Somewhere in the last two years, Claude stopped being the chatbot people mentioned in the second half of a sentence. It’s now the default tool at plenty of engineering teams, the model a growing number of companies have wired into production, and the one writers keep open in a side window while they work. Anthropic AI, the company behind it, went from a research lab nobody outside safety circles had heard of to a business valued at roughly $61.5 billion in under four years.
That speed hides something odd. Anthropic is a safety research organisation that happens to sell a frontier model, or a frontier lab that happens to publish safety research, depending on who you ask. The order matters more than it sounds, because it explains nearly every strange decision the company has made.
Who is actually behind Anthropic
Anthropic was founded in 2021 by Dario Amodei and his sister Daniela Amodei, alongside a cluster of researchers who left OpenAI. Dario had been OpenAI’s VP of Research. Daniela had run operations and people. The departure wasn’t a dramatic falling out so much as a disagreement about pace: the group that left wanted more time for alignment and interpretability work than a company racing to ship products could justify.
That story keeps repeating. Ilya Sutskever walked away from OpenAI in 2024 to start a lab built around the same worry. The pattern is worth noticing: the people who have spent the most time inside these models tend to be the most nervous about them.
Claude, without the marketing gloss
The first Claude shipped in March 2023. Claude 2 arrived that summer with a 100,000-token context window, which was enormous at the time. The Claude 3 family landed in March 2024 across three sizes (Haiku, Sonnet, Opus), and 3.5 Sonnet in June that year became the model that genuinely changed how developers talked about the company. Claude 3.7 introduced extended thinking in February 2025, and the Claude 4 models followed in May.
If you want the underlying mechanics rather than the version numbers, this plain-English guide to how AI models actually work covers the transformer basics well. What matters practically is that Claude’s reputation rests on three things: holding long documents in view without losing the thread, writing code that runs, and prose that doesn’t read like a press release.
Constitutional AI, explained simply
Most early chatbots were tuned by humans ranking pairs of answers. That process is slow, expensive, and requires an army of contractors. Anthropic’s alternative, published in 2022, gives the model a written set of principles and has it critique and revise its own output against them. Humans still write the constitution. They just don’t have to sit in the loop for every single response.
The upside is consistency and auditability: you can read the principles and argue with them. The downside is that a constitution is a value choice nobody voted on, and models trained this way can get preachy or hedge when a direct answer would serve better. Anyone who has asked Claude a blunt question and received three paragraphs of careful qualification has met the trade-off firsthand.
What the safety research actually produced
Anthropic’s interpretability team has done work that is verifiable rather than rhetorical. In 2024 they pulled millions of interpretable features out of Claude 3 Sonnet using dictionary learning, including one that represented the Golden Gate Bridge. Amplify that feature and the model would start steering every conversation toward the bridge, unprompted. It’s a small result with a large implication: researchers are beginning to see inside the black box rather than guessing from outputs.
The company also publishes a Responsible Scaling Policy, which ties model capabilities to safety levels (ASL-2, ASL-3 and so on) with defined thresholds that trigger stricter safeguards. You can disagree with the thresholds. You can’t say they weren’t written down in advance.
Where the criticism lands, fairly
Sceptics make three arguments worth taking seriously.
- Safety as a moat. If compliance rules are strict and expensive, they favour whoever can afford them. A lab that lobbies for them benefits, whether or not that was the intent.
- Commercial pressure is real. Anthropic has taken billions from Amazon and Google, both of whom need Claude to justify the compute. Warning about risk while selling the risky thing is a tension the company lives with rather than resolves.
- Capability still outruns caution. Claude 4 scores near the top of most benchmarks. The safety framing has not slowed the frontier down, which raises the question of what it is actually buying.
Meanwhile the wider field is splitting. Mira Murati’s new lab is built on a bet against the standard training recipe, and other founders are making similar wagers. Anthropic’s bet is narrower: keep the architecture, fix the alignment.
Claude in real work
The clearest place Claude earned its reputation is code. Claude Code, released in 2025, runs in a terminal, reads a repository, edits files and iterates against test failures. Developers who were sceptical of autocomplete-style assistants tend to find agentic coding a different experience entirely.
The other quiet win is the Model Context Protocol, which Anthropic open-sourced in late 2024. It’s a standard way for models to talk to outside tools and data sources. It sounds dull and it’s the reason a lot of enterprise integration work got simpler overnight, because vendors could build one connector instead of five.
The business, briefly
Amazon has committed $8 billion. Google has invested too, and supplies significant compute. Revenue is overwhelmingly enterprise rather than consumer, which is a deliberate contrast with OpenAI’s app-store-of-everything strategy. If you want the fuller picture of how the two rivals compare, this breakdown of OpenAI’s models, funding and roadmap in 2025 puts the numbers side by side.
For teams evaluating vendors, the practical advice is unglamorous: run a real pilot on your own data, get pricing in writing, and check what happens to your data. This seven-step playbook for buying from AI companies walks through the procurement side properly.
If you just want to test it yourself
The free tier gives you a capped number of messages on a mid-size model. Pro costs about $20 a month and unlocks the stronger models plus higher limits. The API is metered per token, and for most small projects the bill stays under the price of a sandwich per day.
Give it a task with a clear right answer before you judge it. Paste a 60-page contract and ask for every clause that shifts liability. Drop in a broken script and ask what fails. The subjective tests people run on social media mostly measure vibes. A structured 45-minute bake-off against your own workload tells you far more about whether it fits your job.
What’s worth watching next
The interesting question for Anthropic isn’t whether Claude gets better. It will. The question is whether the interpretability work matures fast enough to matter. Being able to identify a feature inside a model is a long way from being able to certify that the model is being honest with you, and that gap is where the next few years of argument will happen.
Watch three things. Whether ASL-3 triggers get crossed and what the company does when it’s their own revenue on the line. Whether agentic coding tools move from developer novelty to standard infrastructure. And whether the safety positioning survives a genuine price war, because that’s the moment every stated principle gets tested against a spreadsheet. Anthropic has been unusually explicit about where it draws lines. The next 18 months will show how much those lines cost.