GILD verdict
Standardize on the Claude API when Claude belongs inside your own product and engineers will own the integration.
Opus 5.5 holds the top score on the independent Artificial Analysis Intelligence Index v5.0 (Oct 27, 2026), so capability rarely blocks the decision. Cost does, so set a model per task, an organization spend limit and a retention arrangement before production traffic grows.
Claude API
AI Model API by Anthropic
+25 PositiveFit and limits
Who it fits
- You are building AI features or agents into your own product or internal tools, and have engineers to own the integration.
- You want the same models on Anthropic's API, Claude Platform on AWS, Microsoft Foundry, Amazon Bedrock, or Google Cloud, so procurement can run through a cloud provider's billing.
- Your workloads repeat long prompts or can run offline, where cache reads (5% of the input price on Opus 5.5 and Sonnet 5.5) and the 50% Batch API discount apply.
- A platform team wants to build an internal agent with the Agent SDK, which packages Claude Code's tools and agent loop for Python and TypeScript.
Where it falls short
- You buy model capability and own the product around it: evaluation and the user interface are your team's work.
- Fable 5.1, Fable 5, and the Mythos models require 30-day data retention and are not available under zero data retention unless Anthropic expressly authorizes it.
- On Amazon Bedrock and Google Cloud (Google managed), the Message Batches API, Files API, MCP connector, and code execution are not available.
- Each self-serve usage tier carries a monthly spend cap, from US$500 on Start to US$200,000 on Scale, and API usage pauses at the cap until the next month or a tier increase.
Implementation reality
Choose the model per task
Anthropic recommends Opus 5.5, at US$4 input and US$20 output per million tokens, for most workloads. It points demanding long-horizon work to Fable 5.1 and high-volume classification, extraction, and routing to Haiku 5.5. The Artificial Analysis Intelligence Index v5.0 (Oct 27, 2026) prices the spread: Haiku 5.5 scores 36 at US$0.41 per task, Sonnet 5.5 54 at US$10.02, and Opus 5.5 58 at US$10.04. Sonnet 5.5 generated 340 million tokens on the index, against a median of 63 million, so it costs about as much per task as Opus at half the list price. Run your own evaluation set on two adjacent models and compare cost per completed task.
Design for the cache
Cache reads cost 5% of the input price on Opus 5.5 and Sonnet 5.5, 2.5% on Fable 5.1, and 10% on most other models, and for most models they do not count toward input-token rate limits. Keep stable content such as the system prompt, tool definitions, and reference documents at the front of the prompt. The default cache lifetime is 5 minutes; a 1-hour cache costs 2x the input price to write. Hit rates differ by platform: Requesty's gateway traffic in April 2026 hit the cache on 77% of input tokens sent to Anthropic, 57% on Bedrock, and 24% on Google Cloud.
Put a ceiling on spend
Each usage tier carries a monthly spend cap: US$500 on Start, US$1,000 on Build, and US$200,000 on Scale. Set your own organization limit below the cap, and give each workspace its own spend and rate limits so one team's traffic cannot exhaust another's. Budget server tools on top of tokens: web search costs US$10 per 1,000 searches.
Settle retention and platform
By default, Anthropic deletes API inputs and outputs within 30 days and does not use them for training without permission. Zero data retention is available by request, per organization, and HIPAA readiness through a BAA. On Amazon Bedrock and Google Cloud (Google managed) the cloud provider is the data processor, so its terms apply instead.
What to measure in your pilot
- Track retries and output tokens per completed task on Haiku 5.5 for two weeks. If its cost per task passes half of Sonnet 5.5's on your workload, move that workload up a tier.
- Log cache reads as a share of input tokens per workload each week; a stable prompt prefix that stays below 50% means the prompt order needs fixing.
- GILD found no named team's cost split between the Agent SDK and Managed Agents; run one agent both ways for two weeks and compare session-hour charges with your container bill.
Sources
- Official documentation
- 20
- Public source
- 6
- Claude API docs: Models overview
- Claude API docs: Pricing
- Claude pricing
- Claude API docs: Prompt caching
- Claude API docs: Batch processing
- Claude API docs: Rate limits
- Claude API docs: API and data retention
- Anthropic Privacy Center: How long do you store my organization's data?
- Claude API docs: Data residency
- Claude API docs: Features overview
- Claude API docs: Claude Platform on AWS
- Claude API docs: Claude in Microsoft Foundry
- Claude API docs: Claude in Amazon Bedrock
- Claude API docs: Claude on Google Cloud
- Claude API docs: Claude Models on Google Cloud (Anthropic managed)
- Anthropic: Claude Haiku 5.5
- Claude API docs: Tool use
- Claude API docs: MCP connector
- Claude API docs: Managed Agents overview
- Claude Code docs: Agent SDK overview
- Artificial Analysis: LLM Leaderboard (Oct 28, 2026)
- Artificial Analysis: Changelog (Oct 28, 2026)
- Artificial Analysis: Claude Haiku 5.5 analysis (Oct 28, 2026)
- Artificial Analysis: Claude Sonnet 5.5 analysis (Oct 28, 2026)
- Artificial Analysis: Claude Opus 5.5 analysis (Oct 28, 2026)
- Requesty: Prompt-cache hit rate per provider, April 2026 (May 9, 2026)