Anthropic named its three main model sizes after forms of writing: Opus, Sonnet, Haiku. Opus is positioned as the most capable of the three, Sonnet as the balance of capability and speed, Haiku as the fastest and cheapest. It reads like a minor branding detail, but it points at something the company has repeated in different forms since 2024: a tiered product line built for buyers who want to know in advance what they are going to get, rather than a single flagship model chasing the top of a leaderboard. That framing runs through most of what Anthropic has shipped since.

From Claude 3 to Claude 4

Claude 3 launched in March 2024 as the first release to use the Opus/Sonnet/Haiku split, and it set the pattern the company has followed since: ship a family, not a single model, and let buyers pick a point on the capability-versus-cost curve rather than accept one default. Claude 4 followed in May 2025, carrying the same three-tier structure forward while raising the ceiling on each tier. Sonnet 5 arrived in June 2026, and alongside it Anthropic introduced a new tier above Opus, called Mythos, built around a model named Fable 5. The addition of a tier above what had been the top of the line is itself a claim: that the ceiling for “raw capability” hadn’t been reached, and that some customers wanted something past it even at higher cost and latency. Whether that turns out to be a durable market or a narrow one is not yet settled from the outside.

Built for agents, not just chat

Anthropic has been explicit that coding and agentic work, meaning models that take multi-step actions and call external tools rather than just answering a single prompt, sit at the center of its roadmap. That focus shows up in the tooling built around the models as much as in the models themselves: Claude Code, a command-line tool for working with Claude inside a codebase, and Claude Cowork, aimed at longer-running collaborative tasks. Neither tool is a model in itself; both are bets that the value of a capable model increasingly shows up in how well it can be wired into a workflow, not just in how it scores on a static benchmark. This is a common industry direction, not one unique to Anthropic, but the company has leaned on it as a primary talking point rather than a secondary feature.

Safety as a design constraint, not an add-on

The other consistent thread in Anthropic’s public positioning is a stated approach to safety: that desired behavior should be built into a model during training, through methods like Constitutional AI, rather than patched on afterward with an external filter or moderation layer sitting between the model and the user. The distinction matters because it changes where the behavior is claimed to live. A bolted-on filter can be swapped, tuned, or bypassed independently of the model underneath it; a behavior trained into the model itself is harder to separate out and, in theory, harder to route around. It’s a design philosophy closer to how a car manufacturer might market crash structure engineered into the frame itself, rather than a manufacturer that markets airbags added late in assembly. The frame-level approach is the harder and slower one to build, and it is also the one that is harder for an outside party to verify by inspection alone.

That last point is the crux of Anthropic’s broader bet. Most labs in this series compete primarily on benchmark scores: reasoning tests, coding evals, context-window size. Anthropic’s stated positioning instead competes on predictability of behavior, the idea that a model doing roughly what you expect, consistently, across many runs and edge cases, is worth more to an enterprise buyer than a few extra points on a leaderboard. The bet underneath that positioning is specific: that once a model clears some threshold of raw capability, the customers who matter most start optimizing for trust and consistency rather than for squeezing out the last fraction of benchmark performance. It’s a wager on where the market’s attention shifts next, not a settled advantage. How safety behavior gets implemented in models generally, including the more common bolted-on approach, is covered in Guardrails and Moderation, and the deeper problem with any of these claims, that they are hard to verify from outside with a comparable public score, is the subject of Evaluating Safety: What Red Teaming Actually Measures.