Technology · Agentic → LLM Gateway · Commercial

Claude API

Taxonomy

Claude API is Anthropic’s hosted Messages API and Claude Platform surface for Claude models (claude-opus-5-5, Opus 5, Sonnet, Haiku, Fable/Mythos where available). It is a provider endpoint, not a multi-provider gateway: one vendor, adaptive thinking, tools, computer-use toolsets, prompt caching, batch, and first-party safeguard classifiers with optional fallbacks. Gateways such as LiteLLM and PortKey sit in front of it when you need routing, spend keys, or multi-vendor failover.

DocsStack: agentic-llm-gateway

Why it matters here

Opus-class models are the brain many coding agents and personal gateways call. Migration cost lives on this contract: thinking always adaptive, no forced tool_choice, thinking blocks bound to model + conversation prefix, and computer_toolset_20260801 on Claude API / Google Cloud. Keep Claude API on observe/propose. Safeguard refusals and opted-in fallbacks are ledger events — not silent chrome. tech-llm-gateway · tech-guardrails.

How it works

Clients send Messages (system, tools, content blocks) to Anthropic or cloud-hosted Claude ids. The model returns text, tool_use, and thinking blocks. Prompt cache reads/writes are billed separately. Effort steers adaptive thinking depth. Classifiers may refuse (stop_reason: refusal) or, when fallbacks are configured, hand work to an older Opus for cyber / bio / frontier-dev classes.

Related: Claude Code · LiteLLM · PortKey · Ollama · Browser Use · OpenClaw · tech-llm-gateway · tech-guardrails

Flow

When to reach for it

  • Use when: you need Anthropic’s hosted Claude models (Opus 5.5-class agentic coding / knowledge work) with first-party tools, cache, and published migration rules.
  • Skip when: you only need a multi-provider proxy — prefer LiteLLM / PortKey; or you need weights on your metal — prefer Ollama / vLLM.
  • Prefer instead: LiteLLM for fleet routing; Claude Code when the job is repo-shaped CLI, not raw Messages.

Limits

  • Contract breaks: disabling thinking, forced tools, replaying thinking after system/tools edits (new accounts), and legacy computer_20251124 on Claude API/GCP → HTTP 400.
  • Display trap: mid-tool text can sit in thinking blocks empty under default display: omitted — UIs go quiet between tools.
  • Safeguards: transparent blocks; cyber → Opus 4.8, bio/frontier-dev → Opus 5 when fallbacks opted in; no covert rewrite. Log answering model ≠ requested id.
  • Scope: provider API is not the company ledger or approval authority.

What we checked

Claims below are backed by science sources on disk.

Opus 5.5 product + pricing

Claude Opus 5.5 announcement · STRONG

“We're introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.”

Breaking changes / migration

Claude Opus 5.5 what's new · STRONG

“Four breaking changes affect code already running on Claude Opus 5: thinking can't be disabled, forced tool use returns an error, thinking blocks are tied to the model and the conversation, and, on the Claude API and Google Cloud, the earlier computer_20251124 computer use tool is not accepted.”

Specs / model ids

Claude Opus 5.5 overview · STRONG

“Model ID: claude-opus-5-5”
Research inventory

9 tags · 6 out · 7 in · 3 artifacts · 1 gaps · 0 corpus docs

Catalog tags

landscape.layer
Agentic
landscape.subcategory
LLM Gateway
license_tag
Commercial
maps.dm
absent
maps.features
0
one_liner
Agentic
review.depth
science
slug
claude-api
title
Claude API

Artifacts

  • dm_map · absent
  • features_map · present · technologies/claude-api/features.md
  • readme · present · technologies/claude-api/README.md

Out · alternative_to

In · alternative_to

In · in_stack