AI models roundup, October 2026

On September 28, a few hours after our September edition went live saying Claude Sonnet 5.5 would “follow in the coming weeks,” Anthropic shipped it. The next morning, at DevDay, OpenAI replaced a model it had launched seven days earlier. A monthly digest written once, at the end, was always going to be a day late somewhere. So this one is written as the month happens.

This is edition four and the first live one. It opened on October 12, covering everything since September’s sweep closed. It is revised weekly, most recently on October 23, and gets a summary on November 1 once the month is over. The method has not changed: every date, price and spec below was checked against a primary source and linked in place, vendor benchmarks are labeled as such, and the loudest non-events get named at the end. Seventy-three entries from forty-five vendors made the ledger in twenty-six days. Three of them should change decisions already.

The TL;DR so far: three things to act on

If you buy mid-tier tokens, Claude Sonnet 5.5 is still the pick, even after Gemini 4. Sonnet scores 56 on Artificial Analysis’s index at $2 per million input and $10 per million output. Gemini 4 Argon arrived on October 13 at the same price, scores 53, and doubles in price on December 1. For classification and routing, Claude Haiku 5.5 still costs 10 cents a million tokens in and scores 43. If you make images on your own machine, FLUX 3 Image Klein is a new open FLUX that fits an 8 GB graphics card. It is 5.6B parameters with downloadable weights, free for commercial use under $5M in yearly revenue, and costs $0.009 a 1K image on BFL’s API if you would rather not run it. If any of your prompts end in a yes, a no, a label or a number, a decision model now answers them for $0.01 to $0.15 per million tokens with no output charge. The chart below lists sixteen hosted ones that arrived or changed price this month, and nine of them come with weights you can download. OpenAI’s version left beta on October 22 at the same price. Everything else so far is either a price cut, a preview, or a promise with a date on it.

Text: Gemini 4 arrived at Sonnet’s price and scores three points lower

Claude Sonnet 5.5 is the month’s first entry by a few hours and its most useful one. Anthropic’s launch page claims output more than 30% faster than Sonnet 5 and up to 30% fewer tokens per task; the pricing page keeps $2/$10 across the full 1M context, and on October 7 cut cache reads from $0.20 to $0.10 per million. On the independent index it scores 56 at maximum effort, 52 at the next tier down and 47 at high, against Opus 5.5’s 58 at $4/$20. If your work does not need the last two points, the bill halves. Read the vendor table the way vendor tables deserve; the independent number stands on its own.

OpenAI’s DevDay on September 29 brought GPT-6.1 Sol, which the API changelog describes as a GPT-6 Sol upgrade with “performance close to Astra at a lower price.” GPT-6 Sol had shipped on September 22. The new model keeps the $2/$10 rate up to 272K input tokens, doubling input and adding half to output past that, and scores 52 on the index, one behind Astra at a fifth of the price. It cannot turn reasoning off. The same week added an Ultrafast service tier at six times list, first on Astra and from October 8 on 6.1 Sol, and on October 1 the deprecations page scheduled gpt-5.1, gpt-5.3-codex and gpt-5.4-nano for April 1, 2027, and tts-1 and tts-1-hd for January 6. One note on names: the “GPT-6.1 Sol Pro” and “GPT-6 Astra Pro” ids on gateways are the same models with a reasoning mode set, not separate releases.

Claude Haiku 5.5 (October 7) is the first Haiku with an effort parameter and adaptive thinking. Anthropic’s launch page prices it at $0.10/$0.50 for prompts up to 100,000 tokens and $0.50/$2.50 above, with the whole request billed at the higher tier once it crosses the line, and cache reads at one cent. It scores 43 on the index, where Gemini 3.8 Flash is 41 and GPT-6 Luna is 38. Further down, Mistral Large 4 (October 6) is a 1-trillion-parameter mixture of experts with 52B active, served as a preview at $0.68/$2.09 on Mistral’s API and OpenRouter against a $1.36/$4.18 list price; it scores 38, and its weights are the subject of the open-weights section. Two more previews round out the tier: Reflection AI’s Beam (October 5), 501B total and 23B active with vendor-run scores of 80.9 on SWE-bench Verified, waitlist only until its weights and API arrived on October 16, and inclusionAI’s Ling 3.1 Flash, a 560B/25B-active model served free on OpenRouter since October 5 with no weights. And one correction: StepFun’s Step 5 Preview, a 600B model from September 18 that our last edition missed, reached OpenRouter on October 8 at $1/$2.70 and scores 44.

The second week brought the model September kept waiting for. Gemini 4 Argon went generally available on October 13, per Google’s launch post, with a 1M-token window shared by input and output and text, image, video, audio and PDF input. Google’s pricing page lists $2/$10 “through November 30, 2026” and $4/$20 from December 1, with batch at half price and no free tier. The index puts it at 53: above GPT-6.1 Sol’s 52 and below Sonnet 5.5’s 56, both at the same $2/$10 today. Meta’s Muse Wonder (October 12) is closed and scores 52 at $4/$20 on OpenRouter; its larger sibling, Muse Wonder Full, scores 56 but runs only inside Meta’s Muse Code subscription, so there is no rate card to compare.

Further down the price list, MiniMax M3.1 Flash landed on October 19, the day of this pass, at $0.15/$0.60. MiniMax says the weights “will be released later.” It is free on OpenRouter for now and scores 42, a point under Haiku 5.5 at half again Haiku’s input price. DeepSeek’s V4.1 Flash 1017 (October 17) is a refresh with “shorter outputs than the previous version,” MIT weights on Hugging Face, and a score of 45; DeepSeek opened its search and fetch APIs the same day at $0.75 per 1,000 calls. Two newcomers round it out: Unbiased’s Pareto 26.10 (October 15), a system that blends several models behind one endpoint at $0.80/$3.20, and Elorian’s Glimpse (October 16), a closed visual-reasoning model at $0.60/$3. The cheap tier now has four models between 41 and 45 on the index: Gemini 3.8 Flash, M3.1 Flash, Haiku 5.5 and DeepSeek’s refresh. At that spread, test price and latency on your own prompts before the score.

The fourth week added one cheap model with an unusual engine. Inception’sMercury 3(October 19) is a diffusion language model: it refines a whole draft in parallel instead of writing one token at a time. Inception’smodel page lists $0.20/$0.70 per million tokens, halved during the launch, with a 1M-token context, and the company says it “outperforms GPT-6 Luna and Claude Haiku 5.5” while running four to eight times faster. That is a vendor claim; the independent index has not scored it yet. Recursive’s Flint, a coding model with open weights, is covered in the open-weights section. Anthropic and Google shipped no new text model this week.

The real text story is a category that does not write text. A decision model takes your state and a typed question, and returns a probability, a choice among the options you listed, or a score on a scale you defined, in one forward pass. September had two of them. On October 6 OpenAI put GPT-6 Luna behind a Decisions API in public beta: $0.10 per million input tokens, “no cache-read, cache-write, or output-token charges,” and answers “about 10x faster than the Responses API.” Cloudflare released Clef 27B and Clef-flash 9B on October 1 under Apache 2.0, then Clef-omni on October 9, a 30B/3B-active model that decides over text, images, audio and video, while cutting hosted Clef-flash from $0.09 to $0.038. Microsoft’s Decision-1 (October 9) is a post-trained Qwen3.5-9B at $0.042 with output free; Perplexity’s Decider v1.1 27B (October 7) is Apache 2.0 at $0.02; Liquid AI’s d1-3B and d1-omni-600M (October 7) answer in 8 ms a question on an RTX 4090. The app’s Decision model type runs most of them, local and hosted, through the same wire; Clef’s page is the place to start.

The second week added eight more and moved the floor. Respan’s Span-02 (October 15), a fine-tune of Gemma 4 26B with open weights, and Perplexity’s Decider Flash v1.2 (October 16), a 4B model under Apache 2.0, both list at $0.01 per million input tokens; the 27B Decider v1.2 replaces v1.1 at $0.02. Liquid’s d1-ultra (October 15) and TypeSafe’s Jev 1.14 (October 14) are closed at $0.042. Nace.AI listed Drex v2.0 (October 12), a 35B/3B-active open model at $0.06, then Drex Max on October 19, served on its own platform. ScaleDown’s Classify-1 (October 13) is $0.04 and OpenRouter-only. The category also has an independent scoreboard now: on Artificial Analysis’s Decision Index, which mixes classification, tool calling, knowledge and reasoning, d1-ultra leads at 65, nearly eight points clear of Jev; GPT-6 Luna’s Decisions API scores 48.5, below Microsoft-Decision-1 at less than half the price. Span-02, Decider v1.2 and the Drex models are not on it yet.

On October 22 theAPI changelogtook the Decisions API out of beta: it is “now generally available with gpt-6-luna,” takes up to 400 questions a request instead of 200, and adds prompt caching, images inside a question’s instructions, and structured input. Theguidekeeps $0.10 per million uncached input tokens with output free. Two neighbours arrived the same week. Cloudflare’sClairic(October 19), in preview on Workers AI, is the extraction sibling of a decision model: Cloudflare calls it a model that “answers with values instead of text,” returning typed fields with a confidence score for $0.10 per million input tokens and no output charge, with weights promised “soon.” Baseten’s Mammoth 1 and Elephant 1, decision models on Qwen3.5 27B and 9B bases withMIT weights, reached OpenRouter for free on October 21. Drex Max, covered last pass, reached OpenRouter the same day at $0.06.

Image: Google halved its Flash price, and FLUX shipped one you can run

Nano Banana 2.1 (October 6) is the image release that moves a default. Google’s pricing page bills image output at $30 per million tokens, which works out to $0.0336 for a 1K image, $0.0504 at 2K and $0.113 at 4K, against $0.067, $0.101 and $0.151 for Nano Banana 2. The model page takes up to 14 reference images, video and PDF input, search grounding and three thinking levels. On Arena’s text-to-image board it sits sixth, below OpenAI’s two GPT Image 2.5 models, GPT Image 2, Grok Imagine Image 2.0 and MAI-Image-2.6; on the multi-image edit board it is fourth. The same changelog entry deprecates Nano Banana 2 with no shutdown date. The cost arithmetic in what one image costs now has a new floor at the Flash tier.

Ideogram 4.5 and its Precise Edit variant (September 30) have no launch post, only an X post that promises “open weights soon.” On Runware both run from $0.008 to $0.20 an image across four quality tiers, with up to five source and reference images. Alibaba’s Qwen-Image-2.1-Turbo (October 9) is the 8-step checkpoint of September’s Qwen-Image-2.1, $0.0141 an image hosted and downloadable from Hugging Face under the same research-only license as its parent. SperidLabs’ Iris-3B (October 5) is the odd one out: a 3B generator that works in pixel space with no VAE, Apache 2.0, whose own paper reports no significant benefit from the design; stable-diffusion.cpp added it four days later.

Black Forest Labs shipped twice, and the first pass missed one of them. FLUX 3 Image (October 1, per BFL’s release notes) is closed and API-only, takes up to ten reference images and bounding boxes inside the prompt, and costs $0.048 for a 1K image and $0.607 at 4K. Its small sibling matters more to most readers. FLUX 3 Image Klein (October 15) is a 5.6B model with downloadable weights. BFL’s launch post says they need “11.2 GB in BF16 and 5.6 GB in FP8,” and that an INT8 build runs on RTX cards “with as little as 8GB of VRAM.” Hosted, it runs from $0.008 at 768px to $0.036 at 4K. The FLUX Kommunity License allows free commercial use under $5M in yearly revenue; Klein Base, a 50-step version, is the one to fine-tune.

Product still life of a brushed-titanium espresso tamper standing on wet black slate, a scatter of coffee beans around it. Single hard light source from the upper right casting a crisp shadow, cool grey studio background, macro detail on the metal grain and the water droplets.

Nano Banana 2.1via Runware

FLUX 3 Imagevia Runware

FLUX 3 Image Kleinvia Runware

Real outputs from each model’s page, generated through Runware. Every model in the catalog gets this same prompt and keeps its first result; click a name for the full set.

Krea 3 (October 15) is an editing model first. Krea’s post says most edits come back “in under 8 seconds, end to end, at native 2K,” at $0.06 an image at 2K and $0.03 at 1K, on OpenRouter, Replicate and Krea’s own app, with no weights. Qwen-Image-2.1 Pro and Turbo reached OpenRouter on October 16, a new route for models covered earlier rather than a new release.

Two corrections to September. ByteDance’s Seedream 5.0 Flash reached Runware on September 28 at about $0.018 an image with up to ten references, and Tencent’s HY Image 3.5 Preview launched September 22 at $0.024 for a 1K or 2K image, closed weights, and reached OpenRouter on October 5. Neither was in last month’s ledger; both should have been. One retirement: Amazon’s Nova Canvas reached end of life on Bedrock on September 30, as scheduled.

The fourth week’s image release came from Bria. FIBO 2.0 and FIBO 2.0 Edit reachedRunware on October 20 at $0.04 a run, and the weights are on Hugging Face asFibo-2.0-dev and an Edit Turbo build, under Bria’s own license rather than an open-source one. stable-diffusion.cpp added both a day later. Read the license before you ship anything made with them.

Video: two models built on others’, and Kling 4.0 without an API

HeyGen Video 1.0 (September 30) is, by HeyGen’s own catalog page, “built on MiniMax H3 and post-trained by HeyGen”: text, first-frame and reference-driven clips of 5 to 15 seconds at 480p or 768p, with dialogue, ambience and effects in the same call, at $0.02 a second and $0.01 through October. xAI’s Grok Imagine Video 1.5 Lite is a distilled tier of September’s 1.5 at $0.02 a second for 1 to 15 seconds with native audio; xAI published no post and no date, so the ledger uses October 1, the day it first appeared. The price of a first draft keeps falling, which is the whole argument of draft mode.

The rest is previews. Kuaishou opened Kling 4.0 Flash early access on September 28 to top-tier subscribers inside its own app, and says the full 4.0 “will officially launch in October” with no day, no API and no price. Reka’s Rho-1 (October 5) generates text, images, video and robot actions from one model, by email request. Tavus’s Griffin (October 1) is a full-duplex video-to-video conversation model, waitlist only. Two dates matter more than any of them: Google’s deprecations page shuts down every Veo 3.1 preview endpoint on October 22, pointing at Gemini Omni 1.1 Flash, and Nova Reel left Bedrock with Nova Canvas on September 30.

Kling 4.0 kept its October promise, halfway. As of this pass, Kling’s homepage leads with “All-New Kling 4.0” and a Create Now button, but its API update notices list no 4.0 model, so developers still have no endpoint and no rate card. Two corrections to the first pass: HeyGen added 2K and 1080p tiers to Video 1.0 on October 5 and 8, per its changelog, and Runware lists all four tiers at half price through October 31. Pruna’s P-Video-2 Pro reference mode (October 2) takes up to nine images, three videos and three audio clips and bills from $0.015 a second at 480p.

Kling 4.0 reached more people this week, still not through Kling’s own API. On October 23Luma put it inside its app, “available to every Luma customer.” fal lists Kling 4.0 endpoints with aplayground priceof $0.157 a second at 720p and $0.235 at 1080p, but says “API access is coming soon,” andRunware’s page still reads “Coming Soon.” Kling’s own update notices list no 4.0 model. October 22, the shutdown date for every Veo 3.1 preview endpoint, has passed with Google’s deprecations page unchanged; if you still call one, move to Gemini Omni 1.1 Flash.

Audio: speech you direct in brackets, and a 17 MB transcriber

Microsoft AI shipped three audio models on October 1. MAI-Voice-2.1 and its Flash variant are text-to-speech at $22 and $15 per million characters, one voice across 23 languages, cloning from a few seconds of audio; MAI-Transcribe-2-Streaming is the company’s first streaming speech-to-text, first partials in just over 100 ms, at $0.54 an hour as an introductory rate through 2026. That is more than five times what its batch sibling charged in September, the usual premium for streaming, and above the $0.36 whisper-1 charged in the transcription shootout. Amazon’s Nova 2.5 Sonic (October 5) is speech-to-speech for voice agents in seven languages at the Nova 2 Sonic price. Suno’s Speech beta (October 1) puts a spoken voice and its backing music in one track, inside Suno, with no API.

Two releases are small enough to miss. Cactus Compute’s Whistle (September 30) is an on-device transcription model in a single 16.9 MB file, Apache 2.0, with word timestamps and keyword biasing in English, German, French, Spanish, Italian, Dutch and Polish, running on CPUs down to a watch. Mistral’s Voxtral Mini 4B Realtime Arabic (October 8) streams Arabic dialects and code-switched speech under Apache 2.0. On the hosted side, OpenRouter added ElevenLabs on October 7: Eleven v4 and v3 at $0.04 per thousand characters, v4 Turbo and Flash v2.5 at $0.02, Scribe v2 at about $0.11 an hour, all half price through October 19. And xAI retired Grok Voice Transcribe 1.0 on October 2, routing requests to 2.0 at the same rate.

The second week’s voice model is Fish Audio’s Drama 3 (October 13). You direct it with plain-language notes in square brackets, in any of 83 languages, and Fish’s launch post puts the weights on Hugging Face under a new Community License: free commercial use below US$500,000 in yearly revenue, a separate license to resell it as a speech service, and no cloning of a real person’s voice without consent. The API costs $10 per million characters until November 12, “80% off the $50 list price.” HeyGen opened a Voice API on October 9 and added a Turbo tier on October 13; OpenRouter lists the two at $15 and $4 per million characters. xAI’s Grok Voice 3.0 (October 14, per its release notes) replaced 2.0 for speech-to-speech at the same $0.08 a minute. On October 16 OpenAI’s changelog opened prompt-to-voice to all API customers: describe a voice in text and GPT-Live speaks with it, with creation free. AssemblyAI’s Universal-3.6 Pro appeared on OpenRouter on October 14 at $0.30 an hour, but AssemblyAI’s own pages still list 3.5 Pro, so its accuracy claims stay unconfirmed. The ElevenLabs half price on OpenRouter ended on October 19.

The fourth week belonged to Mistral.Voxtral 3 (October 20) is a speech family in public preview:Voxtral Mini TTS 3at $13 per million characters, with preset voices and cloning;Mini Transcribe 3at $0.0035 a minute, a realtime version at $0.008, andSmall Transcribe 3at $0.004; and a voice-agent API from $0.04 a minute. Mistral says the transcriber “outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe,” a vendor claim we have not tested, and there are no weights this time. At $0.21 an hour, Mini Transcribe 3 undercuts the $0.36 whisper-1 charged inour shootout.

Smaller items. Amazon puta new model behind Polly’s Generative engineon October 20, with 14 new voices plus Hindi and Finnish. ElevenLabs’changelog opened a Video to Music API on October 19 (upload up to ten videos, get a fitting score with optional sound effects and stems) and an enterprise realtime speech translation service, neither with a public price.Cartesia sunset Sonic-2, Sonic-turbo and an October 2025 Sonic-3 snapshot after October 20. And Gradium’s Phonon speech models, covered under local below, are the smallest audio release of the month.

Open weights: the big ones arrived, and none beat September’s leader

Mistral’s announcement says it in one sentence: “We will release the weights by the end of the month.” The docs tag the model open and name no license, and on October 19 Mistral’s Hugging Face organization still has no Large 4 repository. Reflection AI kept its word: on October 16 its post said “Today, we have released Beam’s model weights” under Apache 2.0, ungated, with llama.cpp support merged the same day. Ideogram still says “soon,” and Black Forest Labs has still not dated FLUX 3 Dev, shipping the smaller Klein instead. The hero strip above marks October 31 for Mistral.

What actually shipped with a download link is smaller and more specific. Aleph Alpha’s Kolibri-1 (October 2) is a 78B mixture of experts with 3.5B active, trained for German and English, Apache 2.0, 78 GB in FP8 and sized for one H200. JetBrains’ Mellum2.1 (October 8) is a 12B/2.5B-active coding model under Apache 2.0 with a 7 GB MXFP4 build. Google’s EmbeddingGemma 2 (October 6) embeds text, images and audio from 740M parameters, Apache 2.0, 191 MB quantized for text only. Then the decision models: Clef 27B, Clef-flash and Clef-omni under Apache 2.0, Perplexity’s Decider v1.1 under Apache 2.0, Liquid’s d1 pair under its LFM Open License, and Nace.AI’s Drex v1.5 under RAIL-M. Also on Hugging Face this fortnight, listed for the record: Microsoft’s AesCode slide-and-poster models, Meta’s MetaEncoder-30B scorer, NVIDIA’s PixelUMM, Amazon’s ALoDLM-8B, Tencent’s Youtu-Parsing-Omni, LightOn’s OCR-3 family, and a crowd of community decision fine-tunes from H2O, vLLM’s semantic-router team and others.

The second week’s downloads are far bigger. Meta’s Muse Spark Open (October 15) is an 816B mixture of experts with 42B active and a 1M-token context, released “under the Muse Model License with limited commercial terms” behind a manual approval gate on Hugging Face. Meta calls it “the most capable US open-weight model to date.” On the index it scores 39, above Beam at 30 and NVIDIA’s Nemotron 3.5 Super VL at 26, a 120B/12B-active model under the permissive OpenMDW 1.1 whose sibling, Bolt 74B-A9B, uses a native 2-bit format to fit in 24 GB. The strongest open release of the week is the quietest: DeepSeek V4.1 Flash 1017 under MIT at 45, one point behind September’s open leader, MiMo-V2.6-Pro, which keeps its place.

Smaller and more specific again: Liquid’s LFM2.5-24B-A2B (October 13) is a reasoning model with about 2B active parameters under the LFM Open License, priced at $0.10/$1 on Liquid’s API and $0.05/$0.70 through OpenRouter. FLUX 3 Image Klein and Drama 3 are covered above; Span-02, both Decider v1.2 sizes and Drex v2.0 extend the open decision shelf. Also on Hugging Face, listed for the record: China Telecom’s Xing4.0-29B-A4B under Apache 2.0, Bespoke Labs’ Nimble-v4-max under CC BY-NC 4.0, and the Strands Decider pair (2B and 26B), which Amazon posted on October 9 and the first pass missed.

The fourth week brought one promise kept and one still pending. MiniMax putM3.1 Flash on Hugging Faceearly on October 23, four days after the hosted launch: about 220B parameters, ungated, under MiniMax’s own community license. Mistral’s Large 4 still has no repository, with eight days left on “by the end of the month,” and NVIDIA’s Nemotron 3.5 Bolt E4B, which itsOctober 15 postsaid was “coming Oct. 23,” was not on Hugging Face when this pass closed.

The new download worth trying is Recursive’sFlint(October 21), a coding model “built for fast agentic coding on local hardware”: 40B parameters with about 2.5B active, a 262K context, andApache 2.0 weights. Recursive reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, both vendor-run; it is free on OpenRouter for now. Also with weights this week: Baseten’s Mammoth 1 and Elephant 1 (MIT), Bria’s FIBO 2.0 pair, and Gradium’s Phonon (Apache 2.0).

Local-runnable: an image model for 8 GB cards, and 128 GB laptops

The window’s local story is that the new category runs at home. Ollama’s v0.35.1 (September 29) serves Clef and Clef-flash with image input through the same System One route the hosted ones use; the 27B tag is 18 GB and the 9B fits a 6.5 GB Q4 GGUF, with llama.cpp support merged October 3. Liquid’s d1-3B is 1.7 GB at Q4 and answers a question in 8 ms on an RTX 4090 or 30 ms on an M5 Pro, by Liquid’s own clock. Convai’s Laya, 421M parameters in 846 MB, is on Ollama for Apple silicon only. Qwen-Image-2.1-Turbo arrived with GGUFs from 2.5 to 7.6 GB the next day, and the research license travels with them.

Two runtime notes. Ollama v0.40.2 (October 8) upgrades models downloaded by earlier versions in the background on first run, with on-disk backups, then v0.40.3 (October 11) paused those upgrades to protect embedding and request performance. stable-diffusion.cpp’s master-951 (October 9) adds Iris-3B, and master-956 runs the conditional and unconditional halves of classifier free guidance in one batched pass. The arithmetic for what fits has not changed; the RAM and VRAM guide and the flowchart still do it.

The second week moved the runtimes again. Ollama v0.40.4 (October 12) answers decision questions straight from the command line, v0.40.5 (October 14) adds both Nemotron 3.5 models, and v0.40.6 (October 16) adds a 2-bit GGUF type. New libraries include d1 at 3.7 GB, d1-omni at 850 MB, Perplexity’s Decider at 17 GB, LFM2.5 24B at 15 GB, Nimble v3 at 5.6 GB, and Nemotron 3.5 Super at 90 GB; the 25 GB Bolt build says Apple silicon “is not currently supported.” stable-diffusion.cpp’s master-968 (October 15) runs FLUX 3 Image Klein, with GGUFs from 3.7 GB at Q4_K_M, and later builds add MOSS-Image; llama.cpp merged Muse Spark Open, Nemotron 3.5, Beam and Kolibri-1 between October 15 and 17.

The hardware news is a laptop. NVIDIA’s RTX Spark machines went on sale on October 16 from ASUS, Dell, HP, Lenovo, Microsoft and MSI, “starting at $2,599,” with up to 128 GB of unified memory shared by a Grace CPU and a Blackwell GPU. That puts a 90 GB model on a Windows laptop, and it is worth rerunning the Apple silicon versus NVIDIA comparison with it.

The fourth week’s runtimes. Ollamav0.40.8 (October 21) added Flint on macOS, and v0.40.9 (October 22) brought it to Linux and Windows alongside d3, a decision family from vLLM’s semantic-router team in 2B to 27B sizes with image input, and Gradium’sPhonon, text-to-speech models of 31M and 40M parameters at about 100 MB, for Apple silicon only. TheOrnith 1.5 122Btag landed on October 20 at 77 GB, a third size for a family whose smaller models date from August. stable-diffusion.cpp added FIBO 2.0 on October 21, and llama.cpp merged Flint the same day.

Skipped on purpose

What we cut, and why. GPT-6 Cyber: the pre-DevDay reports did not become a model on OpenAI’s pages; the cyber models ship inside a Codex product instead. GPT-6.1 Astra: a reported cancellation, secondary sources only. Grok 4.8, Kimi K2.8: no release, no card, no price in the window, and no Moonshot source for K2.8 at all. Kling 4.0: in the ledger, not ranked until the API has a price. Rho-1, Griffin: in the ledger as announcements, not ranked until there is a weight file or a rate card. Pro reasoning ids: GPT-6.1 Sol Pro and GPT-6 Astra Pro are gateway slugs for a mode, not models. Tev1 4B: reached OpenRouter on September 30 on weights that predate the window. Community decision fine-tunes with self-reported scores and download counts in the millions after a week: listed above, not scored. ChatGPT plans, Dots, marketplaces: DevDay product news, not model news. Moondream 3.1, Pareto 26.9, Scribe v2 Medical: new OpenRouter listings for models that launched before the window, Moondream 3.1 as early as July; a listing is not a launch. Xing4.0, Nimble-v4-max: weights with no post and no price yet, listed above for the record. ElevenLabs audio effects, Runware’s FFmpeg tool: features, not models. Mistral’s open family: the sentence from July, August and September now has a date attached. That is progress, and we will hold them to it. Added on October 23: Ling 3.1 Flash 1019: on the independent index only, with no inclusionAI post, weights or route. Canva Craft: reported on October 22, but Canva’s own page was unreachable to us and there is no API or price yet. VideoGen Model 1: a press release and a private partner rollout. PixVerse V6 Max, GLM 5.3 Flash, Nemotron 3.5 Embed: new routes on Replicate, Mistral’s API and OpenRouter for models that launched before the window. ElevenLabs Realtime v1: agent configurations, not new voice models.

Watching: September’s open items and October’s dates

Added in October and still open: Kling 4.0’s API; the CLM-35B that the CLM-8B card still says is “coming in early October”; Ideogram 4.5’s weights; Clairic’s weights, promised “soon”; NVIDIA’s 4 GB Nemotron 3.5 Bolt E4B, due October 23 and not out when this pass closed; Hume’s TTS and voice-agent APIs, which end on November 13; Anthropic’s retirement of Claude Sonnet 4.5 on November 30; Nano Banana 2’s undated deprecation; Google’s TTS and Live previews ending November 17; Drama 3’s launch price ending November 12; Gemini 4 Argon’s price doubling on December 1; and Fish Audio’s S1 retiring on December 31. Beam’s weights left the list on October 16, and MiniMax M3.1 Flash’s weights and the Decisions API’s general availability left it this week. Between passes, the sortable model comparison table stays current.

Update log

October 12: edition opened, covering September 28 to October 12.

October 19: added twenty-four entries from October 12 to 19, among them Gemini 4 Argon, Muse Wonder and Muse Spark Open, Nemotron 3.5, Beam’s weights, FLUX 3 Image Klein, Krea 3, Drama 3, Kling 4.0 on the web, DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash and eight decision models. Added FLUX 3 Image (October 1), which the first pass missed, and corrected the video section for HeyGen’s 1080p and 2K tiers and P-Video-2 Pro’s reference mode. Decision prices and the September carry-forward table re-read on October 19.

October 23: added fifteen entries from October 19 to 23: Mercury 3, Clairic, Voxtral 3, FIBO 2.0, Flint, Mammoth 1 and Elephant 1, Phonon, MiniMax M3.1 Flash’s weights, the Decisions API’s general availability, Kling 4.0 in Luma, Polly’s new Generative model, ElevenLabs’ Video to Music API, and the Cartesia, Veo 3.1 preview and OpenAI legacy retirements. Carry-forward table re-read on October 23. Replaced the three illustrations with real samples from six handpicked comparisons, one prompt each, drawn from the models’ own pages.