State of AI 2026: What Actually Works

After three years of hype, AI is finally meeting reality.

The gap between a cool demo and a working business is closing fast. LLMs are writing production code, robots are landing on doorsteps, and autonomous taxis are actually turning a profit. We have moved from “what if” to “what is.”

Here is the current state of the frontier: from the infrastructure of open source to the rise of physical world agents, and the inevitable shift toward AGI and monetization.

The models

Most people still conflate two very different things when they say “AI.”

Conversational AI is a smart analyst. You ask it something, it responds. Text in, text out. This is what 99% of people use.

Agentic AI is a persistent worker. It clicks buttons, runs code, executes multi-step tasks over hours or days without hand-holding. This is where things are getting interesting.

Most of the model comparisons below are about the first kind. Agents get their own section.

Which model for what

Before the full breakdown, here is the quick version based on benchmarks and months of real usage:

Context window = how much text the model can hold in “working memory” at once. 200K tokens is roughly 150,000 words. 1M tokens is most of a novel.

I do not use Grok anymore. Perplexity replaced Google for me.

Claude: the enterprise bet

Anthropic did one thing right. They made a model calm and reliable enough that companies would feed it real data and sign real contracts.

The numbers: zero enterprise customers in 2023 to 300,000 by mid-2025. Revenue from $1B to $5B in eight months. Now a $350B company.

Why it worked: Claude’s first draft almost always holds. You ask for a pitch deck, you get a pitch deck. You ask for a contract review, it finds what matters. Other models require three rounds of refinement to get there. That difference is what enterprise buyers pay for.

The deals followed: Snowflake brought Claude to 12,600 customers for $200M. Accenture trained 30,000 people on it. The US Department of Defense signed a $200M contract (now on hold). Even Google uses Claude models in their Antigravity IDE.

The strategy was simple: do not build seventeen products. Just be the best conversational AI and embed yourself where work already happens. It worked.

ChatGPT: the everything app

OpenAI went the opposite direction and it is also working. They are building an operating system for AI, not just a model.

Recent launches: Prism (a free LaTeX workspace for scientists that drafts entire papers). Atlas (an AI-native browser). Codex App (multi-agent coding on autopilot). ChatGPT Go (an $8/month tier). Shopping integration with Shopify.

The moat is ecosystem lock-in. If ChatGPT is your browser, your code editor, your research tool, and your calendar, you are not switching to Claude even if Claude is technically better at one task.

They are also launching ads. Free and Go tier users will see ads at the bottom of responses. They are charging around $60 per thousand impressions, similar to NFL streaming. The pitch: the most targeted ads ever created because the AI knows exactly what you asked about.

Google: everything they need, not using it

Google makes their own chips, owns the internet, and employs more AI researchers than almost anyone. Gemini still feels one step behind.

It wins benchmarks. The context window is over a million tokens. But it does not have the conversational fluency that Claude and ChatGPT nailed. The Antigravity IDE launched in November, it is free, and most users run it specifically to access Anthropic’s models at no cost. That says everything.

The hardware story is more impressive. Their TPU v7 (Ironwood) delivers 4,614 teraflops per chip. Two to three times better performance per watt than GPUs. Four to ten times more cost-effective for large-scale training. And they just opened TPU access to third parties for the first time. Anthropic is now running Claude on Google TPUs.

Google has all the pieces. They are just slow to execute.

Open source and China

Most people still think Chinese AI is knockoff models trained on stolen data. That was maybe true two years ago.

Since August 2025, Chinese AI models have had more cumulative open-source downloads than western models. Not because they are free. Because they are good.

Moonshot AI’s Kimi K2.5 beats Gemini 3 Pro on SWE-Bench Verified. Outperforms GPT-5.2 on multilingual coding. Beats Claude Opus 4.5 on video understanding. Trained with roughly 1% of the resources top US labs use.

The strategy: efficiency plus open source. Train smaller, faster, cheaper models and release them publicly. It is driving adoption across the Global South where OpenAI and Anthropic pricing is prohibitive. That builds ecosystem lock-in at a geopolitical scale.

Western labs are spending billions to squeeze out marginal gains. Chinese labs are spending millions to get 80% of the performance. For most use cases, 80% is enough.

What you can actually do with this: by role

For professionals (accounting, legal, consulting, finance)

You have used AI to write emails faster. That is fine. It is also about 5% of what it can do.

Document automation. Give Claude a client brief with the problem, scope, deliverables, timeline, and budget, and it generates a full proposal. You edit instead of build from scratch. The time savings are real: two hours of editing versus eight hours of drafting.

For legal work, Claude is particularly strong at contract review. Upload the contract, ask it to flag unusual clauses and identify risk areas. It will not replace a lawyer. It will save the lawyer from reading 50 pages of boilerplate to find the three paragraphs that matter.

For financial models, describe what you need in plain language: revenue projections, scenario analysis, sensitivity tables. The model writes the formulas and structures the sheets.

Research workflows. ChatGPT with search integration wins for breadth: summarize recent industry developments, compare competitor strategies, pull regulatory updates. Claude wins when you already have the documents: upload five analyst reports, ask for consensus view on market size and key trends.

The key discipline: be specific. Do not ask it to “write a proposal.” Tell it the client name, the problem, the scope, the deliverables, the timeline, the budget, and give examples of past work. The more context you provide, the less you edit.

One rule on data: if you are handling sensitive client information, use paid enterprise tiers with data retention policies. Free tiers train on your inputs.

For developers

The last two years shifted the bottleneck. It used to be writing code. Now it is knowing what to build.

The three tools that matter:

• Cursor ($20/month) is the best all-purpose agentic IDE. Uses Claude Opus 4.6 under the hood. Best autocomplete in the industry. Strong at understanding large codebases across multiple files. Downside: $20 burns fast on complex refactors.
• Windsurf ($15/month) has faster context retrieval and handles large codebases better. Best choice for monorepos or legacy systems with sprawling, undocumented code. Less polished than Cursor, more cost-effective for heavy use.
• Claude Code (free with Anthropic Pro) is exceptional for agentic tasks. You describe what you want to build, it writes the code, tests it, debugs it, iterates. The catch: it consumes tokens aggressively and you will hit rate limits on complex projects.

What you can automate right now:

• boilerplate code generation: fully automated — CRUD operations, API endpoints, database migrations, test scaffolding
• refactoring: mostly automated — converting class components to hooks, migrating frameworks, adding TypeScript types
• debugging: partially automated — works well for clear errors, struggles with race conditions and memory leaks
• architecture and system design: not automated — the agent implements what you describe, it cannot make strategic decisions for you

The real cost: these tools are expensive at scale. Claude Opus 4.6 costs roughly $15 per million input tokens and $75 per million output tokens. A complex feature generation might cost $7.50 per run. Do that ten times a day and you are spending hundreds monthly before subscription fees.

The honest tradeoff: an 8-hour task costs the agent 2 hours and about $20. If your time is worth more than $10/hour, it pays. If you are a student on nights and weekends, it is prohibitive.

For non-developers

This is the most interesting category because the tools are making real automation accessible to people who do not code.

Cowork (Anthropic, launched January 2026) gives Claude access to your desktop. You grant folder permissions and it reads files, moves things, organizes downloads, creates documents, updates spreadsheets, drafts emails.

Practical use cases: planning travel by researching and outputting an itinerary directly to a Google Doc. Cleaning up your downloads folder by categorizing and moving files. Updating a budget spreadsheet by reading receipts from email. These are not demos. They work.

The security consideration is real: you are giving an AI access to your file system. Use it for low-stakes tasks first. Automate your downloads folder before your client communications. Build trust gradually.

Zapier Agents connect to 8,000+ apps and run 24/7 on natural language instructions. Strong use cases: lead research that scrapes LinkedIn and adds contacts to your CRM. Invoice reminders that check for overdue payments and send follow-ups. Content repurposing that turns a blog post into Twitter threads and LinkedIn drafts.

The limitation: these agents are narrow. They are excellent at repetitive tasks with clear rules. They break down when tasks require judgment or context that changes frequently.

For students and researchers

The challenge with AI for learning is that it makes it too easy to skip the hard parts. The hard parts are where learning happens.

Prism (OpenAI, 2025) is a free LaTeX-native workspace that drafts entire papers. You provide the research question, methods, data, and findings. It generates the full paper with introduction, literature review, methodology, results, and discussion.

For researchers, it is genuinely useful for the tedious parts: formatting equations, managing citations, structuring sections to journal guidelines. The intellectual work still requires humans.

For students, this is dangerous. If you use it to write your thesis without understanding the underlying research, you have learned nothing. You have outsourced the part of the process that builds critical thinking.

The right framework: use AI to accelerate tasks you already know how to do. Do not use it to replace tasks you are trying to learn. Use it after you have done the research. Not instead of.

The infrastructure wars

Nvidia’s cracking monopoly

Nvidia generated $200B in AI chip revenue in 2025. Their H100 delivers 156 teraflops. They have dominated for years.

That is changing.

Google’s TPU v7 (Ironwood): 192GB HBM memory, 7.2 TB/s bandwidth, 4,614 teraflops per chip. Two to three times better performance per watt than GPUs. Four to ten times more cost-effective for large-scale training. And now open to third parties for the first time. Anthropic is running Claude on Google TPUs.

Amazon Trainium generated multiple billions in 2025, growing at 150% per quarter. Trainium 3 just launched with a 50% cost reduction versus alternatives. Anthropic invested $4B in Amazon and is running Claude on Trainium chips in Indiana data centers.

The non-Nvidia alternatives market is growing faster than Nvidia’s market itself. The monopoly is not broken. It is bending.

Scaling laws and the four-month doubling

In October 2025, Andrej Karpathy said on Dwarkesh’s podcast: “It is the decade of agents, not the year of agents. They just do not work.”

He was right about the present. He underestimated the trajectory.

The length of tasks that AI agents can complete at 50% success rate has been doubling every 7 months from 2019 to 2025. In 2024 and 2025, that rate accelerated to every 4 months.

When ChatGPT launched in late 2022, it could handle 30-second tasks. Today, agents reliably complete tasks that take humans a few minutes, and can handle 2-hour tasks at reasonable success rates.

If the 4-month doubling continues: agents doing month-long tasks by 2027.

This is already starting. Cursor ran agents for close to a week building a browser. Over a million lines of code across a thousand files. Hundreds of agents working on a single codebase. Anthropic gave an agent control of a vending machine business and let it run for months managing cash flow and procurement. OpenAI is running agents continuously for two weeks to see what unlocks when you do not shut the model off.

Microsoft’s CEO called it “a new scaling law.” Not just bigger models. Longer inference time. More compute during the task itself, not just during training.

The physical world

Robots shipping to homes

For years, humanoid robots were impressive Twitter videos. That is changing in 2026.

Figure AI is targeting home robot shipments by end of year. OpenAI invested $5M. Their founder Brett Adcock predicts humanoids will perform unsupervised, multi-day tasks in new homes using entirely neural networks by end of 2026.

Tesla Optimus Gen 3 is launching in Q1 2026, targeting a million units of annual production. Right now it does simple factory tasks trained on footage of workers. By end of 2027, Musk says they will start selling to the public.

Boston Dynamics Atlas went fully electric in April 2024. 56 degrees of freedom. 7.5-foot reach. Lifts 110 pounds. Partnering with Google DeepMind’s Gemini Robotics AI. Deploying at Hyundai’s metaplant in Georgia in 2026.

1X Neo is the most interesting. You can pre-order it now. $20,000 upfront or $499/month. 5’6”, weighs 66 pounds, lifts 150 pounds. Runs at 22 decibels, quieter than a whisper. Autonomous for basic tasks (opening doors, fetching items, scheduling chores). Human takes over via VR telepresence for complex tasks. Ships to the US in 2026.

The wildest detail: 1X signed a deal to ship 10,000 units to industrial companies. They explicitly said their goal is collecting training data. With 10,000 robots running in homes, they will amass more daily data than YouTube uploads. They are making customers pay to be their R&D lab.

China’s robotaxi lead

China is not testing robotaxis. They are profitable.

Baidu Apollo Go: 1,000+ fully driverless robotaxis across 15 cities. 11 million completed rides. Per-vehicle profitability achieved in Wuhan. Expanding to Dubai, Germany, and the UK in 2026.

Weride: 1,023 vehicles. Fully driverless in Guangzhou. Deploying 1,200 robotaxis with Uber in the Middle East. Cut their autonomous kit cost by 50%.

Pony.ai: 1,159 vehicles. Operating in Shenzhen and Beijing. Partnership with Toyota. Per-vehicle profitability in Guangzhou in late 2025.

Goldman Sachs estimates the robotaxi market at $25B by 2030. Chinese companies are years ahead. They are not in beta. They are running a business.

Agents and what is next

Agents that work (sort of)

Simple definition: agents are AI models smart enough to predictably do anything on your laptop that a human would do. Mimicking clicks, typing, navigating apps, executing tasks over hours without supervision.

The most remarkable agent story of 2026 is OpenClaw. An Austrian developer named Peter Steinberger built an open-source agent that connects to Claude or ChatGPT and runs locally. It integrates with WhatsApp, Telegram, Discord, iMessage and more. Persistent memory. Feels like a digital employee. It got 145,000 GitHub stars and 20,000 forks in weeks.

Then the security researchers arrived. Palo Alto Networks called it a “lethal trifecta”: private data access, exposure to unreliable web content, and ability to communicate externally. They found tens of thousands of publicly exposed instances. The vast majority had critical authentication bypass vulnerabilities and were running outdated versions. Cisco’s summary: “the capability is groundbreaking, the security is a nightmare.”

So that is the current state of agents. Incredible potential. Massive adoption. Catastrophic security posture.

The enterprise-grade alternatives are safer but narrower. ChatGPT’s agent (launched July 2025) runs on a virtual computer, can work for up to an hour, and asks for confirmation before anything consequential. Zapier Agents connect to 8,000 apps and run 24/7. Both work well for repetitive, well-defined tasks.

The rule: agents work best when the task is specific and the stakes of failure are low. “Build a REST API with JWT authentication and rate limiting” works. “Make the app better” does not.

AGI: where we actually stand

AGI means AI capable of the full range of human-level intellectual tasks. Fluid movement between domains. Abstract reasoning. Learning from novel experience. Solving problems it was never trained on.

We are not there. Not close.

What we have: Claude Opus 4.6 approaching human-level at coding. AI performing junior-intern-level work across most knowledge fields. Agents completing multi-step tasks autonomously over hours.

What is missing: genuine continual learning (you cannot tell a model something once and expect it to remember). Reliable computer use across arbitrary interfaces. Consistent reasoning outside the training distribution. True multimodal integration.

The scaling law of task length doubling every four months is real and measurable. If it holds, we will see month-long autonomous tasks by 2027. But task completion is not general intelligence. Getting better at longer tasks is still narrow AI getting better at longer tasks.

Honest timeline: AGI is probably a decade away. Maybe more. The progress is real. The hype is ahead of the reality. Both things are true.

Ads coming to AI

OpenAI is launching ads. Free and Go tier users will start seeing them at the bottom of responses. Charging $60 per thousand impressions, the same as NFL streaming, triple what Meta charges.

The pitch: the most targeted, contextually relevant ads ever built, because the AI knows exactly what you asked about.

The concern: this changes the incentive structure. Right now, ChatGPT tries to give you the best answer as fast as possible. With ads, there is pressure to keep you in the interface longer. To give partial answers. To link to the advertiser’s source instead of the best source.

Google Search went through this exact transition. It started trying to get you off their site as fast as possible. Then ads became the business model and results got longer, more ambiguous, designed to keep you clicking.

ChatGPT could go the same way. Anthropic and Google are both exploring ads too. Anthropic needs a path to profitability. Google already runs the world’s largest ad network.

The question is not whether AI ads are coming. They are. The question is whether users will pay for ad-free access the same way they pay for YouTube Premium and Spotify. We are about to find out.

Conclusion

2026 is the inflection point. Not because everything suddenly works. Because enough works that the gap between demo and deployment is closing across every category at once.

Here is what the evidence shows:

1. Multiple strategies are winning simultaneously. Anthropic’s enterprise embedding and OpenAI’s consumer ecosystem are both working. Being the best at one thing and showing up where work already happens is sufficient.
2. Efficiency beats scale. Chinese labs are proving you do not need infinite resources. 1% of the compute to reach 80% of the performance is enough for most use cases.
3. The four-month doubling is the most important number in AI right now. If it holds, the capabilities in 2027 will look unrecognizable from today.
4. Custom silicon is viable. Google’s TPUs and Amazon’s Trainium are real alternatives to Nvidia. The chip monopoly is bending.
5. Deployment is real. Robots, robotaxis, coding tools. Not beta. Production.

The bottom line: we are still in the junior intern phase. The models are good enough to be genuinely useful and not good enough to replace expertise. The next three years will determine whether we get to Karpathy’s decade of agents or something faster.

What is certain: AI is no longer just chatbots. It is in our code, our cars, our factories, and soon, our homes.

The hype was early. The reality is catching up.

CT