AI Model Releases in October 2026: Every New Model, Date and API Price

October 2026's biggest release is Google's Gemini 4 Argon (October 13), priced at $2 input and $10 output per million tokens until November 30. Anthropic shipped Claude Haiku 5.5 on October 7, Mistral previewed Mistral Large 4 on October 6, Meta launched Muse Wonder on October 12, and OpenAI shipped no new model this month, only API features.

This tracker lists each release with its date on the vendor's own announcement or changelog, its API price, and who can use it today. Every entry was checked on October 17, 2026. The September flagships sit lower on the page, because most buyers are still choosing between those and the October models.

Model	Vendor	Released	API price per 1M tokens (input / output)	Who can use it
Gemini 4 Argon	Google	Oct 13	$2 / $10 until Nov 30, then $4 / $20	Paid API tier; Google AI Pro and Ultra in the Gemini app
Muse Wonder	Meta	Oct 12	Not listed in the launch post	Meta Model API and Muse Code
Muse Wonder Full	Meta	Oct 12	Not offered on the API	Muse Code subscribers
Grok Voice 3.0	xAI	Oct 14	Same rate as Grok Voice 2.0	xAI API, speech to speech
Claude Haiku 5.5	Anthropic	Oct 7	$0.10 / $0.50 up to 100K prompt tokens	Claude Platform, AWS, Google Cloud, Azure
Mistral Large 4	Mistral	Oct 6	$1.36 / $4.18	Public preview API; weights promised later
Nano Banana 2.1	Google	Oct 6	$1.50 input; about $0.0336 per 1K image	Paid Gemini API tier

What counts as a model release here

A model release, for this tracker, is a new model identifier a buyer can call or select, announced on the vendor's own news page or developer changelog. Feature launches, price changes and deprecations get their own section because they change bills too, but they are not counted as models. A model reported by the press but absent from the vendor's own pages stays off the list until the vendor publishes it.

Toolradar data: the AI assistants ranking on Toolradar, the software directory we run, evaluated 1,559 tools in October 2026. Almost all of them sit on top of a handful of vendor models, so a release on this page usually reaches the apps you already pay for within weeks.

What a typical workload costs on each new model

We priced one fixed workload on every text model above: 1,000 requests, each with 10,000 input tokens and 1,000 output tokens, at standard list prices with no caching or batch discount. That shape fits a support bot reading a short document or an agent step with a modest context.

Model	Cost of the 1,000-request job	What drives it
GPT-6 Luna (Sep 22)	$1.50	Lowest list price in this group
Claude Haiku 5.5 (Oct 7)	$1.50	Same price as Luna while prompts stay under 100K
Mistral Large 4 (Oct 6)	$17.78	Open-weight model at a mid-tier price
Grok 4.7 (Sep 21)	$26	Output costs less than on Sonnet 5.5 or GPT-6.1 Sol
Gemini 4 Argon (Oct 13)	$30, then $60 from Dec 1	Introductory price ends November 30
Claude Sonnet 5.5 (Sep 28)	$30	Same list price as GPT-6.1 Sol
GPT-6.1 Sol (Sep 29)	$30	Same list price as Sonnet 5.5
Claude Opus 5.5 (Sep 22)	$60	Anthropic's top Opus tier
GPT-6 Astra (Sep 3)	$150	OpenAI's most capable model

The spread is 100x from the cheapest to the most expensive model, for the same job. That is the main reason to read a tracker like this one: a team that routes routine requests to a small model and saves the flagship for hard tasks pays a fraction of the single-model bill. The arithmetic is ours, on the vendors' published per-token rates.

Gemini 4 Argon is the release that changes budgets

Google released Gemini 4 Argon to general availability on October 13, describing it on its Gemini API changelog as its most capable model for complex coding, agentic, math and science tasks. It is not on the free API tier.

The price is the part to plan around. The Gemini API pricing page keeps the launch rate through November 30, 2026, then doubles both input and output prices from December 1. A project budgeted on the launch price doubles in cost on December 1 unless it moves work to a cheaper model first. Batch pricing is half the standard rate on both dates.

In the consumer app, Google's announcement says Google AI Pro and Ultra subscribers can use Argon. Free Gemini users do not get it, so a free account is not a way to try it. Our Gemini pricing breakdown covers what those plans cost, and the Gemini review covers where Gemini fits against the other assistants.

Claude Haiku 5.5 is the cheap model with a prompt-size trap

Anthropic released Claude Haiku 5.5 on October 7 as claude-haiku-5-5, its fastest and cheapest model. The launch post puts the list price at a tenth of Haiku 4.5's on short prompts, and Anthropic's own estimate is that real workloads cost around 75% less to run. It is the first Haiku with an adjustable effort setting, so a developer can trade cost for quality per request.

The catch is the 100,000-token threshold. Anthropic's pricing docs say any request whose prompt passes it pays the higher rate for the whole request, including cached tokens: $2.50 per million output tokens instead of the short-prompt rate. A long-document job can cost five times more per token than the headline suggests. Anthropic says about 90% of requests to its previous Haiku stayed under that line.

The same day, Anthropic cut Sonnet 5.5 cache reads in half, which it says lowers Sonnet's cost on most agentic work by around 20%. For the subscription side, see our Claude pricing page, and the Claude review for how the apps compare.

Mistral Large 4 is the big open-weight release

Mistral released Mistral Large 4 in public preview on October 6, and the Mistral announcement describes a 1 trillion-parameter multimodal model with 52 billion active parameters. Its model card lists a 1M-token context window and a list price of $1.36 input and $4.18 output per million tokens, with a launch offer at half price for two weeks.

The weights are the reason to watch it. Mistral says open weights are coming, which would let a company run the model on its own hardware without sending data to an API. Until they ship, it is a preview API like any other, so treat production plans as provisional. Our list of open-source LLMs covers the models whose weights you can download today.

Meta's Muse Wonder is a frontier model with limited access

Meta introduced Muse Wonder and Muse Wonder Full on October 12, calling them its most capable models to date in its newsroom post. Muse Wonder is available to developers through the Meta Model API and Muse Code. Muse Wonder Full, the larger version, is available only to Muse Code subscribers.

The post names two features: Active Memory, which carries context across conversations, and Active Focus, which keeps an agent working through a hard task. Meta says Muse Wonder will power its Muse personal agent app, rolling out to subscribers first. The launch post lists no API price, so cost comparisons with the table above are not possible yet.

Smaller October releases: images and voice

Google released Nano Banana 2.1 on October 6, an update to its Nano Banana 2 image model with better text rendering, character consistency across edits, and panoramic aspect ratios up to 8:1. The pricing page converts its image output to about $0.0336 per 1K image and $0.113 per 4K image. The older gemini-3.1-flash-image is deprecated. Our AI image generators list covers the consumer apps built on models like it.

xAI made Grok Voice 3.0 available on October 14, according to its release notes, at the same price as Grok Voice 2.0. The grok-voice-latest alias now points to it, so apps using that alias were upgraded automatically, and Grok Voice 2.0 is deprecated.

OpenAI shipped features, not models, in October

OpenAI's last model launch was GPT-6.1 Sol on September 29. Its API changelog for October lists the Decisions API in beta on October 6, running on GPT-6 Luna and described as returning typed answers 10x faster than the Responses API. It also lists an Ultrafast mode for GPT-6.1 Sol on October 8, plus new voices for GPT-Live on October 16.

The account-level change is the usage tiers, cut from five to three (Build, Launch and Grow) on October 6. Rate limits now follow total credit purchases, so a team that hit a limit should check which tier it landed in. Our ChatGPT pricing page covers the consumer plans.

The September releases most buyers are still comparing

September carried the larger launches, and most of them are still the newest model in their tier:

• GPT-6 Astra (OpenAI, September 3): OpenAI's most capable model, at $10 input and $50 output per million tokens on the OpenAI pricing page.
• GPT-6 Sol and GPT-6 Luna (OpenAI, September 22): Luna lists at half the input price of the GPT-5.6 Luna it replaces and well under half its output price, which is why it ties Haiku 5.5 in the cost table.
• GPT-6.1 Sol (OpenAI, September 29): same price as GPT-6 Sol with cheaper cached input, aimed at coding and professional work.
• Claude Opus 5.5 (Anthropic, September 22): Anthropic says it matches Claude Fable 5.1 on most work and costs 40% less to run than Opus 5, which makes it the default for teams that were paying Fable rates for routine agent work.
• Claude Sonnet 5.5 (Anthropic, September 28): described as 30% faster than Sonnet 5 and up to 30% cheaper for most work, so it suits everyday coding and documents where Opus is more than the task needs.
• Grok 4.7 (xAI, September 21): a 500K context window at $6 output per million tokens below 200K prompt tokens.
• DeepSeek V4.1-Flash (DeepSeek, September 10): the smallest model in DeepSeek's new architecture family, with native image understanding.

For a side-by-side of the two most-used assistants, see Claude vs ChatGPT, and Claude vs Gemini for the Google matchup.

Retirements and dates to put in a calendar

New releases push old models out, and a few dates are close enough to break a working integration:

• October 23: Google shuts down the deep-research-pro-preview-12-2025 agent. Requests must move to a newer Deep Research version.
• October 31: Z.ai GLM 5.2 retires on Mistral's platform, replaced by GLM 5.3 at the same price, per the Mistral changelog.
• November 2: xAI retires grok-imagine-image-quality, with requests served by grok-imagine-image-2.0.
• December 1: Gemini 4 Argon's API price doubles.

Google also routes requests for gemini-3.7-flash and gemini-3.5-flash to newer Flash models since October 8. Routing keeps an app running, but the output of a routed model can change without a code change, so rerun your evaluations.

How to use this tracker when choosing a model

Start with the cost table, not the benchmark charts. Most teams do not need the most capable model for every request, and the cheapest two models here cost a hundredth of the most expensive for the same job. Route classification, extraction and short summaries to a small model, then test whether the hard 10% of requests needs a flagship.

Then check access. Two of October's headline models are not available to everyone: Argon needs a paid Google plan or API tier, and Muse Wonder Full needs a Muse Code subscription. For agent work, our explainer on agentic AI covers what those agent features do, and the AI chatbots list compares the apps built on these models.

How we track releases

Each entry comes from the vendor's own announcement, model card, pricing page or developer changelog, read on October 17, 2026. Prices are standard list rates per million tokens for short prompts, without batch, caching or regional surcharges. The cost table is our arithmetic on those rates. Third-party release trackers were used only to find candidates, never as the source of a date or price. The editorial rules for this site are on the AIpresso about page.

FAQ

What AI models were released in October 2026?

As of October 17, 2026, the main October releases are Google's Gemini 4 Argon (October 13) and Nano Banana 2.1 (October 6), Anthropic's Claude Haiku 5.5 (October 7), Mistral Large 4 in preview (October 6), Meta's Muse Wonder and Muse Wonder Full (October 12) and xAI's Grok Voice 3.0 (October 14). OpenAI released no new model in October; its last was GPT-6.1 Sol on September 29.

What is the newest Claude model?

Claude Haiku 5.5, released October 7, 2026, is the newest. It completes the Claude 5.5 family after Opus 5.5 on September 22 and Sonnet 5.5 on September 28. Haiku is the low-cost model, and Opus 5.5 is the most capable of the three, at $4 input and $20 output per million tokens.

What is the newest OpenAI model?

GPT-6.1 Sol, released to the API on September 29, 2026, is OpenAI's newest model. GPT-6 Astra, released September 3, remains its most capable and most expensive model. In October OpenAI shipped API features, including the Decisions API and an Ultrafast mode, rather than a new model.

How much does Gemini 4 Argon cost?

On the Gemini API, Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens through November 30, 2026, rising to $4 and $20 from December 1. It is not on the free API tier. In the Gemini app, it is available to Google AI Pro and Ultra subscribers.

Which new AI model is the cheapest?

GPT-6 Luna and Claude Haiku 5.5 have the lowest list prices of the models on this page, both at $0.10 input and $0.50 output per million tokens on short prompts. Haiku 5.5 charges more once a prompt passes 100,000 tokens, so Luna is cheaper for long documents.

Is Mistral Large 4 open source?

Mistral describes Mistral Large 4 as an open-weight model, but as of October 17, 2026 it is a public preview on Mistral's API and the weights have not shipped yet. Mistral says they are coming. Once released, you could run the model on your own servers instead of paying per token.

How often is this tracker updated?

Each update re-reads the vendor changelogs and pricing pages and moves the dated "checked on" note. The September and October entries were last verified on October 17, 2026. A model appears only once its vendor publishes it.

Cite this: AIpresso, "AI Model Releases in October 2026: Every New Model, Date and API Price", October 2026.

Some offers on this page may be paid placements or contain affiliate links.