Agentic AI news today, part 1: Muse, Le Chonk and the 1T-parameter wave

Agentic AI news today splits into two piles this week: models shipping, and agents already at work.

Part one covers the model wave. Releases, weights, and prices announced between October 14 and October 22, 2026.

Meta Muse Wonder, Mistral Large 4, MiniMax M3.1 Flash, DeepSeek DeepInfer, and Recursive Flint all landed inside nine days.

Every number here comes from company announcements, benchmark pages, or dated reporting this month.

Part two of this report picks up the agents that lie, browse, and get hired.

AI model news: Meta Muse Wonder and Mistral Le Chonk

Meta launched the Muse Wonder series, saying the models deliver state of the art results on personal and professional agentic tasks, memory, and coding.

The models were trained at Meta’s gigawatt scale Prometheus data center, a scale claim no consumer lab matched this month.

Muse Wonder runs through the Meta Model API and Muse Code. Muse Wonder Full is limited to Muse Code subscribers.

Meta also paused Muse chat sharing on October 21 after roughly 6,800 links surfaced, a moderation stop inside the launch week.

Mistral previewed Mistral Large 4, nicknamed Le Chonk, a one trillion parameter model with 49 billion active parameters.

It is natively multimodal and was trained on 3,800 NVIDIA Grace Blackwell GPUs inside Mistral’s European datacenters.

The preview API is live on Mistral Studio. Open weights are scheduled to drop at the end of October.

Artificial Analysis scores Le Chonk at 38, up from 9 for Mistral Large 3, the jump llm updates news tracked all week.

Mistral and NVIDIA separately plan AI workflows for industrial engineering, with public datasets and benchmarks for agent driven CAE and chip design work.

Multimodal ai assistant picks: MiniMax, DeepSeek, Flint

MiniMax launched M3.1 Flash, a 428 billion parameter mixture of experts model with about 23 billion active parameters and a one million token context.

Input pricing sits at 15 cents per million tokens, the cheapest agentic tier announced this month.

It tops MiniMax’s Automation-Bench v1.0.6 at 62.2 against 50.3 for Opus 5, but trails on Terminal Bench 4.0, 38.1 to 51.8.

MiniMax claims 70 percent lower agent costs for teams, and says weights arrive October 23.

DeepSeek open sourced DeepInfer, the inference engine behind its API, under an Apache 2.0 license.

DeepInfer runs DeepSeek-V4.1-Flash on Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. One cluster serves 155 trillion tokens per day.

DeepSeek withheld the KV cache, prefill and decode disaggregation, and load balancing code, so the release covers offline inference only.

Recursive released Flint, a 40 billion parameter model with about 2.5 billion active per token, built largely by its own automated research system.

Flint runs 125 tokens per second on a MacBook Pro and 150 on NVIDIA’s RTX Spark, fitting in about 23GB at 4-bit.

Flint scores 74.2 percent on SWE-bench Verified but 54.9 on Terminal-Bench 2.1, behind dense Qwen 3.8 at 73.0.

Sage-A1.5 shipped a 35B and 3B mixture agent that beats Agents-A1 by more than six points on BrowseComp and SWE-Bench Pro.

Voice ai news inside the agent stack

HeyGen Voice Turbo took first place in Artificial Analysis’ controlled-voice ranking with an Elo score of 1,202.

Its API pricing is 4 dollars per million characters, against 30 dollars for the previous HeyGen voice model.

Mistral priced a full voice agent at 4 cents a minute, another line where the unit cost collapsed in one release cycle.

Voice ai news now reads like pricing news: the models stopped being the expensive part of a voice assistant.

GM wants the next OnStar assistant to run errands for drivers, debuting in the 2027 Cadillac Lyriq.

Long-horizon agents: Gemini 4 Argon and GPT fleets

Google released Gemini 4 Argon globally on October 14 for developers and enterprises through the Gemini API, AI Studio, and the Gemini platform.

Google positioned it for complex, long horizon work: code migration, enterprise research, and cybersecurity, at 2 dollars per million input and 10 dollars per million output tokens.

A task mode in the Gemini app keeps working when the user steps away, connected to more than 40 applications including Instacart, OpenTable, Adobe, and Shopify.

Gemini 4 Argon agents also beat Minecraft in 27 hours of continuous play, a result published October 19.

OpenAI models are coming to Google Cloud Model Garden, with frontier models and Codex purchasable through the Marketplace and existing spending commitments applicable.

OpenAI also deployed thousands of GPT-Rosalind agents against mouse brain data, finishing in 32 minutes what researchers estimated would take months to a year.

At the largest workload, distributed agents cited more than 700 references while a single agent cited 11, and blinded experts preferred their reports 75 percent of the time.

That experiment is the clearest ai model news of the week for anyone running agent fleets on real datasets.

What the model wave changes

Prices fell across the board: 15 cents per million input tokens, 4 cents a minute for voice, 4 dollars per million characters for synthesis.

Weights opened up: Flint is live on Hugging Face, Le Chonk ships at month end, MiniMax weights follow October 23.

Context hit one million tokens on the cheap tier, which is what a multimodal ai assistant needs to remember a workday.

Distribution moved too: Google Cloud sells OpenAI models, Meta runs its own API, Mistral Studio carries the preview.

Benchmark names shifted with it. Automation-Bench, BrowseComp, and Terminal Bench replaced chat leaderboards in llm updates news.

The practical read for buyers: agentic inference is now a commodity line, and the premium moved to orchestration and control.

Voice ai news and agentic AI news today meet in Mistral’s per minute agent pricing.

Multimodal ai assistant and AI model news buyers ask the same benchmark questions.

AI model news and llm updates news both flagged MiniMax’s October 23 weights.

Agentic AI news today and voice ai news share the OnStar 2027 timeline.

Voice ai news and multimodal ai assistant products both leaned on task modes.

Llm updates news read Flint’s release as local ai model news first.

AI model news and agentic AI news today linked Google Cloud’s OpenAI listing.

Multimodal ai assistant and AI model news overlap on Muse Wonder’s memory claims.

Voice ai news and llm updates news now quote cents per minute together.

Agentic AI news today and ai model news counted the same five launches twice.

Llm updates news and multimodal ai assistant reporting both track one million token contexts.

AI model news and voice ai news covered the same HeyGen price collapse.

Voice ai news and multimodal ai assistant launches shared Mistral’s October pricing event.

Agentic AI news today and llm updates news both ranked Le Chonk first this week.

Agentic AI news today part one closes on prices, weights, and context windows.

AI model news for buyers starts with licensing: Apache 2.0, preview, or API only.

The multimodal ai assistant category absorbed voice, memory, and long context this month.

Llm updates news now reports agent cost per task alongside raw token prices.

Voice ai news ends where product teams feel it: cents per minute of talk.

Flint and Le Chonk anchor the ai model news segment of every October roundup.

Agentic AI news today also noted OpenAI’s models joining Google Cloud Model Garden.

Multimodal ai assistant picks this month favor models with published agent benches.

AI model news and llm updates news agreed on Le Chonk as headline.

Voice ai news and multimodal ai assistant coverage share the OnStar launch beat.

Llm updates news and voice ai news both covered Mistral’s pricing sheet.

Agentic AI news today analysis tied the model wave to part two’s agent incidents.

Every ai model news roundup this week ended with a per token price.

AI model news for researchers cited OpenAI’s 32 minute biological reasoning run.

AI model news readers got DeepSeek’s offline inference release as a separate item.

AI model news columns split preview APIs from downloadable checkpoints in one table.

AI model news this week paired every launch with an open weights date.

AI model news watchers noted Meta’s gigawatt Prometheus data center behind Muse.

AI model news for enterprise starts with Le Chonk’s European training footprint.

AI model news counted five major releases between October 14 and October 22.

OpenAI models arriving on Google Cloud changed how multimodal ai assistant tools get bought.

Google’s task mode reframes the multimodal ai assistant as something that works unattended.

Multimodal ai assistant buyers now compare Automation-Bench before signing contracts.

The multimodal ai assistant race added MiniMax, Sage, and Flint inside one week.

Muse Wonder targets the multimodal ai assistant seat with memory and coding claims.

Multimodal ai assistant launches this month all carried one million token contexts.

The voice ai news beat now overlaps the multimodal ai assistant beat at product launches.

Voice ai news readers saw unit prices fall faster than model quality rose.

Voice ai news columns tied GM’s OnStar errand plans to the cheaper voice stack.

Voice ai news framed Mistral’s 4 cent per minute agent as the real milestone.

Voice ai news led with HeyGen’s Elo 1,202 and its price cut to 4 dollars.

Llm updates news also counted Sage-A1.5 beating its predecessor by six points.

Flint’s 74.2 percent SWE-bench result entered llm updates news as a local model bar.

Llm updates news this week quoted Automation-Bench before MMLU for the first time.

Llm updates news noted Terminal Bench, not chat scores, now separates agent models.

Llm updates news covered DeepInfer serving 155 trillion tokens a day on mixed hardware.

Llm updates news tracked MiniMax dropping input costs to 15 cents per million tokens.

Llm updates news measured Le Chonk’s jump from 9 to 38 on Artificial Analysis.

Mistral’s preview dominated the agentic AI news today column on October 6 follow ups.

Agentic AI news today roundups now list open weights and API prices in the same table.

Part one of the agentic AI news today series stays with benchmarks, not with safety.

The agentic AI news today desk sorted five launches by weight count, then by price.

Agentic AI news today readers got Muse Wonder, Le Chonk, and M3.1 in one news cycle.

Agentic AI news today trackers marked Muse Wonder as Meta’s first frontier agent play.

Llm updates news published a weight release calendar for October within days of Le Chonk.

Voice ai news added Mistral’s per minute pricing to its comparison chart.

Multimodal ai assistant buyers now ask for Automation-Bench scores in requests for proposal.

AI model news kept DeepSeek’s withheld inference components on its checklist.

Agentic AI news today part one treats every price as a per token number.

Llm updates news saw Terminal Bench separate agent models from chat models.

Voice ai news watched HeyGen cut its own price by six times.

Multimodal ai assistant launches this month all ship with memory features.

AI model news counted Meta’s Prometheus center as the largest training claim yet.

Agentic AI news today links each launch to an enterprise buying channel.

Llm updates news compared Sage-A1.5’s BrowseComp gain against the prior Agents-A1 run.

Voice ai news and AI model news both file under cents per unit now.

Multimodal ai assistant coverage noted task modes on Gemini and proactive agents on Claude.

AI model news buyers read Le Chonk’s European GPU footprint as supply chain news.

Agentic AI news today summary: five launches, four open weight dates, one price war.

Llm updates news added Flint’s laptop throughput to its local inference tracker.

Multimodal ai assistant shoppers can now benchmark context length before they benchmark chat.

Voice ai news expects OnStar’s 2027 rollout to use the cheaper agent stack.

AI model news for OpenAI this week meant Model Garden distribution, not a new chatbot.

Agentic AI news today data all traces to October 2026 announcements.

Llm updates news reported Artificial Analysis publishing the Le Chonk score within a day.

Multimodal ai assistant and voice ai news increasingly share the same product launches.

AI model news and agentic AI news today both file MiniMax’s weights date.

Voice ai news readers get the clearest signal from per minute agent quotes.

Llm updates news counted Le Chonk’s score jump as the week’s biggest move.

Llm updates news priced every October launch in dollars per million tokens.

Llm updates news filed MiniMax, DeepSeek, and Sage benchmark notes together.

Llm updates news this month reports weights dates before feature lists.

Llm updates news now leads with Terminal Bench and Automation-Bench tables.

Llm updates news treated Flint’s laptop run as a local inference update.

Llm updates news and AI model news share the same release calendar now.

Llm updates news for October also tracks Google Cloud’s OpenAI Model Garden listing.

Agentic AI news today and voice ai news share Mistral’s per minute agent story.

Agentic AI news today and the multimodal ai assistant beat both watch task modes.

Voice ai news and the multimodal ai assistant column covered HeyGen together.

FAQ: this week’s agentic AI news today

What is the biggest ai model news in October 2026?

Mistral Large 4, a one trillion parameter multimodal model trained on 3,800 Grace Blackwell GPUs, previewed with weights due at month end. Meta Muse Wonder and MiniMax M3.1 Flash followed within days.

Which multimodal ai assistant models are open weight this month?

Recursive Flint is on Hugging Face now under Apache 2.0. MiniMax releases M3.1 Flash weights October 23, Mistral ships Le Chonk weights at month end, and DeepSeek’s DeepInfer engine is already open.

How cheap is agentic inference right now?

MiniMax asks 15 cents per million input tokens with a one million token context. Mistral prices a voice agent at 4 cents a minute. HeyGen’s voice API dropped from 30 dollars to 4 dollars per million characters.

Is voice ai news actually getting cheaper?

Yes. HeyGen Voice Turbo leads the controlled voice ranking at 1,202 Elo while charging 4 dollars per million characters, a sixth of its predecessor’s price.

What does Gemini 4 Argon change for agents?

It ships globally at 2 and 10 dollars per million tokens for long horizon tasks, adds a task mode that keeps working while the user is away, and connects to more than 40 applications.

Where can I read part two of this agentic AI news today report?

Part two, covering agents at work plus the safety incidents, runs at agentic AI news October 2026 part 2.

Related reading on Guide Globe: our AI News October 2026 hub, Latest agentic AI news today roundup, Latest LLM news today, AI infrastructure news and LLM updates, Agentic AI News category. Questions go through the contact page.