This course reaches its final chapter at last. Here, we'll take a sweeping look at what's happening on the front lines right now. Going multimodal, reasoning models, AI agents, open source, robotics, and AGI — let's grasp the "current position" of a world whose scenery changes every six months, without jargon. Each number comes with its source and date; product and model names are as of September 2026.
A 5-minute overview of AI's "now"
Feeling the speed of AI's evolution
As of September 2026, when this chapter was last updated, "the common sense of six months ago no longer applies" has become the watchword of the AI industry.
Numbers make the pace easy to see. According to the research firm CB Insights, private AI companies raised a record $225.8 billion in 2025, nearly double the 2024 total[1].
Source: CB Insights
Gartner forecast (August 2025)
Official MCP blog (December 2025)
Let's look back over the past year and a half through the major events, in chronological order.
From here, we'll dig into four especially important trends.
Multimodal AI — AI with senses
Multimodal AI means AI that can handle multiple formats in an integrated way — text, images, audio, video, and more. Early LLMs were "things that could only read and write text," but today's AI can see photos, hear voices, and make videos.
What multimodal AI can do
| Field | Service | What it can do |
|---|---|---|
| Image generation | ChatGPT native image generation | Accurately generates images with text in them. Right after its March 2025 release, demand was so high that OpenAI CEO Altman described the GPUs as "melting" |
| Video generation | Google Veo 3 | Generates videos with sound. In Flow, Google's filmmaking tool built on Veo, more than 275 million videos had been made by October 2025 (Google) |
| Long-context understanding | Claude Fable 5.1 | Handles a 1 million-token context (more than a full book) and up to 128K tokens of output (as of September 2026, per Anthropic's official docs) |
| Voice conversation | ChatGPT Advanced Voice mode | Natural real-time voice conversation without going through text. Can even be used like an interpreter |
Meanwhile, in March 2026 OpenAI announced it would shut down its video-generation AI "Sora": the web and app experiences ended on April 26 and the API on September 24 (OpenAI Help Center).
💡 A practical hint: Image analysis (photo → text) is available in almost every company's free tier. Try it in everyday life — reading receipts, turning handwritten notes into text, extracting numbers from charts, and so on.
The reasoning-model revolution — the arrival of AI that "thinks"
From late 2024, a new category was born in the world of AI: reasoning models.
Conventional AI had a "answer instantly when asked" style. Reasoning models are different. They take "time to think" before answering. It's the same as when a human solving a math problem doesn't just write the answer straight away, but thinks step by step while taking notes.
Why it matters
With the arrival of reasoning models, performance improved dramatically in the areas AI was bad at — math, science, and complex programming.
Scored 92.7% on AIME 2025, a qualifying exam for the US Math Olympiad team, and 99.5% when given a Python interpreter (OpenAI).
DeepSeek's technical report puts the GPU cost of the final training run of its base model, V3, at about $5.58 million (excluding earlier research and experiments). R1 was released on January 20, 2025 under the MIT license, billed as performing on par with OpenAI o1[2].
A unique feature letting developers freely set a "thinking budget." Implements "interleaved thinking," continuing to think even while using tools.
In 2026, reasoning moved from a special mode to standard equipment. On 1 September 2026 Anthropic added Claude Fable 5.1 for demanding reasoning and long-running agent work, and on 22 September it released Claude Opus 5.5. Anthropic's official model overview says "if unsure, start with Opus 5.5 for most workloads; use Fable 5.1 for demanding reasoning and long-horizon agent work" (Anthropic). Today's top models from every major vendor are built to think before they answer. According to the official documentation as of 5 October 2026, Opus 5.5 has thinking always on; OpenAI's flagship GPT-6 Astra (announced 3 September 2026) and the balanced GPT-6.1 Sol let you choose reasoning effort from "low" to "max" (GPT-6.1 Sol has no no-reasoning option); and Google's Gemini 3.8 Flash also thinks by default[4].
That said, newer does not mean better at everything. As our Opus 5 review sets out, Opus 5 pulls ahead on work that runs autonomously for a long time, while on single-patch coding tasks (DeepSWE v1.1) the order flips: Opus 5 scores 68.8% and GPT-5.6 Sol 72.7%. Strengths stay split even as generations advance, so when a new model arrives, test it on your own work before switching.
⚡ Point to watch: inference-time compute
The discovery that "letting it spend more time thinking yields more accurate answers" added a new axis to AI's evolution. In addition to the conventional "increase training data" and "make the model bigger," it turned out that increasing the amount of computation at inference time also improves performance.
AI agents — toward an era of "delegating"
As of 2026, the hottest keyword is AI agents.
Until now, AI was an advisor that "answers when you ask." AI agents are different. Tell it a goal, and it makes a plan, uses tools, and carries out the task autonomously. It's close to the feeling of "delegating" work to a secretary or assistant. If you want to know the mechanism in more detail, see "What is an AI agent," and for how multiple AIs work in coordination, "What is multi-agent" as well.
Concrete examples of AI agents
| Agent | What it can do | Characteristics |
|---|---|---|
| Claude Code | Autonomously handles code generation, execution, and debugging end to end | Reached $1 billion in annualized revenue in November 2025, about six months after its public launch (Anthropic announcement) |
| ChatGPT Work | Takes a goal and uses web search, files, and connected tools to deliver reviewable output such as documents and analyses | Successor to the line that ran from the browser agent "Operator" (2025) to "ChatGPT agent" (July 2025). Agent mode is no longer available as of September 2026, and the official help page directs users to Work |
| Manus AI | Executes complex tasks asynchronously in the cloud | Appeared in 2025. Chinese authorities blocked Meta's roughly $2 billion acquisition; after a temporary interruption of access, Manus announced on 1 September 2026 that it had resumed independent operations |
| Devin | An AI engineer that autonomously carries out software development | At its 2024 launch it resolved 13.86% of SWE-bench issues (Cognition). As of September 2026 it has a free plan and paid plans from $20/month (official pricing) |
🛠 If you want to seriously use a code-writing "partner," our sister course "AI Coding in Practice" explains everything from getting started with Claude Code and the like to mastering it. For people who want to build apps solo, we also recommend "Indie Development with AI."
MCP — the "common language" of AI agents
As the standard for connecting agents to external tools, MCP (Model Context Protocol), developed by Anthropic, is spreading rapidly. In December 2025 it was donated to the Agentic AI Foundation, a new foundation under the Linux Foundation; at that point it had over 97 million monthly SDK downloads, with support in major AI products such as ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, and VS Code (official MCP blog).
Gartner predicts that 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025[1].
⚠️ The challenges of agents: Agents are convenient, but at present they have several important limitations. Complex judgment errors, security risks (such as sending information on their own), cost (repeatedly making API calls autonomously), and murky accountability. The iron rule is "delegate and verify," not "delegate and forget."
The rise of open-source AI
Commercial AI like ChatGPT and Claude isn't everything. Open-source AI — free to use and free to modify — is evolving at an incredible pace.
Representative open-weight models today (October 2026)
| Model | Developer | Characteristics |
|---|---|---|
| Muse Glimmer | Meta | Released in August 2026. A 30B model distilled from Muse Spark, Meta's main model, and it can read images. Designed to run on your own hardware, with weights under Apache 2.0 (Meta, model card) |
| DeepSeek V4.1 Flash | DeepSeek (China) | Released 10 September 2026. The smallest model in DeepSeek's new architecture family, yet still a 552B-parameter MoE. Reads images and handles up to 1 million tokens of context. Weights under the MIT license (DeepSeek, model card) |
| Qwen3.8-27B | Alibaba | A compact 27B model from Qwen3.8, which Qwen calls the most capable generation of its open models. Reads images and video, with thinking mode on by default. This 27B model is Apache 2.0 licensed (Qwen model card). The Qwen family overtook Llama in cumulative downloads in September 2025 (ATOM Report) |
Meta's Llama 4 and DeepSeek R1, which put open models on the map in 2025, are now a generation behind. Open source turns over even faster than commercial models — the line-up changes within months. To see which model is strongest right now, check download counts on Hugging Face Models or the leaderboard at LM Arena directly.
🖥 Want to run it on your own PC? Open-source models can run in your own environment too. For how to get started, "How to get started with local LLMs" is helpful, and for which model to choose, "Local LLM model comparison."
Why open source matters
Open-source AI has five significances.
You can inspect how the model works and assess its safety.
You can build a specialized model on your own data.
Run it on your own servers and API fees drop to zero.
You can use AI without sending data outside.
Prevents AI monopoly by a handful of big companies.
According to the ATOM Report (2026), in late July 2025 models from China overtook models from the US in cumulative open-model downloads on Hugging Face, with most of the growth coming from Alibaba's Qwen. The geopolitical balance of AI development is shifting.
💡 A point for general users: Open-source AI is mainly for companies and developers, but its benefits reach everyone indirectly. As competition intensifies, commercial AI's prices fall and its performance rises. DeepSeek R1, for example, claimed o1-level performance yet launched at a low API price of $2.19 per million output tokens[2].
The AI ahead — the outlook for the next few years
Finally, let's look at where AI is headed over the next few years through three lenses: robotics, AGI, and national AI strategies.
Figure 03 is at work at BMW's Spartanburg plant in the US. 1X NEO takes home orders at $20,000 (or $499 a month), with shipping planned for 2026. Tesla is building an Optimus production line at its Fremont factory. Japan's AI Basic Plan also names physical AI among its "paths to win."
Anthropic expects powerful AI in "late 2026 or early 2027," Google DeepMind finds it "plausible by 2030," and OpenAI CEO Sam Altman wrote that "we know how to build AGI." A survey of AI researchers puts the median at 2047. The new scaling axis of investing compute at inference time is also drawing attention.
The US pushes deregulation with its AI Action Plan, the EU applies the risk-based AI Act, China pairs its "AI+" drive with rules for generative AI, and Japan puts "trustworthy AI" and physical AI at the center. The balance between promotion and regulation differs by country.
AI × robotics — putting "moving AI" into practice
Humanoid robots that combine the intelligence of LLMs with the body of a robot are finally starting to be active in real-world settings.
- Figure 03 — at work in logistics at BMW's Spartanburg plant in the US (June 2026, Figure). Figure raised over $1 billion in September 2025 (Figure)
- 1X NEO — a home robot 1X calls "the world's first consumer-ready humanoid." Preorders opened in October 2025 at $20,000 (or $499 a month), with shipping planned for 2026 (1X)
- Tesla Optimus — a production line is under construction at the Fremont factory, with production anticipated later in 2026 (Tesla Q2 2026 Update)
Japan's AI Basic Plan also names the development and deployment of "physical AI" — AI that moves robots and other machines in the real world — as one of Japan's paths to win[3].
The road to AGI — experts' predictions
On when AGI (artificial general intelligence — AI with intelligence equal to or greater than a human's) will arrive, industry views are split. For the meaning of the term, see "What is AGI (artificial general intelligence)," and for the "technological singularity" where AI surpasses humans, we explain in detail in "What is the technological singularity."
| Position | Prediction |
|---|---|
| Anthropic | "Late 2026 or early 2027" — expects "powerful AI" with intellectual capabilities matching Nobel Prize winners across most disciplines (March 2025, recommendations to the US government) |
| OpenAI | "We know how to build it" — CEO Sam Altman wrote in January 2025 that "we are now confident we know how to build AGI as we have traditionally understood it," without giving a date (Altman's blog) |
| Google DeepMind | "Plausible by 2030" — while highly uncertain about timelines, finds it plausible that powerful AI systems will be developed by 2030 (April 2025, safety paper) |
| Survey of AI researchers | "2047" — aggregating predictions from 2,778 researchers, the chance of machines outperforming humans in every task reaches 50% by 2047 (conducted in 2023, paper) |
Even when people say "AGI is coming," it doesn't mean life will change dramatically overnight. Still, it's a fact that the range of what AI can do is expanding almost every month. The assumption that "AI probably still can't do this" may well be out of date six months from now.
AI strategies of major countries and regions — the US, EU, China, and Japan
Governments are competing over AI too, and their approaches differ sharply: some put promotion first, others set rules according to risk. Here is how four major players compare (as of September 2026).
Main policy: America's AI Action Plan (July 2025) — over 90 federal actions across three pillars: accelerating innovation, building AI infrastructure, and leading in international diplomacy and security (White House). A January 2025 executive order revoked the previous administration's 2023 AI order (executive order)
Stance: "Win the AI race" by removing regulation. A December 2025 order set up a task force to challenge state AI laws (executive order), and in March 2026 the White House proposed a national legislative framework to Congress (White House)
Main policy: AI Act — in force since August 2024. It sorts AI into four risk levels: bans such as social scoring have applied since February 2025, rules for general-purpose AI models since August 2025, and most of the rest since August 2, 2026 (European Commission)
Stance: One law with obligations scaled to risk. An amendment (the "AI Omnibus," in force since July 2026) moved the rules for high-risk uses such as hiring and education to December 2027
Main policy: Interim Measures for Generative AI Services (in force since August 2023) — services that can shape public opinion must pass a security assessment and file their algorithms (CAC). Rules requiring AI-generated content to be labeled took effect in September 2025 (CAC)
Stance: "Equal emphasis on development and security." The State Council's "AI+" action (August 2025) targets adoption of new intelligent devices and AI agents above 70% by 2027 and above 90% by 2030 (State Council)
Main policy: AI Promotion Act (2025) and AI Basic Plan — first adopted in December 2025 and revised in July 2026. The goal is to become "the most AI-friendly country in the world" with "trustworthy AI" at its core[3]
Stance: Promotion first. The revised plan names domain-specific "vertical AI" and physical AI as Japan's winning strategy, and the government AI platform "Gennai" is used by about 180,000 civil servants (Cabinet Office)
In short, the US leans on deregulation, the EU on rules by risk level, China on state-led promotion alongside rules for content, and Japan on promotion built around "trustworthy AI." Many other countries have national AI strategies of their own. For what these rules mean for you as a user, see "Regulatory trends worldwide" in AI Risks and Ethics.
What to do "now" in the age of AI
Actually experience it with free AI tools. One try beats a hundred hearsays.
AI is a tool. It becomes valuable combined with your expertise and creativity.
An era where common sense changes every six months. Staying curious is the greatest skill.
- AI has gone multimodal, able to handle not just text but images, audio, and video.
- With reasoning models, AI that "thinks" became real, and with AI agents, AI you can "delegate" to.
- Open-source AI is surging, and the geopolitical balance is shifting too.
- The next battlegrounds are robotics, AGI, and national strategies. Common sense changes every six months.
- What matters are three things: touch it and get used to it, combine it with your strengths, and enjoy the change.
References
- CB Insights. "State of AI 2025 Report." January 13, 2026. cbinsights.com / Gartner. "Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025." Gartner Newsroom, August 26, 2025. gartner.com
- DeepSeek. "DeepSeek-R1 Release." DeepSeek API Docs, January 20, 2025. api-docs.deepseek.com
- Cabinet Office, Government of Japan. "Artificial Intelligence Basic Plan." Cabinet decision, December 23, 2025; revised July 14, 2026. www8.cao.go.jp
- Anthropic. "Models overview." Claude Docs, 2026. / Anthropic. "What’s new in Claude Fable 5.1." September 1, 2026. / OpenAI. "Models" / "GPT-6.1 Sol." OpenAI API Docs, 2026. / Google. "Gemini thinking." Gemini API Docs, 2026.
Related links:
- Hugging Face Models — a hub for open-source AI models
- LM Arena — a leaderboard for comparing AI model performance
From the fundamentals of AI to the latest trends, you've built up a systematic body of knowledge. AI evolves by the day. Building on the foundation you learned in this course, keep touching the tools while continuing to catch up on the latest information. If you want to master AI further as a partner for writing code, head to our sister course "AI Coding in Practice"; if you want to build a service solo, to "Indie Development with AI."