Claude
Why it matters
Deep Dive
Claude is not one model but a product line that Anthropic has iterated on since the first release in March 2023. Claude 2 arrived later that year and pushed the context window to 100,000 tokens when most rivals handled a fraction of that. Claude 3, released in March 2024, reorganized the family into the Haiku, Sonnet, and Opus tiers still used today and added image input, while the 3.5 and 3.7 refreshes that followed cemented Sonnet's reputation as a leading coding model. Claude 4 continued the line in 2025 with a heavier focus on agentic workloads — models that plan, call tools, and work for extended stretches rather than answering single prompts. Underneath the branding, every Claude model is a transformer-based LLM trained with Anthropic's Constitutional AI recipe, which the company treats as a core product feature rather than a research sideshow.
One Family, Three Tiers
Within each generation the three tiers trade latency and price against capability, and learning to route between them is most of the cost-engineering game. Haiku is the small, fast, cheap option, aimed at classification, extraction, and high-volume chat where milliseconds and cents matter. Sonnet is the balanced workhorse that most production workloads default to, and it is usually where Anthropic's coding strength shows up first. Opus is the largest and most deliberate, reserved for hard reasoning, long-running agents, and tasks where the quality gap justifies a several-fold price difference. Because the tiers share an API and prompt format, a common pattern is to prototype on Opus, ship on Sonnet, and demote each step of a pipeline to Haiku as soon as measurements say it can cope.
Constitutional AI in Practice
Claude's characteristic behavior comes from a training method Anthropic calls Constitutional AI. Instead of relying only on human contractors to rank outputs, as in the standard RLHF pipeline, Anthropic writes down a set of principles — a constitution — and has the model critique and revise its own responses against them, then trains on the revisions. Part of the preference signal comes from the model itself rather than people, a variant known as RLAIF, which scales feedback far beyond what human labelers can produce. The visible result is a model that tends to explain its refusals instead of flatly declining, holds a consistent persona across long conversations, and follows nuanced instructions about tone and format. Developers often describe this as steerability, and it is a large part of Claude's appeal in products where the model's voice is customer-facing.
Long Context Is the Signature Feature
Long context has been Claude's calling card since the 100k window, and current models accept a context window of 200,000 tokens — several hundred pages — with larger windows available on some tiers. That changes what is worth building: instead of standing up a full RAG pipeline, you can hand the model an entire repository, a contract portfolio, or a book-length manuscript and query across all of it in one shot. The tradeoff is that every request re-processes the whole input, which is where prompt caching comes in: it stores a reusable prefix server-side and cuts both cost and latency sharply for repeated queries against the same material. Position still matters, since models recall information near the beginning and end of a long input more reliably than material buried in the middle. For genuinely large corpora, retrieval plus long context works better than either alone.
Beyond Chat: Code, Computers, and MCP
Anthropic has pushed Claude past the chat window in three directions worth knowing. Claude Code is an agentic coding assistant that runs in the terminal, reads a repository, edits files, runs tests, and iterates under supervision, and it has become a reference point for the coding-agent category. Computer use lets the model perceive a screen and operate a mouse and keyboard, an early but real step toward models that drive ordinary software instead of depending on purpose-built integrations. The third is the Model Context Protocol, an open standard Anthropic released in late 2024 for connecting models to external tools and data sources, which has since been adopted by other major AI providers and tooling vendors. Together these turn Claude from a text-in, text-out API into a platform for building agents.
Safety Training Is Not a Force Field
A common misconception is that Claude's safety training means you can skip your own defenses. Constitutional AI shifts the distribution of failures; it does not remove them. Claude still hallucinates plausible-sounding wrong answers, still drifts toward telling users what they want to hear, and remains jailbreakable with enough effort, and its knowledge cutoff guarantees it will eventually describe an outdated library or API with full confidence. The practical reading is that the training buys a lower base rate of bad behavior and more predictable refusals — genuinely useful when wiring a model into a product — but evaluation, input filtering, and output checks are still your job. Teams that treat vendor alignment as a substitute for their own guardrails tend to learn this the expensive way.