AI model releases in October 2026: what shipped, what it costs, and what we could not confirm

The short answer: October's real news for a small business is about price, not a new champion. Claude Haiku 5.5 matched the lowest list price on the market at $0.10 / $0.50 per 1M tokens, Gemini 4 Argon went on sale with a promotional price that doubles on 1 December, and DeepSeek and MiniMax shipped $0.15 / $0.60 models. Nobody has dethroned Claude Opus 5.5 on the independent index we track (1 October capture). We also list six things we could not confirm on a primary source.

Every entry links the company's own page. Dates are the ones the company states; where we could only find a press-reported date we say so. Written 20 October 2026. Earlier months: September 2026 roundup.

What shipped

Claude Haiku 5.5 (7 October). Anthropic's cheapest and fastest model: $0.10 input and $0.50 output per 1M tokens for prompts up to 100K tokens, and $0.50 / $2.50 above that. Independent testing found it writes about three times as much per task as GPT-6 Luna, so the per-token price flatters it (Anthropic pricing, Artificial Analysis). Compare it with Luna in Haiku vs Luna.

Mistral Large 4 (6 October). An open-weight mixture of experts with 1 trillion parameters, 52 billion active. At launch it is a public API preview at $1.36 / $4.18 per 1M tokens, and Mistral says the weights follow at the end of October (Mistral).

Nano Banana 2.1 (6 October). Google's image model went generally available with better text rendering and panoramic aspect ratios, at $0.0336 per 1K image, $0.0504 per 2K and $0.113 per 4K (Gemini API pricing, changelog).

GPT-6 Luna Decisions (6 October, beta). Not a new model: an API that has GPT-6 Luna return typed answers (a yes/no, a pick from options, a score) instead of text. OpenAI says it is 10x faster than its Responses API and lists no price yet (OpenAI changelog).

Gemini 4 Argon (generally available 13 October). $2 / $10 per 1M tokens through 30 November, then $4 / $20 from 1 December. See what it costs after the promo.

Grok Voice 3.0 (October). xAI's new speech-to-speech model is now the default behind its voice alias and is priced the same as Voice 2.0; xAI's model page lists speech to speech at $0.08 per minute (xAI release notes, models). The release notes give only the month.

Muse Spark Open (15 October). Meta's open-weight model: 816 billion parameters, 42 billion active, 1M-token context, under the Muse Model License with limited commercial terms. It runs through hosts such as Together AI and Fireworks, not Meta's own API (Meta).

GPT-6 Pro parallel agents (16 October). A ChatGPT update, not a model: GPT-6 Pro can split a task across several agents at once, for eligible Pro users first (ChatGPT release notes).

DeepSeek V4.1 Flash 1017 (17 October). A refresh behind the existing deepseek-flash id, with open weights; DeepSeek says it is faster and writes shorter outputs, and announced no new price (DeepSeek).

MiniMax M3.1 Flash (19 October). $0.15 input, $0.60 output and $0.015 cached input per 1M tokens. MiniMax says the weights will be released later, and the benchmark figures in its post are its own (MiniMax).

Grok Imagine Video 1.5 Lite. xAI lists it at $0.020 per second of video against $0.080 for Grok Imagine Video 1.5 (xAI models). We found no dated xAI announcement, so we do not state a launch day.

Meta Muse Wonder 1.0. Meta's developer docs list it as a preview model with a 1,048,576-token context window (Meta). The docs carry no launch date or price; press coverage dates it to 12 October, which we could not confirm.

What we could not confirm

• "GLM-5.3 Fast". Z.ai's pricing page lists GLM-5.3, GLM-5.3-Flash and GLM-5.3-FlashX. No Z.ai page lists a "Fast" model, only third-party trackers do, and they disagree. We left it out.
• A price for Muse Wonder 1.0. Press reports quote figures; Meta's own docs do not.
• A price for GPT-6 Luna Decisions. Aggregators list one; OpenAI's changelog does not.
• A dated xAI launch for Grok Imagine Video 1.5 Lite. Social posts imply early October.
• Qwen-Image-2.1-Turbo's launch day. Alibaba's model card on Hugging Face describes a 7B image model that needs 8 steps, under the Qwen Research License; the 9 October date comes from press coverage (Hugging Face).
• Independent scores for most of the new models. Almost every benchmark number above is the vendor's own.

What it means

• Cheap got cheaper and more crowded. Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash and MiniMax M3.1 Flash all list between $0.10 and $0.15 for input. Measure cost per finished job, not per token.
• Promotions have end dates. Gemini 4 Argon doubles on 1 December, and Gemini 3.8 Flash is introductory through 31 December (Google).
• The top spot did not change. Claude Opus 5.5 still leads the Artificial Analysis index we track (58 on 1 October).

Every entry is on the AI model tracker, and head-to-head pages are on model comparisons.

Sources: Anthropic pricing, Mistral, Gemini API changelog, Gemini API pricing, OpenAI API changelog, xAI release notes, xAI models, Meta: Muse Spark Open, Meta developer docs, ChatGPT release notes, DeepSeek, MiniMax, Z.ai pricing, Hugging Face: Qwen-Image-2.1-Turbo