AI Model Releases: October 2026 Tracker and Dated Ledger

Checked October 28, this ledger now includes GPT-6.1 Terra, Genie 4, LFM3, Hy-SimulMT and Unlimited-OCR 2, alongside the earlier Kimi, Voxtral, vector, retrieval and decision releases. It separates vendor announcement dates from later weight uploads, API availability changes and router listing times.

October 2026 opened without a frontier launch and with a cluster of small, specialised models instead. On October 1 Cloudflare and Amazon each released open-weight decision models, Microsoft AI released two text-to-speech models and a streaming transcription model, and Tavus previewed a full-duplex video model to invited testers. On October 2 Bilibili’s Index team published the preview of a 35B-total, 3B-active translation model. This page records each one on the vendor’s own date and is refreshed in place through the month.

For current API prices, access limits, and workload selection, see our maintained frontier model guide.

Written October 3, 2026 and dated October 1, 2026. Rows and listings dated October 2 postdate the dateline and were collected on October 3 with the rest; rows added after October 3 carry the date they were added. Facts are as of October 3 unless a row says otherwise.

Each row records a model or jointly announced family, dated from the vendor’s announcement or an explicit release date on its model card. The primary column names the page each row was read from; the linked sources are collected in the method section.

The two decision-model rows are the month’s first story, and they have their own page: open decision models compared puts Clef, Clef-flash and Decider 2B beside TypeSafe’s Jev on licence, size, context and the vendor’s benchmark rows.

Four of the five rows added on October 20 are covered in our own posts. Our Reflection Beam launch post covers the licence and hardware, and Beam versus Muse Spark Open compares the two US open-weight releases. DeepSeek’s update is in our DeepSeek V4.1 Flash 1017 post, MiniMax’s model in MiniMax M3.1 Flash: price and benchmarks, and a three-way comparison puts it beside DeepSeek’s update and Mercury 3. Glimpse has no page of its own; its scores, like Beam’s, are the vendor’s.

Microsoft’s post attaches two claims to the Flash voice model that are the vendor’s own and unverified here: that inference is 55% faster and roughly 60% cheaper than comparable models, and that it produces 45 seconds of audio at 150 milliseconds end to end. Our September ranking of text-to-speech models is where the two voice models will be scored once listener data exists.

Availability update, October 26: MiniMax M3.1 Flash weights arrived on October 23 under the custom MiniMax Community licence: the base grant is non-commercial, and commercial use carries attribution plus notice or, where the product or service exceeds US$20 million yearly revenue, prior written authorisation. FIBO 2 Dev and Edit Turbo weight repositories date October 18 and 19; their cards state CC BY-NC 4.0, with commercial licensing separate. Nanbeige 4.3 33B and Tencent Hy-Nomas also have public weights, but their repository timestamps alone are not vendor announcement dates under this ledger's rule.

Seedance 2.5's new 4K tier is a selected-enterprise trial; it is not the same access offer as the existing base-model API. Our video API access comparison separates that trial from Kling 4's unavailable API and currently callable alternatives. OpenAI Decisions reaching general availability is an API change, so it receives its own adoption guide without being counted as a new model here.

Gemini 4 Argon was announced in September, so its October general availability sits outside the release ledger. Google announced Gemini 4 Argon on September 30 as rolling out to a set of trusted cyber defenders through its Fairwind programme, with an introductory price of $2 per million input tokens and $10 per million output tokens and cached input at 95% off. On October 13 Google made it generally available in the Gemini API as gemini-4-argon, with the introductory price running through November 30. Our Argon page records the announcement, and the migration checklist covers the general release.

Three September releases sit just outside this ledger and are not counted here: Together’s Tev1 4B experimental, published on GitHub and Hugging Face on September 23 and listed on OpenRouter on September 30; Inception’s Mercury Decide, listed on OpenRouter on September 30 with no vendor page located; and Bilibili’s dense Index-Translate 2B and 9B, on Hugging Face since September 28, which the October 2 preview extends. The two decision models are covered in decision models that return probabilities, not text. The month before this one is in the September tracker.

The changes below, dated October 15 to 17, sit here rather than in the ledger, because each is a feature, a product or an update to an existing model. On October 15 Anthropic’s Claude Platform release notes took fast mode for Claude Opus 5.5 out of research preview: no beta header is needed, and Opus 5.5 has a default fast-mode rate limit in the Start, Build and Scale tiers, while fast mode for Opus 5 and Opus 4.8 stays in research preview. Our guide to model speed tiers covers what the premium buys.

On October 16 OpenAI’s API changelog released prompt-to-voice to all API customers, which builds a custom GPT-Live voice from a text description and charges nothing to create it, and added nine GPT-Live voices in German, French, Hindi, Japanese and Korean. Our prompt-to-voice guide covers the limits. The same day, OpenAI’s ChatGPT release notes announced ChatGPT Slides, rolling out to eligible Pro, Business and Enterprise users with Space access (see which plans get ChatGPT Slides), and an update to GPT-6 Pro in ChatGPT with faster responses, better tool use and several agents working on parts of a task in parallel. DeepSeek’s Search and Fetch APIs, opened on October 17 with the 1017 update, are APIs rather than models: its pricing page lists each at $0.75 per 1,000 successful requests, with server-side calls through DeepSeek Harness, Claude Code or Codex free until 00:00 Beijing time on November 1, 2026.

The pattern is the inverse of September’s opening, which brought three frontier models in two days. This month began with models that do one thing: decide, speak, transcribe, translate, or hold a face-to-face conversation. Two of the five vendors are infrastructure companies, Cloudflare and Amazon, releasing weights rather than selling tokens. That is consistent with where decision models sit in an agent: close to the tool call, where a hosted round trip is the cost.

The opening-days releases largely added specialist options: a cheap check before an action, a voice across languages, or a translator with open weights. Later releases include general-model successors such as Kimi K3.1 Pro, so replacement decisions now belong in the comparison too. The question to ask of each is not “is it better than the frontier” but “what does it let me stop sending to the frontier”. Our AI transformation work starts from that question.

A dated ledger is built from rules, and each rule exists because a reader would otherwise draw a conclusion the rows do not support. The illustrative readings below are the four the method rows guard against. They are patterns of interpretation, not things any particular reader has done.

The listing as a release: an OpenRouter entry that appeared on October 1 is cited as a model released on October 1. The dating rule says a listing never sets a date on its own, which is why the original October 3 observations for Pareto 26.10 Preview, Apodex 1.1 Mini and Ling 3.1 Flash remain in the corroboration section and outside that opening sample. The count as a market signal: eight releases in three days becomes a statement about the pace of the industry. The count is of vendor-dated pages in a three-day window, counted by model so that Clef and Clef-flash are two, and the same rule excludes availability changes, so September-announced Argon’s October availability change is not in that opening sample. The vendor claim as a measurement: Microsoft’s figures for its Flash voice model, 55 percent faster, roughly 60 percent cheaper and 45 seconds of audio at 150 milliseconds, are quoted as benchmarks. They are the vendor’s own and unverified here, and the ranking page where they would be scored waits for listener data. Argon as released: Gemini 4 Argon, with a published introductory price, is added to the ledger. It was announced on September 30 and reached general availability on October 13, which is an availability change the exclusions row keeps out.

The ledger is useful because the rules hold. A reader who quotes a row with its date, its primary source and the rule that put it there has quoted the page; a reader who quotes the count alone has quoted a number the page defines more narrowly than the sentence will.

A source-based ledger of selected vendor announcements. No model on this page was run by Digital Applied.

What was collected. The original October 1–3 sample records eight models found on vendor-dated release pages. Later additions extend the ledger with selected, source-verified releases through October 28. Each records the vendor-stated price, context, licence and access. Jointly announced sizes or families may share a row, so the row total is not a count of individual models.

Sources. Cloudflare’s Clef announcement and Workers AI model pages; Amazon’s Strands Decider post on the Strands Agents site, dated October 1; Microsoft AI’s transcription and voice post on its news site, dated October 1; the Index-Translate-35B-A3B-preview card; Hugging Face API records for licence tags, base-model tags and creation dates; Tavus’s announcement page and its Business Wire release; the OpenRouter models API for listing times.

October 26 additions: Kimi's launch post; Mistral's changelog; Ideogram's model page; Recursive's announcement; NVIDIA's retrieval announcement; Bolt's model card; Base Labs' official model cards; and Kling's API availability page. Nace AI's Introducing Drex Max (October 19), Kuaishou's company release (October 19, describing October 15), and VideoGen's company announcement (October 20) supply the other announcement dates.

October 28 additions: OpenAI launch; DeepMind announcement; Liquid AI launch; Tencent card; Baidu card. These primary sources were checked on October 28.

As-of date. October 28, 2026 for the five latest rows; October 26 for the preceding ten additions. Older rows retain their stated observation dates. The ledger was opened on October 3, 2026 and covers October 1–3; later rows carry the date they were added. On October 20, 2026 we read Reflection’s launch post and model card with the Hugging Face blog, Elorian’s launch post, DeepSeek’s API change log and pricing page, MiniMax’s launch post and API reference, Inception’s launch post and models page, OpenAI’s API changelog and ChatGPT release notes, and Anthropic’s Claude Platform release notes.

Dating rule. A row’s date is the vendor-stated announcement or release date, including an explicit release date on an official model card. Where a later company announcement reports an earlier launch, both dates are shown in the primary-source note. A Hugging Face card created before the announcement does not move the date earlier; an OpenRouter listing never sets a date on its own.

Exclusions. Availability changes to models announced earlier, such as Gemini 4 Argon; products, programmes and APIs that are not models; OpenRouter listings with no located vendor page.

Refresh. Rows are added as vendors publish through October 31, each marked with the date it was added. Last refreshed October 28, 2026, with five additional rows. A row may group model sizes or a jointly announced family; the row count is not a count of distinct models. Corrections are made in place with a dated note. The month closes with a count of releases by type.