senn-techAI Radar: Anthropic's Claude-4 Family (Opus, Sonnet, Haiku)
Anthropic tiers its current Claude-4 family into three classes — and exactly this tiering is what makes it interesting for practice.
The Three Sizes
- Claude Opus 4.8 (
claude-opus-4-8) — the strongest model, excellent for large codebases, multi-step reasoning, and agentic workflows. - Claude Sonnet 4.6 (
claude-sonnet-4-6) — the balance of quality, speed, and price for daily business. - Claude Haiku 4.5 (
claude-haiku-4-5) — fast and inexpensive for bulk tasks (classification, extraction, routing).
Which Size Goes Where
The tiering follows no marketing logic, just the costs. Opus pays off only where its reasoning makes a real difference — complex code refactors, deep document analysis, multi-step agentic runs. For most everyday text, summaries, and standard coding tasks, Sonnet suffices. Haiku takes on the high-volume routine, where speed and price matter. Routing every call through the largest model pays a premium for capabilities it does not need.
Reasoning and Cost Control
The models support an extended thinking mode: on hard tasks, Claude can run a visible or hidden chain of thought before producing the final answer. That raises hit rates on multi-step plans, code refactors, and complex document analysis — and costs additional tokens and latency. This is exactly where the tiering pays off: Opus with extended thinking where it counts; Sonnet without the deep mode where a solid draft is enough; Haiku for the many small calls where every millisecond and cent matters. Sending everything through Opus costs more without getting more.
License: Closed, API-Only
Like Gemini and GPT, Claude is proprietary and API-only — no open weights, no self-hosting, no local fine-tuning on company data. The entire family runs in Anthropic's cloud. For teams that care about data sovereignty there is no official Claude path beyond the API; the answer to on-premise lies with other vendors.
Why It Concerns Us
This website is built with Claude Code — Anthropic's agentic coding tool that builds on the Claude family. In everyday use, the tiering means: deploy the expensive model only where it counts, and handle the rest with Sonnet or Haiku. Sticking to that consistently keeps costs for coding-agentic workflows noticeably lower, without a real loss in quality. In practice it means: routine tasks like test generation or documentation run on Sonnet, long refactors and architecture decisions on Opus, and pure classification or routing steps in a pipeline on Haiku.
The GDPR Caveat
Claude runs as a cloud API. For sensitive data, we combine the best of both worlds: cloud models for non-critical tasks, own on-prem models (vLLM, Ollama) where data must not leave the premises. Which model for which purpose — exactly what we plan with our customers. One criterion is always the sensitivity of the inputs: public marketing material may go to the cloud; customer and contract data stays in-house. Drawing that line consistently lets you use the strengths of the Claude family without putting data sovereignty at risk.
Further Reading
When is the expensive Claude Opus worth it, and when does Sonnet suffice?+
Opus pays off only where its reasoning makes a real difference — complex code refactors, deep document analysis, and multi-step agentic runs. For the overwhelming majority of everyday text, summaries, and standard coding tasks, Sonnet is the better choice. Haiku covers high-volume routine work like classification or routing. Routing everything through Opus pays for capabilities you do not need.
Can we run Claude models on our own hardware?+
No. Claude is proprietary and API-only — no open weights, no self-hosting, no local fine-tuning. For teams that care about data sovereignty there is no official Claude path beyond the API. For sensitive data we combine cloud models with our own on-prem models (vLLM, Ollama); on-prem then lies with other vendors.
What does extended thinking mean for cost?+
On hard tasks Claude can run a chain of thought before the final answer — that raises hit rates but costs additional tokens and latency. The tiering helps: Opus with extended thinking where it counts; Sonnet without the deep mode for solid drafts; Haiku for the many small calls. Routing all three blindly through Opus burns budget.