Claude is Anthropic’s family of AI models: Haiku, Sonnet and Opus. Haiku is the fastest and cheapest of the three, Sonnet is the one most people should be using, and Opus is for the hardest work. Claude is best at coding and long-running agent work, and after that at safety, which Anthropic builds in during pre-training rather than adding on afterwards.
Haiku, Sonnet or Opus: which one do you use
The three tiers have been in place since March 2024. They exist so you can trade speed and cost against depth.
| Model | Speed and cost | What you run on it |
|---|---|---|
| Haiku | Fastest, cheapest | The same small job thousands of times: sorting inbound email, tagging jobs, one-line replies |
| Sonnet | Middle on both | The default for most work, including most day to day coding |
| Opus | Slowest, most expensive | Long multi-step runs, and reasoning you cannot afford to get wrong |
Default to Sonnet. Move up to Opus when a task runs for hours or a wrong answer costs real money, and down to Haiku when volume matters more than depth. The video does not give prices, and what you pay turns on whether you are on a subscription or usage-based plan.
Who makes Claude, and who owns Anthropic
A group of OpenAI researchers left because they did not think the models were being built safely enough. They founded Anthropic, started building to rival OpenAI, and shipped the three models we use today. The company is registered as a public benefit corporation rather than the for-profit or non-profit structures OpenAI has run under. Its backers include Google, which owns Gemini, and Amazon.
What constitutional AI actually means
Anthropic writes an ethical constitution, a set of rules about what the model is and is not allowed to do, into pre-training. Pre-training is the stage where the model is still building its understanding of the world, so those rules sit underneath everything it later learns to say.
OpenAI does the equivalent job after training, in the system prompt. That difference has one practical consequence: a system prompt can be overridden. A constitution embedded at pre-training is much harder to talk the model out of.
Every Claude model, in order of release
| When | What shipped | What it changed |
|---|---|---|
| Early 2023 | Claude 1 | Arrived about four months after ChatGPT |
| Late 2023 | 100,000-token context window | Could digest some of the largest documents around at the time |
| March 2024 | Haiku, Sonnet and Opus | The three-tier line-up |
| Late 2024 | Sonnet 3.5 | The turning point for coding agents |
| Mid 2025 | 1,000,000-token context window | Long-running tasks instead of small point solutions |
| Early 2026 | Models up to 4.8 | Multi-hour work, with the model checking and iterating on itself |
Claude’s context window has been 1,000,000 tokens since mid 2025. That number is what turned coding agents from something that patched one function into something that works through a whole problem.
Fable 5 arrived about two weeks ago and was then taken off the market, so as of June 2026 everyone is waiting for it to come back. We have used it. It is an excellent coding model with strong applications for agentic work and agentic ops.
What Claude is best at
Coding, and the reason is partly historical. Claude was one of the first models behind a product aimed directly at software work. The coding tools before that were wrappers: Cursor and Windsurf, with Claude or Codex doing the actual work under the hood. Claude Code was Anthropic’s own coding product, the one you point at your codebase and use directly instead of through somebody else’s app, and running it gave them the data to train the next models on.
Two things make coding a good fit. Long-running tasks, which is what most real software work is, and reasoning. Those are the two capabilities Anthropic has pushed hardest on, and they are the pillars coding rests on.
The other reason people choose Claude is safety and reliability. If you work in finance or healthcare, Claude is the safer model to reach for. You may not be a bank, but a property manager holding tenant records and card details has a smaller version of the same problem. Some people also rate Claude highest for creative writing.
Claude vs ChatGPT: which should you use
Pick on the job, not the brand. Day-to-day drafting and summarising come back at much the same standard from every frontier model, so for that work the answer is whichever seat your team already has. Two things genuinely separate Claude, and neither shows up until you push past a single reply.
The first is software work that runs long. Anthropic put a developer product in front of engineers before its rivals did and trained the next generation on what came back through it, which is why Claude holds up over runs measured in hours. The second is where the guardrails sit. Anthropic embeds them before the model has finished learning, while OpenAI applies its equivalent afterwards at the instruction layer, which is the easier of the two to talk a model out of. That difference earns its keep in regulated work and almost nowhere else. Gemini barely enters the argument, and Google is one of Anthropic’s backers, so those two are not the clean rivals the branding suggests.
There is a less technical factor, and it decides more of these choices than benchmarks do. We started using Claude in 2024 because it was better than anything else available at the time, and we are still on it. Other models might be ahead this month. A year of knowing how a model responds to the way you write prompts is usually worth more than that.
