Anthropic Claude, used where it earns its place
Claude is Anthropic's model family — Opus, Sonnet, and Haiku. We reach for it when an agent needs strong tool use, adaptive reasoning, and a million tokens of context. We never claim it's the only answer.
- Tool use & agentic loops
- Adaptive thinking + effort
- 1M-token context window
- Vendor-honest, no lock-in
The capabilities we lean on
Not a feature tour — these are the parts of Claude that change how an agent is built.
Adaptive thinking
Claude decides how deeply to reason per request and interleaves thinking between tool calls. We tune the effort level per route instead of hand-budgeting tokens.
Structured tool use
Typed tool schemas, parallel calls, and strict output formats let an agent take real action in your systems rather than just answer.
Long context
A 1M-token window holds a whole repo, a contract set, or a session's history in view — fewer retrieval round-trips, more coherent decisions.
Tiered models
Opus for the hard reasoning, Sonnet for the balanced middle, Haiku for cheap sub-tasks. We route by workload, not by habit.
Prompt caching
Caching the stable prefix cuts repeated-context cost to roughly a tenth. We design the prompt so the volatile part sits last.
Refusals & stop reasons
We branch on every stop reason — refusal, max-tokens, tool-use — so the agent degrades gracefully instead of crashing on an edge case.
Where Claude sits in an agent
The model is one component behind a clean interface — not the architecture itself.
Reason
Claude plans the next step with adaptive thinking, weighing tool results and conversation state.
Act
It emits typed tool calls; our harness executes them against your systems, with gates on anything risky.
Observe
Tool results, errors, and stop reasons flow back; context editing and compaction keep the window lean.
Govern
Every call is logged with model, tokens, and decision lineage — the same audit trail regardless of which model ran.
A model choice, not a dependency
Every Automatic.co agent talks to a model through an interface we own. Claude lives behind that interface alongside OpenAI and open-source models. Switching a route from Opus to GPT or to a self-hosted model is a config change, not a rebuild.
That's deliberate. Model leadership moves quarter to quarter, pricing shifts, and some data can't leave your perimeter at all. We pick the model per workload on the merits — and we'll tell you when Claude isn't the right call.
- Route per workload, swap by config
- Run on Anthropic API, Bedrock, or Vertex
- Pair with private/on-prem models where required
When we reach for Claude vs. when we don't
A working heuristic, not a rule. We validate per engagement.
| Reach for Claude | Reach for something else | |
|---|---|---|
| Task shape | Multi-step reasoning, heavy tool use | High-volume classification or extraction |
| Context | Whole codebase / document set in view | Short, bounded prompts |
| Data residency | API, Bedrock, or Vertex is acceptable | Must stay fully on-prem / air-gapped |
| Cost profile | Quality justifies the per-token rate | Cheapest token-per-task wins outright |
Frequently asked questions
Are you locked into Anthropic?
No. Claude is one model family we reach for, not a dependency. Our agents route to a model behind an interface, so we run Claude where its reasoning and long context earn it, and swap in OpenAI or an open model where those fit better.
Why Claude specifically?
Three reasons most often: strong multi-step tool use, adaptive thinking that scales reasoning to the task, and a 1M-token context window that lets an agent hold a whole codebase or document set in view. We pick it on the merits per workload, not by default.
Where does our data go?
Wherever the engagement requires. We can run Claude through Anthropic's API, through Amazon Bedrock or Google Vertex inside your cloud, or pair it with private/on-prem models for the parts that can't leave your perimeter.
How do you control cost and latency on Claude?
Prompt caching on the stable prefix, an effort setting tuned per route, streaming for long outputs, and the cheaper Haiku tier for sub-tasks. We measure tokens per workflow rather than guessing.
Bring the workflow. We'll pick the right model.
One working session to see whether Claude — or something else — is the right engine for your agent, and how it fits the rest of the stack.