Claude API Models Overview and Selection Guide for 2026

I run this blog on a CMS and AI writing pipeline I built for my own site, so deciding which model handles which step is a recurring decision for me. The mistake I'd steer you away from is picking a model on benchmark headlines instead of workload fit.

The Claude API lineup keeps moving: Claude Sonnet 5, Claude Fable 5.1 and Claude Opus 5.5 all shipped between June and September 2026. Choosing now means weighing capability tier, effort tuning, deployment platform, and fast mode. This guide covers all of it, as of Anthropic's documentation in late September 2026.

The Current Claude API Models

The lineup is built around distinct workload tiers, not a simple "good, better, best" ladder. The official models overview lists four current models (prices per million input / output tokens):

• Claude Fable 5.1 (claude-fable-5-1), $10 / $50: built for demanding reasoning and long-horizon agentic work, and Anthropic's most capable widely released model.
• Claude Opus 5.5 (claude-opus-5-5), $4 / $20: built for long-running agentic coding and knowledge work, and the default starting point for most workloads. It also offers fast mode in research preview: up to 2.5x higher output speed at premium pricing.
• Claude Sonnet 5 (claude-sonnet-5), $2 / $10: positioned as the best combination of speed and intelligence.
• Claude Haiku 4.5 (claude-haiku-4-5-20251001), $1 / $5: positioned as the fastest model with near-frontier intelligence, with a 200K-token context window against 1M for the others.

Claude Mythos 5.1 offers the same capabilities as Fable 5.1, but only to Project Glasswing participants. The models overview now lists Claude Fable 5, Opus 5 and Opus 4.8, plus Opus 4.5 to 4.7 and Sonnet 4.5 and 4.6, as legacy models that are still available. And fast mode no longer works on Claude Opus 4.7: per the fast mode docs, requests to claude-opus-4-7 with speed: "fast" return an error, though the model itself still runs at standard speed.

Which Model Do You Actually Need?

Anthropic now gives a direct default in the models overview: "If you're unsure which model to use, start with Claude Opus 5.5 for most workloads. Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short." Its choosing a model guide adds: "Tuning effort is often a better lever than switching models."

So before you jump a tier because your outputs feel shallow, raise the effort level on the model you already run and rerun your evals. The exception is Haiku 4.5, which doesn't support effort.

The guide describes two ways to start, efficiency-first on Haiku 4.5 or capability-first on Opus 5.5. Its selection matrix maps use cases to starting models like this, with an upgrade trigger added to each row:

The word "starting" is doing real work in that table. Build your benchmark tests first and change models only when the evaluation data tells you to. As Anthropic puts it, "having a good evaluation set is the most important step in the process."

Your First Messages API Call

The Messages API is the core integration point for all Claude models. A minimal request, following the pattern in Anthropic's official quickstart:

import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "Summarize this contract in three bullet points."}
    ],
)

for block in message.content:
    if block.type == "text":
        print(block.text)

message.content is a list of blocks, not a string. Opus 5.5 always runs adaptive thinking, so a response can start with thinking blocks; Anthropic advises selecting blocks by type, not position.

A few things to get right from the start:

1. Install the SDK via pip install anthropic and set your ANTHROPIC_API_KEY environment variable.
2. Leave room in max_tokens. It's required, and thinking tokens count toward it.
3. Use system prompts to define behavior, persona, and output format, passed in the top-level system parameter.
4. Know what your model ID pins. From the 4.6 generation on, an ID like claude-opus-5-5 is itself a fixed snapshot. Haiku 4.5's alias claude-haiku-4-5 points to the latest dated snapshot, so pin the dated ID.

For streaming, call client.messages.stream(...) as a context manager and iterate over stream.text_stream. Streaming matters for any user-facing application, because users see text as it arrives instead of waiting for the full response, and the SDKs require it for large max_tokens values.

What All Current Models Support

The shared baseline, per Anthropic's models overview: "All current models support text and image input, text output, multilingual capabilities, vision, and tool use." Past that, support depends on the model and the platform:

• Thinking: adaptive and always on for Fable 5.1 and Opus 5.5, adaptive and on by default for Sonnet 5 (which lets you turn it off), and the older manual-budget extended thinking for Haiku 4.5.
• Effort: Fable 5.1, Opus 5.5 and Sonnet 5; not Haiku 4.5.
• Structured outputs: all four models on the Claude API, but Amazon Bedrock doesn't offer them for Fable 5.1, Opus 5.5 or Sonnet 5.
• Web search: Claude API, Claude Platform on AWS and Microsoft Foundry; basic version only on Google Cloud; not on Bedrock.
• Files API: Claude API, plus beta on Claude Platform on AWS and Microsoft Foundry; not on Bedrock or Google Cloud.

Structured outputs deserve a callout: they constrain Claude's response to a JSON schema, so any pipeline that feeds Claude output into downstream systems gets consistent JSON enforced at the API level instead of prompt-engineered.

Vision on every tier also has a cost angle: test Haiku 4.5 on your screenshot or diagram extraction before routing everything to Fable 5.1.

Can You Build Real Agents with Claude?

Yes. Three of the five areas in Anthropic's features overview cover agent work: tools, tool infrastructure, and context management for long-running sessions.

Claude Code is, in Anthropic's words, "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools." It runs in the terminal, VS Code and JetBrains IDEs, a desktop app, and the browser. If you're evaluating it against other AI coding tools, the Claude Code vs Cursor vs Windsurf comparison covers the practical differences in depth.

Computer use takes it further, giving Claude "screenshot, mouse, and keyboard control of a desktop environment." The computer use tool is generally available on the Claude API and Google Cloud (beta elsewhere), a real option for workflows without APIs, like legacy desktop software. On those two platforms, Opus 5.5 accepts only the newer computer_toolset_20260801 toolset. For tasks that stay inside webpages, Anthropic calls its browser use tool the closer fit.

MCP (Model Context Protocol) is the integration layer that connects Claude to external tools and services. Anthropic's Claude Code docs call it "an open source standard for AI-tool integrations," with examples like turning a Jira issue into a GitHub PR or working from Figma designs posted in Slack. Through the API, the MCP connector (beta) calls tools on remote MCP servers without a separate MCP client. It supports tool calls only, needs servers publicly reachable over HTTP, and isn't available on Bedrock or Google Cloud.

Platform Options: Where Do You Run It?

Claude models are available across multiple deployment surfaces, and the right choice depends on your infrastructure and compliance requirements:

For most developers getting started, the direct Claude API is the right call: every feature is available there (some in beta), and authentication is a single API key.

The cloud options make sense when you want to reuse existing cloud contracts, identity and billing, or when compliance points to one provider (see the Claude Platform on AWS docs for how it differs from Bedrock). Check features first: web search, the Files API and the MCP connector aren't on Bedrock, and Opus 5.5 isn't on Bedrock's list of models open to all customers.

Third-party gateways are a separate route. OpenRouter serves Claude behind an OpenAI-compatible API with its own keys and model IDs (anthropic/claude-opus-5.5, not claude-opus-5-5); my guide on how to use OpenRouter covers that setup.

How Does Effort Tuning Work?

The effort parameter controls how many tokens Claude spends when responding, trading thoroughness against token efficiency within one model. It applies to every output token, including text, tool calls and thinking, and Anthropic calls it a behavioral signal, not a strict token budget. Higher effort buys deeper reasoning on complex tasks at the cost of more tokens and slower responses.

Defaults differ: high on Fable 5.1 and Sonnet 5, medium on Opus 5.5, and no effort support on Haiku 4.5. A request that omits effort on Opus 5.5 runs one level lower than it did on Opus 5, so set it explicitly after any model change.

Anthropic's effort docs suggest matching effort to task complexity: low for speed-sensitive or simple tasks, higher for agentic coding and complex reasoning. One catch: setting a new top-level effort value mid-conversation starts the prompt cache over. On Fable 5.1 and Opus 5.5, a per-message effort change (in beta) keeps the cache.

Effort tuning pairs well with model routing: simple intent classification on Haiku 4.5, complex queries on Opus 5.5 with effort set explicitly. The choosing guide names the two common multi-model patterns, an executor that escalates hard decisions to an advisor and an orchestrator that delegates bulk work to cheaper workers.

Production Pitfalls to Avoid

Swapping the model string and nothing else. Opus 5.5 returns a 400 error for settings earlier models accepted: disabled thinking, manual extended thinking with budget_tokens, forced tool use (tool_choice of any or tool) and, on the Claude API and Google Cloud, the older computer_20251124 tool. Read what's new in Claude Opus 5.5 first.

Ignoring model lifecycles. Anthropic doesn't change the weights behind an existing model ID, but IDs retire. The model deprecations page lists each schedule (Haiku 4.5: not sooner than October 15, 2026), with at least 60 days' notice before a publicly released model retires.

Skipping prompt caching. In Anthropic's internal benchmarks, caching was the largest cost lever, cutting agent-loop cost by a factor of 2.7 to 5.3. It's not enabled by default; add a top-level cache_control field or set breakpoints yourself. Cache reads cost 10% of the base input price (5% on Opus 5.5, 2.5% on Fable 5.1).

Treating every failure the same. The official SDKs already retry connection errors, rate limits and 5xx errors with exponential backoff, twice by default. Refusals are different: Fable 5.1 and Opus 5.5 include safety classifiers that can decline a request with a normal HTTP 200 and stop_reason: "refusal", so check stop_reason before reading the text.

Upgrading models without benchmarks. Switching from Haiku to Opus because outputs "feel better" is expensive and unmeasurable. Run both on your eval set and compare cost per completed task, not per token.

Treating thinking as an on/off switch. Adaptive thinking can't be turned off on Fable 5.1 and Opus 5.5, and thinking tokens bill as output tokens. The control is effort: at lower levels, Claude can skip thinking entirely on simpler problems.

The n8n vs Zapier AI workflow tools comparison makes a similar fit-the-tool-to-the-job argument at the platform level: simple linear automations and multi-step AI pipelines need different tools.

Real Workloads, Real Models

To make this concrete, here's how the model selection maps to actual use cases:

• Customer support triage: Haiku 4.5 for intent classification and simple FAQ responses at low cost, with complex escalations routed to Sonnet 5.
• Document extraction: Haiku 4.5 or Sonnet 5 with structured outputs. PDF support sends each page as an image alongside its extracted text, so charts and tables come through.
• Agentic coding assistant: Opus 5.5 as the base model, with Claude Code as the agent in your terminal or IDE.
• Research summarization: Sonnet 5 with web search enabled, so the model can pull current information rather than relying on training data.
• Desktop automation: Computer use with Opus 5.5, for workflows that require moving through UIs without an API.
• Multi-step enterprise workflows: MCP plus Opus 5.5 or Fable 5.1, depending on the complexity of the reasoning required.

Before you commit a use case to a platform, check the features overview, which lists every feature with the platforms that offer it.

Frequently Asked Questions

The Decision Framework

Picking the right Claude API model comes down to four variables: capability requirement, latency tolerance, cost budget, and effort tuning headroom. The expensive mistake is over-specifying capability and under-investing in evaluation, which means paying for Opus when Sonnet would have worked fine.

The API has moved well beyond a text endpoint, and the right model for a single-turn summarization task is rarely the right model for a multi-step agentic workflow.

Start with your eval set. Choose efficiency-first or capability-first on purpose. Tune effort before switching models. Track the retirement dates of the model IDs you ship. And if Claude will be one step in a larger automation, the n8n vs Make complex marketing workflows comparison covers how each platform handles LLM calls, AI agents and branching logic.

The lineup has already changed since this guide first ran in July 2026, so treat model selection as an ongoing engineering tradeoff, not a one-time decision.

← Back to the blog