Anthropic and OpenAI Found Product-Market Fit — What It Means for the AI Tools You Use

What Simon Willison Actually Said

On May 27, 2026, software engineer and AI commentator Simon Willison published a widely discussed analysis arguing that Anthropic and OpenAI have finally found product-market fit. After two years of jaw-dropping demos, eye-popping valuations, and legitimate questions about whether anyone would pay for AI at scale, Willison makes the case that both companies now have products people genuinely cannot work without.

The essay, which quickly hit 545 points on Hacker News, isn't just Silicon Valley navel-gazing. It's a signal that the AI tool market is maturing — and that the shakeout between tools worth paying for and tools worth ignoring is accelerating. If you're evaluating AI tools for yourself or your company, understanding this shift is critical.

Willison's core argument: the winners aren't the ones with the flashiest demos or the biggest models. They're the ones that solved specific, painful, recurring problems so well that people reach for them instinctively. That's the definition of product-market fit — and in 2026, only a handful of AI tool categories have genuinely achieved it.

Why Now? The Shift from Demos to Daily Drivers

For most of 2024 and 2025, the AI industry was trapped in what critics called the "demo treadmill" — each month brought a new model with a new trick, but few translated into sustained daily usage. People would try ChatGPT for a week, get bored, and drift back to their old workflows. Enterprise AI pilots gathered dust. The common refrain was: "AI is impressive, but I don't know what to use it for."

Something shifted in late 2025 and accelerated through early 2026. The shift wasn't a single breakthrough — it was the accumulation of dozens of small improvements that collectively pushed AI tools over the usefulness threshold:

• Context windows got massive: Gemini 3.1 Ultra's 2-million-token context and GPT-5.5's 1-million-token window meant you could finally feed AI tools entire codebases, not just code snippets.
• Agentic capabilities matured: Claude Code, Cursor, and Copilot went from suggesting the next line to autonomously completing multi-file tasks across an entire project.
• Reliability improved: Hallucination rates dropped from ~20% to under 5% on structured tasks. Not perfect, but good enough to trust with real work.
• Speed collapsed: Claude Sonnet 4.6 and GPT-5.5 Instant deliver responses in under 500ms. AI tools finally feel responsive, not sluggish.
• Integration went deep: MCP (Model Context Protocol) let AI agents connect to databases, APIs, file systems, and internal tools without custom engineering work.

The result? AI tools went from being "nice to try" to "can't work without them" — at least for specific user groups. That's product-market fit.

AI Coding Tools: The Clearest Product-Market Fit in Tech

If there's one category where AI tools have unambiguous, undeniable product-market fit, it's software development. Willison specifically highlights this as the strongest signal:

Claude Code has become the go-to AI coding agent for senior developers. It doesn't just autocomplete — it understands entire codebases, plans multi-file changes, writes tests, and catches bugs before you commit. A Stack Overflow survey found that 89% of developers now use AI coding tools daily, up from 44% in 2024.

Cursor has carved out a massive niche as the AI-native code editor. Its Agent Engine can take a natural language description like "add OAuth2 login with Google and GitHub" and implement it across 15+ files — including tests and documentation. Companies like Uber burned through AI budgets on tools like this, as we covered in our analysis of Uber's AI spending crisis.

GitHub Copilot remains the most widely adopted AI coding tool, embedded in the workflow of over 20 million developers. Its Copilot Agent mode, launched in late 2025, turned it from a suggestion engine into a true pair programmer.

The data is irrefutable: developers using AI coding tools are 30–55% more productive on measured tasks, and more importantly, they refuse to go back. That's the strongest possible product-market fit signal — the "pulling the product out of their cold dead hands" test.

AI Agents: The Category That Finally Justifies the Hype

2026 is the year AI agents went from science project to production tool. Willison's analysis notes that the "agentic" layer — where AI doesn't just answer questions but actually takes actions on your behalf — is where the real product-market fit lives.

Consider what these tools now do routinely:

• Claude's computer use agent can navigate software, fill out forms, and complete multi-step workflows by controlling a virtual desktop — essentially an AI employee that works 24/7.
• Microsoft's 365 Agents can monitor your inbox, draft responses, schedule meetings, prepare reports, and escalate issues — all autonomously. As we covered in our deep dive on Microsoft's enterprise agent strategy, companies are deploying thousands of these agents.
• Google Gemini Spark runs as a persistent AI agent that can research topics, book travel, manage your calendar, and handle personal admin. Our Gemini Spark review found it genuinely useful for everyday tasks.

The key insight from Willison: product-market fit for agents came not from intelligence but from reliability. Early agents were clever but unreliable — they'd work 70% of the time and fail spectacularly 30% of the time. The current generation works reliably 90–95% of the time on well-defined tasks. That's the difference between a toy and a tool.

Consumer AI: Where ChatGPT and Gemini Actually Deliver Value

Consumer AI has had a rockier path to product-market fit than developer tools. But Willison argues that the subscription numbers tell the story: ChatGPT Plus crossed 30 million subscribers in 2026, and Google One AI Premium is growing even faster thanks to Gemini's integration across Google Workspace.

The consumer use cases that have genuine stickiness are surprisingly specific:

The pattern is clear: AI tools that solve specific, frequent, time-consuming tasks have product-market fit. Tools that try to be general-purpose "smart assistants" for everything don't.

Which AI Tools Are Actually Worth Paying For in 2026

Based on the product-market fit analysis, here's our breakdown of which AI tools justify their subscription costs and which ones are still skating by on hype:

The AI Tool Categories That Still Haven't Found PMF

Willison's analysis also highlights what hasn't worked — and this is where many AI tool companies are quietly bleeding money:

• AI wrapper apps: Thin applications that just pipe prompts to GPT or Claude with a prettier UI. Users increasingly realize they can get the same results directly from ChatGPT or Claude. Churn rates for these apps exceed 60% monthly.
• AI for every business function: Tools promising to "revolutionize HR/marketing/sales with AI" often deliver marginal improvements over existing software. The AI tool rollback trend we covered shows companies are learning this the hard way.
• AI hardware: Humane, Rabbit, and even Apple's Vision Pro AI features haven't found a usage pattern that justifies the hardware cost. The AI hardware graveyard grows monthly.
• General-purpose AI assistants for enterprises: Companies want specialized tools for specific workflows, not a chatbot that can "do anything" but excels at nothing. As we explored in our piece on AI chat fatigue, conversation is giving way to delegation.

How to Spot an AI Tool With Real Product-Market Fit

Whether you're an individual choosing personal tools or a CTO building a company stack, here are the signals that an AI tool has genuine product-market fit — not just a good marketing team:

1. Users Complain About Missing Features, Not Existence

When people have product-market fit, they don't ask "why should I use this?" — they ask "why can't it also do X?" If the complaints are about missing capabilities rather than fundamental value, the tool has PMF.

2. Organic, Unsolicited Testimonials

The strongest signal is when people recommend a tool without being asked. When developers spontaneously tell colleagues "you need to try Claude Code," that's PMF. When someone shares a Cursor-generated code review on social media unprompted, that's PMF.

3. Users Create Workflows Around the Tool

If people are restructuring their work to incorporate an AI tool — changing how they plan projects, organize files, or communicate with teammates — the tool has become infrastructure, not a novelty.

4. Retention Over 6 Months

Most AI tools have great 30-day retention (the novelty effect) but terrible 6-month retention. Tools with genuine PMF maintain 60%+ retention at 6 months. Ask for retention data before committing to enterprise contracts.

5. The Tool Works Even When the Model Isn't Perfect

The best AI tools are useful even when the underlying model makes occasional mistakes. If a tool is only valuable when the AI is 100% accurate, it doesn't have sustainable PMF — because no AI model will ever be 100% accurate.

Frequently Asked Questions