AI Agents Get Cheaper Than Humans: The October 2026 Release Flood

October 2026 AI Roundup: World Models, Open-Source Trillion-Parameter Models, and the Agent Battle

October 2026, the AI world was like a toppled set of dominoes: Google served up Genie 4, which can "run a world"; OpenAI used GPT-6.1 Terra to push the cost of computer operation below human labor; Moonshot AI put an open-source model with 3 trillion parameters on the table; Anthropic filed for an IPO while launching its personal agent Claude Duet. This month, large models collectively shifted from "being able to talk" to "being able to work," with three tracks firing at once: open source vs. closed source, general-purpose vs. specialized, and world models vs. agents.

This piece skips idealism and just rounds up the ten things worth knowing right now, each with "what it is, why it matters, and what it means for you."

📦 Supporting resources: This is a news roundup. It has no source code packages, configuration lists, or chart source files, so the supporting resources section is empty and skipped.

Google Genie 4: From "generating video" to "running a world"
OpenAI GPT-6.1 Terra: Computer operation cost drops below human labor for the first time
Kimi K3.1 Pro: A 3-trillion-parameter open-source model tops the open-source leaderboard
Anthropic: IPO filing and the personal agent Claude Duet
DeepSeek V4.1 Flash and MiniMax M3.1: A double shot on the open-source side
Google Gemini 4 Argon opens worldwide
OpenAI's thousand-agent run cracks months of brain-cell analysis in 32 minutes
Eight-ministry "15th Five-Year Plan": agent adoption rate target exceeds 90%
Comparison table of the ten events
Afterword: What this wave means

1. Google Genie 4: From "generating video" to "running a world"

On October 28, Beijing time, Google DeepMind released its next-generation general world model Genie 4, described on the official site as "the most realistic and interactive world model to date." The most disruptive change is pushing generative AI from "generating video" to "running a world": from a single prompt, the model generates an interactive, dynamic virtual world in real time, where scenes keep adjusting to your input, characters respond autonomously, and the environment evolves on the fly.

What's really new? The most obvious points are two: the world becomes more realistic, and for the first time it has sound. Genie 4 is Google's first real-time world model that natively generates audio; ambient sound and character dialogue are produced in sync with the world. It also supports real-time physics simulation, so character actions, interactive dialogue, and scene modifications all respond instantly to a one-sentence instruction. Users aren't just "watching" an AI-generated world; they can "enter" it and interact with the environment.

Google CEO Pichai called this step the beginning of a "oasis" era of "speak and it happens." Some industry observers remarked that two years ago AI couldn't even render hands properly in static images, and now it can generate playable 3D worlds with real-time physics and live audio on the fly.

Why it matters: The world model has long been seen as the missing piece for robot training and agent evolution; it provides not just visuals but a controllable, interactive training environment with physical rules. Genie 4 pushes "seeing" into "hearing and interacting," laying the groundwork for merging embodied AI with generative intelligence. Google is exploring applications in game design, film production, education, and robot training.

A cold splash: It's still a research preview, open only to trusted testers; regular users can't get it yet. And all audio and video are watermarked with SynthID.

Genie 4 references:

2. OpenAI GPT-6.1 Terra: Computer operation cost drops below human labor for the first time

On October 26, OpenAI launched GPT-6.1 Terra, positioning it as "fast, efficient, built for high-intensity professional work," with particular strength in visual reasoning, computer use, and clear, natural writing. It opened the same day in ChatGPT Work, Codex, and the API, covering Plus, Pro, Business, Enterprise, and Edu tiers.

The real highlight is the cost structure. Terra is priced at $1 per million input tokens and $5 per million output tokens, a direct 50% drop from the previous generation. At the highest intensity setting, the cost per single task is just one-ninth of Claude Opus 5.5. On Terminal-Bench 4.0, Terra achieves 50.3% success, versus 21.5% for Claude Opus 5.5; its factual error rate drops from 53.6% to 16.7%.

Why it matters: These numbers mean computer-use automation has entered, for the first time, the zone where "priced per task it's cheaper than human labor." In the past we talked about token costs; now the industry starts talking about "single-task cost" — cost per completed task rather than cost per token — and it becomes the new yardstick for routing decisions. When the cost of completing one task really is lower than a stretch of manual operation, the economics of agents taking over computer operation at scale finally add up for the first time.

GPT-6.1 Terra references:

3. Kimi K3.1 Pro: A 3-trillion-parameter open-source model tops the open-source leaderboard

On October 23, Moonshot AI released Kimi K3.1 Pro, the first model in its K3.1 series. It has 2.8 trillion total parameters (some say 3 trillion) with 104 billion active, making it the largest publicly released open-source weight model. Context length is 1 million tokens, pricing is $2 input and $10 output (a 33% cut from the previous generation), and the API and Kimi.ai were both available the same day.

Its capability isn't just scale: it scores 55 on the Artificial Analysis Intelligence Index — the highest ever for an open-source model — and ranks fourth overall. It reaches 65.7% on Terminal-Bench 4.0. The weights are scheduled to be public on November 4.

Why it matters: This is a direct open-source response to closed-source flagships. Hitting the 55-point tier with 3 trillion parameters while costing a third less than the previous generation shows Chinese AI vendors have genuinely leveled up on the "parameter scale × cost × capability" balance. For Chinese developers and enterprises, K3.1 Pro means getting an open-source base that can be deployed locally and handles frontier tasks. It performs strongly on long-tail professional tasks like game development (Unreal/Unity/Godot) and Blender/CAD 3D modeling — precisely the vertical deep-water where general-purpose LLMs fall short.

Kimi K3.1 Pro references:

4. Anthropic: IPO filing and the personal agent Claude Duet

Anthropic is this month's "double act."

First, the IPO. On October 27, Anthropic filed its IPO paperwork, revealing its full financials for the first time: Q2 2026 revenue of $11.6 billion, up a staggering 1368% year over year, with operating profit of $974 million in the same period. This is the first time a frontier LLM vendor has disclosed complete financials in an IPO filing, and it's seen as the key sample for testing whether the LLM business model can sustain profitability. The offering could value the company at around $1.5 trillion, but most Asian investors "almost certainly" can't get a direct allocation in the placement; funds that miss out are pivoting to buying shares in its already-public shareholders, crypto derivatives, and semiconductor stocks.

Second, the product. On October 21, Anthropic released its personal AI agent Claude Duet, entering the personal agent track, capable of proactively monitoring apps and executing tasks like meeting scheduling. This is already the seventh or eighth similarly shaped product in the last two months — xAI launched Grok Bot in August, Meta released Muse in September and pushed its stock up 27%. The piece argues that the personal agent track has entered a phase of homogeneous competition, with vendors mostly differentiating on interaction form and entry point; meaningful differentiation isn't visible yet.

Why it matters: The IPO lifts the lid on whether LLM companies actually make money, and Claude Duet shows frontier vendors are collectively shifting the battlefield from "model capability" to "personal agent entry point." One is a ledger that manages money, the other a product that grabs entry points; both point to the same thing: AI is moving from technology narrative to commercial payoff.

Anthropic IPO and Claude Duet references:

5. DeepSeek V4.1 Flash and MiniMax M3.1: A double shot on the open-source side

If Kimi raised the open-source ceiling, DeepSeek and MiniMax together floor the "per-task cost" on the open-source side — in these ten days, open-source domestic models pushed "low price for high frequency" to the extreme.

DeepSeek V4.1 Flash: On October 17 it upgraded to V4.1-Flash-1017, the smallest member of its new model architecture family, with native multimodal visual understanding, positioned as "higher capability ceiling, faster reasoning, greater throughput, more scalable." The official statement calls this upgrade a "performance-cost Pareto improvement": compared to the previous version, HLE (w/o tools) jumped to 42.6, Terminal-Bench 4.0 to 42.3, DeepSWE v1.1 to 77.0 — and it cut task runtime and output length while improving performance, so you can run the same work at lower cost and higher speed. It scores 45 on the Artificial Analysis intelligence index and is specifically optimized for platforms like DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.

MiniMax M3.1 Flash: Released on October 19, with open weights on October 23. What made it stand out isn't parameters but an "alias" — the Space Bunny that anonymously topped the OpenCode leaderboard for 12 days is, in fact, an early version of M3.1 Flash. It's a MoE (Mixture-of-Experts) model with 428 billion total parameters, activating only 23 billion per token, a 1-million context, and native text/image/video input. Pricing is $0.15 per million input tokens, $0.6 for output, and $0.015 for cached reads.

Why it matters: These two pushed the "Agent task unit price" to a new level. MiniMax's pricing matches DeepSeek's off-peak price exactly, but DeepSeek doubles to $0.3/$1.2 during peak hours, while M3.1 Flash stays the same all day — so it's not just "cheaper," it's "cheap around the clock." The most direct test from users: building a hardware launch landing page took M3.1 Flash 23k tokens and $0.014, while Claude Opus 5.5 took 60k tokens and $1.22 — yet the design quality was on par; the bill was nearly 88x different.

A cold splash: Cheap doesn't mean omnipotent. On hard engineering tasks like DeepSWE v1.1, M3.1 Flash scores 73.7, DeepSeek 77; on OSWorld 2.0 M3.1 Flash scores 61.7, clearly behind Claude Opus 5's 70.6. Its hallucination rate is high, and it can fail on tasks with physical constraints. So a more precise take: delegate high-frequency, high-volume work to it, and leave hard refactoring to stronger models.

DeepSeek V4.1 Flash and MiniMax M3.1 references:

6. Google Gemini 4 Argon opens worldwide

On October 29, Google pushed Gemini 4 Argon to the world: developers and enterprises can use it via the Gemini API, and Google AI subscribers can use it directly in Antigravity, the Gemini app, and AI Mode in Search.

What's it strong at? The official positioning is "built for your hardest problems," excelling at complex multi-step tasks in coding, finance, law, cybersecurity, and science. Crucially, it ranks first on the Vals Index — a benchmark that measures models on performance in critical economic domains like finance, coding, law, and tax; Argon is the current cross-domain leading flagship. Google even used it to carry out a Chrome architecture refactor spanning hundreds of files, compressing months of work into weeks.

Pricing: Entry price is $2 per million input tokens and $10 for output (cached input gets 95% off). Note this is a promotional price; after November 30, input rises to $4, and output rises to $20 starting December 1.

Why it matters: Google actually played two cards this month: the first is Genie 4, which "runs a world"; the second is Argon, which "does deep work." One builds worlds, the other handles the most tedious professional tasks in the real world. Together they show Google is no longer just competing on "how much the model knows," but on "whether the model can take over high-value, high-difficulty vertical work for you." For developers, Argon is a flagship entry point available globally (not just in the US).

Gemini 4 Argon references:

7. OpenAI's thousand-agent run cracks months of brain-cell analysis in 32 minutes

In October, OpenAI showed the most impactful demonstration of "agent scaling": together with PerturbAI and Scripps Research, it used thousands of AI agents (driven by GPT-Rosalind) working in parallel to identify biologically relevant leads from data on millions of mouse brain cells in 32 minutes.

What exactly did it do? Traditionally, data from large-scale experiments (like gene perturbation screens) far exceeds what scientists can manually analyze. OpenAI had thousands of agents in parallel "read experimental data + search scientific literature," linking the effects of gene perturbations to existing findings in the literature, thereby surfacing unexpected biological connections that researchers hadn't noticed. These connections can help scientists understand disease and find new therapeutic research directions.

Why it matters: This fight is about whether "scientific discovery can be scaled by agents." Looking at just the 32-minute number, it compresses work that might have taken months of iterative derivation by a small group into "thousands of agents running a bit of code in parallel." More critically, it's not "let AI spout nonsense"; it's having agents "catch" leads buried in lab data that human eyes can no longer see. When AI starts telling scientists "what experiment is worth doing next," agents stop being mere tools and become part of the research workflow.

OpenAI thousand-agent brain-cell analysis references:

Using thousands of AI agents to uncover new leads in biological data -- OpenAI

8. Eight-ministry "15th Five-Year Plan": agent adoption rate target exceeds 90%

If the earlier items are dots on the vendor side, this one is the plane on the national side. At the end of October, the Central Cyberspace Affairs Commission issued the "National Informatization 15th Five-Year Plan", drawing a task list, timeline, and roadmap for the next five years of informatization.

One number worth noting: among the 12 main indicators in the plan, the biggest increase is for the "adoption rate of next-generation smart terminals, agents, and similar applications" — from 40% in 2025 to over 90% by 2030, more than doubling in five years. In other words, the national level has put "agents" on the same footing as 5G, IPv6, and the core digital economy industry, with a clear quantitative target.

Other signals from the plan are also directional: internet penetration rises from 80.1% to 85%, 5G (including 5G-A) base stations per 10,000 people rise from 34 to 50, the digital economy core industry's value-added share of GDP rises from 10.5% to 12.5%, and the plan "will start commercial use of sixth-generation mobile communications (6G) at an appropriate time." It also explicitly calls for accelerating the industrialization of quantum technology, brain-computer interfaces, embodied intelligence, 6G, and other technologies.

Why it matters: This is the first time agents appear as a "national adoption rate indicator," meaning they are no longer just a tech-circle buzzword but have entered the hard constraints of industrial policy. For practitioners, it's a clear "certainty signal": investing in agents has explicit policy support for the next five years. For enterprises, "90% agent adoption" isn't a slogan; it will concretely translate into assessment and support for intelligent transformation across industries.

Eight-ministry "15th Five-Year Plan" references:

Agent application adoption rate exceeds 90%, 6G commercial use to start at an appropriate time; what signals does the national informatization '15th Five-Year Plan' release -- The Paper

9. Comparison table of the ten events

Put the month's main threads into a table and you'll see more clearly where they're heading — from "being able to talk," to "being able to work," to "entering the workplace to work."

Event	One-sentence summary	What it means for you
Google Genie 4	From "generating video" to "running a world"; first real-time world model with native audio	Entry point for merging embodied AI and generative intelligence, but still a research preview
OpenAI GPT-6.1 Terra	Computer operation cost drops below human labor for the first time; single-task cost is one-ninth of Opus 5.5	Computer-use agents finally have a viable economic case
Kimi K3.1 Pro	2.8 trillion (some say 3 trillion) open-source model, 55 points, the highest in open-source history	An open-source base you can deploy locally and tackle frontier tasks
Anthropic IPO	Revenue $11.6B, up 1368% YoY; valuation around $1.5T	Key sample for testing whether LLM business model can sustain profitability
Claude Duet	Personal agent that proactively monitors apps and schedules meetings	Personal agent entry point enters homogeneous competition
DeepSeek V4.1 Flash	Performance-cost Pareto improvement, HLE 42.6, DeepSWE 77.0	Run high-frequency tasks faster and cheaper
MiniMax M3.1 Flash	Space Bunny revealed; MoE 428B / 23B active, all-day $0.15/$0.6	Flooring the Agent task unit price, but hallucination rate is high
Gemini 4 Argon	Open worldwide, first on Vals Index, positioned "for your hardest problems"	Globally available flagship entry point for high-value vertical tasks
OpenAI thousand agents	Thousands of agents in parallel, 32 minutes to find leads from millions of brain-cell data	Agents aren't just tools; they're becoming part of research workflow
Eight-ministry plan	Agent adoption rate from 40% to over 90% by 2030	Agents enter the hard constraints of national industrial policy; a certainty signal

Three sentences to understand this wave: First, open source and closed source are accelerating simultaneously; open source approaches closed source on both parameter scale and unit cost. Second, world models and agents advance in parallel — one builds "playable worlds," the other grabs "workable entry points." Third, state and capital are both entering — agents appear in policy indicators for the first time, and the IPO lifts the lid on commercial payoff.

10. Afterword: What this wave means

Looking back at this month, the rhythm of the AI world has shifted from "a stronger model every month" to "several new tracks opening every month." World models, open-source trillion-parameter models, personal agents, agent scaling, national policy — they aren't isolated events; they're the fruits of the same inflection point.

When should you pay attention: If you're building agents, embodied intelligence, or enterprise intelligence upgrades, this month is a "worth a close look" signal month — whether on cost (Terra, M3.1), capability (Genie 4, K3.1 Pro), or policy (the "15th Five-Year Plan"), investing in agents now has clearer justification.

When not to panic: If you're just watching, most of these releases are still "research previews" or "entry point battles," a long way from large-scale deployment. Don't get dizzy over numbers like "one-ninth of single-task cost" or "32 minutes replacing months" — they describe optimal outcomes in specific scenarios, not that general replacement has already happened.

This column, the "LLM Large Model Technology Column," has gone from the Transformer attention mechanism all the way to LangChain, RAG, and MoE implementation, with a runnable code snippet for every piece, publishing weekly, focused on real usable models and agent practice. To read in order, subscribe on the column page.