison below uses our standing poster prompt, which tests lettering, against 2.5.
Minimalist Swiss-style concert poster on off-white paper. Large bold sans-serif headline reading "NIGHT SIGNAL" stacked over a thin rule, smaller caption below reading "LIVE - SEPT 14 - HALL C". A single geometric orange semicircle in the lower third, generous white space, printed-ink texture.
Two smaller items. Topaz Labs’ Wonder 4 (November 3) restores and upscales photos up to 6x in one click, and on Runware it costs $0.0428 for an output up to 8 megapixels, rising to $1.37 at 256 megapixels. Krea 3, which launched in October, gained a Runware route on November 4 at $0.025 per megapixel, half price until November 10. On the way out, xAI retired its grok-imagine-image quality and pro ids on November 2, per its release notes, and now serves those requests with a cheaper 2.0 tier. The open image model of the week, P-H3-Image, is in the open-weights section because its license is the story.
Video: Google cut Omni by 30%, and 720p now starts at 2 cents
Gemini Omni 1.2 Flash went generally available on November 5. Google’s post says it dropped the price 30% from Omni 1.1, “making it available at $0.07 per second for 720p video.” Clips run 3 to 15 seconds and extend to 30, from 360p to 4K, where 1080p and 4K are upscaled. Runware matches Google at $0.07 and charges $0.02 at 360p for drafts; Replicate charges $0.105 for the same 720p second. Google’s changelog says requests to Omni 1.1 move to 1.2 on December 3.
Kuaishou’s Kling 4.0 Flash reached Runware on November 4, generating 3 to 20 seconds at 720p for $0.094 a second, or $0.126 when you feed it a video, with 15% off until November 10. Our October edition missed MiniMax’s two cheaper H3 tiers from October 29. MiniMax’s own prices are $0.025 and $0.04 a second for H3 Lite at 480P and 768P, and $0.04 and $0.06 for H3 Highspeed. Replicate now lists half of that: $0.0125 and $0.02 for Lite, and $0.03 flat for Highspeed, with no promotion named on the page.
Pruna’s P-Video-2 Pro Replace (November 5) does a different job: it takes one or two photos of a person and puts them into an existing video, keeping the camera, the motion and the soundtrack, and matching the lips to the audio. Source clips can run from 1 to 120 seconds. Pruna and Runware list $0.05 a second at 480p up to $0.18 at 4K, with Runware at half price until November 19, and Replicate lists $0.025 to $0.09. Use it with photos of people who have agreed to appear. Below, three of the week’s video tiers answer our standing establishing-shot prompt.
Slow aerial push in over a coastal cliff road at golden hour, waves breaking on rocks below, a single car travelling along the road. Smooth continuous camera movement, warm cinematic colour grade.
Prices for every tier in this section move often, so the text-to-video guide is the place for the method rather than the numbers.
Audio: sound effects that loop, and open voices for India
ElevenLabs made its third sound effects model, eleven_text_to_sound_v3, generally available on November 2 and the default across its products, with a loop option for seamless background audio. The changelog also says v2 “is deprecated and will be removed on March 1, 2027.” Hume launched an Audio Decisions API the same day: you ask plain-language questions about an audio file, such as whether a caller sounds frustrated, and get probabilities back. It is the decision-model idea applied to sound. Hume’s older TTS and EVI APIs end on November 13, and their documentation has already left Hume’s site.
Maya Research shipped two open models for Indian languages. Maya 2 Native (October 26, missed in October) is a streaming text-to-speech model for 11 Indian languages and English, with 24 kHz output, 20 built-in voices and voice cloning, under Apache 2.0. Maya Transcribe (November 2) is a Hindi and Hinglish transcriber on a Qwen3-ASR 0.6B base, also Apache 2.0, with a self-reported 9.36% word error rate on the Kathbath Hindi test. Both are small enough for a laptop.
Open weights: Kimi kept its date, and the licenses got longer
Moonshot’s Kimi K3.1 Pro repository appeared on November 2, inside the window its launch post promised. It is 2.8 trillion parameters with 104 billion active per token, which makes it a download for a data center rather than a desk. The license is a modified MIT: free to use, but a product with more than 100 million monthly users or $20 million in monthly revenue must show “Kimi K3.1-Pro,” and a hosting company above $20 million a year needs a separate agreement.
Ricursive’s Ripple-65B Preview (November 4) is the cleanest release of the week: a 65B dense model trained from scratch on 12 trillion tokens, with a 1M context, under Apache 2.0. Its launch page claims 80.52 on TerminalBench 2.1, a vendor number with no independent score yet. inclusionAI’s ZwZ2, which our October edition missed, adds visual reasoning to Qwen3.5 and 3.6 bases at 4B, 9B and 27B; the 9B card reports a gain from 51.63 to 56.59 across 25 benchmarks, its makers’ figure.
Pruna’s P-H3-Image (November 3) is an open image model for generation and editing, built on MiniMax H3, that renders in 8 steps. Its model card says the MiniMax H3 Community License “excludes the European Union, the United Kingdom, the Republic of Korea and the United States” unless you hold a separate license from MiniMax, and that products built on it must show “Powered by MiniMax H3.” The same base license sits under Pika’s H3 Flex weights from October, so it is worth reading once before downloading anything in the H3 family.
Local-runnable: an open H3 image model needs about 20 GB
stable-diffusion.cpp added P-H3-Image in master-1049 (November 4), alongside a second, higher-resolution pass for H3. The smallest diffusion file is 13.36 GB, and with the 3.6 GB text encoder and a 3.2 GB VAE the total lands near 20 GB before anything is generated. master-1051 and master-1052 added packed 4-bit models and faster 4-bit kernels on NVIDIA cards. Since master-1044 the Linux CPU and Vulkan builds target Ubuntu 26.04 only, so older distributions need an older build.
Ollama v0.40.15 added two families this week. Xiaomi’s MiMo V2.6 is a 9B model for code and agent work under MIT, 6.6 GB at 4-bit; its weights date from September, so the news is that it now runs with one command. AutoTrust’s JEV decision models come at 9B (11 GB), 27B with vision (21 GB) and 35B (26 GB) under Apache 2.0, on Apple silicon only in this release. Perplexity’s pplx-decider-flash, a 4B decision model from October, also arrived on Ollama as MLX tags from 4.6 to 9.1 GB, so Apple silicon only too.
llama.cpp’s v0.9.0 (November 4) added Chatterbox speech, NVIDIA’s Nemotron 3.5 Transcribe and the Mistral Large 4 architecture, and later builds added Ripple and Fish Audio’s Drama 3. Whether any of this fits your machine is the same arithmetic as before; the RAM and VRAM guide does it.
Skipped on purpose
What we cut, and why. Crouton Alpha: a free, unnamed stealth model on OpenRouter. OpenRouter Auto and Auto Code: routers, not models. Parakeet Ultra on OpenRouter: a new route at $0.001 a minute for a model Moondream launched on September 22. Gemini Omni 1.1 on OpenRouter: a new route for an August model that retires into 1.2 on December 3. Kimi K3.1 Pro on Ollama: a cloud-only tag. Fireworks Forge: faster serving, background jobs and desktop plans, but no new model. Gemini 3.8 TTS voice remixing, Suno’s MIDI input, ElevenLabs Realtime: features, not models. Claude platform updates: message threads, watermarking and IP allowlists, no new model.
One correction to October. That edition skipped Canva Craft because Canva’s page was unreachable and there was “no API or price yet.” In fact Canva’s developer blog opened an image generation API with the craft_v1_2k model on October 28. It is “currently in preview and can change without notice,” and apps that use it cannot pass Canva’s public-app review yet, so it stays out of the ledger until it leaves preview.
Watching: October’s open items and November’s dates
New to the list this edition: Cloudflare’s Workers AI retirements on November 15, which move MeloTTS to Deepgram’s Aura-2 and the old Stable Diffusion models to FLUX.2 klein; BytePlus shutting down Seedance 1.5 Pro on November 11 with no automatic migration; Liquid’s d1 leaving OpenRouter on November 20, with d1.1 as its successor; and Google rerouting gemini-2.5-flash to 3.5 Flash-Lite and gemini-2.5-pro to 3.8 Flash from December 1, per its deprecations page. Anthropic’s deprecations page still shows Claude Sonnet 4.5 retiring on November 30. Between editions, the sortable model comparison table stays current.
Update log
November 6: edition opened, covering October 31 to November 6, plus four October releases the October edition missed: ZwZ2, Maya 2 Native, H3 Lite and H3 Highspeed.
Disclaimer: The license descriptions in this post are plain-English summaries, not legal advice. Model licenses change between releases, and what counts as permitted use depends on your situation. Read the license text that ships with the model, and consult counsel before relying on it in a product. Details reflect sources available as of November 2026.
