Claude 2026 Developer Guide: Opus 5.5, Sonnet 5.5 and the API

Choose Claude Opus 5.5 for demanding development and Sonnet 5.5 for a lower-cost everyday starting point. Evaluate Fable 5.1 when difficult reasoning remains unresolved, and Haiku 4.5 for latency-sensitive, high-volume work. This developer guide connects those choices to API implementation, caching, Claude Code, MCP, and Projects. Anthropic model guidance Sonnet 5.5 specifications

Quick navigation: What is Claude · Models · Pricing · API · Claude Code · Projects · MCP · Patterns · vs ChatGPT · FAQ

What Claude is in 2026

Claude is Anthropic's model family, available through its API and applications. The current general-purpose lineup accepts text and images and produces text. Its models support tool use; model pages specify cloud-platform availability and provider-specific identifiers. Model overview

Choose the surface around the job. Use the API for an application you control, Claude Code for repository work, and Projects for sustained work with shared context. Claude Code runs in terminals, IDEs, desktop, and browsers. Claude Code overview Projects documentation

Models in 2026: IDs, context and output limits

The table covers Anthropic's current general-purpose lineup. Prices are direct API USD per million tokens. M means million; K means thousand. Output limits are for ordinary requests; model names link to specifications.

Haiku's convenience alias is claude-haiku-4-5. It uses manual extended thinking rather than adaptive thinking. Haiku 5.5 is announced as forthcoming; it is not a model to deploy from this lineup yet. Haiku specifications Sonnet announcement

Specialized access and supported legacy models

These models also appear as active in Anthropic's lifecycle documentation. “Legacy” does not mean retired; retain a validated deployment while evaluating migration. Lifecycle reference

Mythos availability is restricted; a published model page does not confer access. Haiku 3.5 is retired on the Claude API. Provider lifecycle schedules can differ, so check the target platform when migrating. Mythos access Deprecations

Pricing: caching and batch processing

Budget separately for uncached input, cache creation, cache reads, and output. The current million-token Claude models include their full context at standard token rates. Fast mode, geographic processing, and tools can have separate charges. API pricing

Use the LLM API pricing calculator to model these published cache rates. Fable and Opus have model-specific read discounts, so a universal “10% of input price” rule is insufficient. Pricing table

Message Batches discounts input and output by 50%. Batches support up to 100,000 requests or 256 MB, whichever limit comes first, and can expire after 24 hours. Use them for asynchronous evaluation or document processing, with handling for individual failures. Batch processing

Selected Opus and Sonnet models, including 5.5, support 300,000 output tokens in Message Batches with output-300k-2026-03-24. That beta limit does not change ordinary requests' 128K ceiling. Fable and Haiku are not on the documented extended-output list. Batch output extension

Anthropic API basics

Use the Messages API with a current model ID. Supply max_tokens, put instructions in the top-level system field, and send user/assistant turns in messages. The Python SDK reads ANTHROPIC_API_KEY from the environment. Messages reference Python SDK

import anthropic

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=4096,
    system="Review the supplied code. Distinguish bugs from suggestions.",
    messages=[{
        "role": "user",
        "content": "Review this function: def double(x): return x + x",
    }],
)
for block in response.content:
    if block.type == "text":
        print(block.text)

Inspect the response's stop reason and usage before treating it as complete. SDK streaming uses messages.stream(); responses can contain content blocks beyond plain text. API reference

With client tools, Claude returns a tool_use block; your application executes the permitted operation and returns tool_result. Keep tool execution and its authorization in your application. Tool-use documentation

Integration changes that deserve explicit tests

Opus 5.5 has always-on adaptive thinking, rejects forced tool use, and binds thinking blocks to their model and conversation. Progress text between tool calls can arrive in thinking blocks whose default display exposes no text. Opus integration notes

Sonnet 5.5 offers between_tools to disable up-front thinking at high effort or below. It also rejects forced tool use and non-default sampling parameters such as temperature. Check its migration notes before carrying request options between models. Sonnet integration notes

I would test streaming progress, tool dispatch, refusal handling, and truncated output before changing a production model. Preserve a small set of real failures as regression fixtures; a successful greeting proves little about an agent loop.

Claude Code: terminal and IDE workflows

Claude Code reads files, edits code, and runs commands. Anthropic recommends its native installer; this example is for macOS, Linux, or WSL. Installation documentation

curl -fsSL https://claude.ai/install.sh | bash
claude --version
claude --model claude-sonnet-5-5

Run the final command from your project directory. Sonnet 5.5 requires 2.1.284+, Opus 5.5 2.1.280+, and Fable 5.1 2.1.255+. Full model names help avoid provider-dependent alias resolution. Model configuration Fable requirements

Both 5.5 models default to medium effort in Claude Code; the Sonnet API defaults to high. Record the actual setting when comparing runs. Code effort settings Sonnet API specifications

Fable on Pro and standard Team seats uses usage credits. Max and eligible premium seats can spend up to 50% of their existing weekly allowance on Fable. Guaranteed weekly message counts are not published in that plan guide. Plan rules

For workflow comparisons, consult the AI coding assistants guide.

Claude Projects: knowledge, memory and availability

Projects provide workspaces with files, instructions, and chats. Free accounts can create up to five projects; paid plans support expanded knowledge through retrieval. Team and Enterprise projects support organizational sharing subject to permissions. Projects overview

Put shared facts in project knowledge and behavioral rules in project instructions. Anthropic says chat context is not shared across project chats unless added to knowledge; its memory features have separate plan and administrator settings. Do not assume every transcript becomes shared reference material. Project management

A Projects beta is rolling out to selected Pro and Max Claude Code users. It organizes work into cloud threads with shared project files, repositories, instructions, and memory. Existing chat projects remain available, so check which experience the account actually has. Beta availability

MCP: connecting Claude to tools and data

Model Context Protocol connects Claude Code to tool and data servers. Configure the server in the client where it will run. Claude Code documents remote HTTP connections, local stdio processes, authentication, and project-scoped configuration. MCP documentation

For a remote server, replace the placeholder with its documented endpoint:

claude mcp add --transport http example https://mcp.example.com/mcp
claude mcp list

Use /mcp inside Claude Code to inspect connection status and authenticate where required. Project-scoped configurations use .mcp.json; choose that scope when the configuration should be shared with a repository. MCP setup

For Cursor, see the Higgsfield MCP setup guide, the Claude × Higgsfield integration, and Meta MCP integrations.

Practical patterns for Claude applications

Cache a stable prefix

Automatic caching uses a top-level cache_control={"type": "ephemeral"}; explicit breakpoints attach that field to content blocks. The default lifetime is five minutes, with an optional one-hour duration. Cache matching covers the prompt prefix, including tools, system content, and messages in that order. Caching guide

Fable 5.1, Opus 5.5, and Sonnet 5.5 require at least 512 cacheable tokens; Haiku 4.5 requires 4,096. Inspect cache_creation_input_tokens and cache_read_input_tokens to confirm caching actually occurred. Cache requirements

Use schemas for machine-readable output

For structured data, Anthropic documents output_config.format with a JSON schema, plus strict tool schemas. Use these features when downstream code needs a predictable shape. Structured outputs

For prose, clear sections and examples may suffice. The prompt engineering guide provides patterns to adapt.

Evaluate critique and routing as separate steps

Try a second review pass for difficult work, but judge it against known defects and reviewer effort. Similarly, test a cheap classification step before adding a larger model to the workflow. Treat both as experiments rather than guaranteed accuracy improvements.

For architecture examples, see the agent patterns guide.

Preserve durable state outside the conversation

Anthropic documents compaction for conversations that approach their context limit. My recommendation is to keep requirements, decisions, and external operation results in application state as well, so the workflow can resume from a verifiable record. Compaction documentation

Claude vs ChatGPT and Gemini for developers

Compare models and clients separately. GPT-6 Astra is OpenAI's flagship; Sol and Luna cover lower-cost workflows. Gemini 3.8 Flash is Google's model for long-horizon software engineering, with broader documented input modalities. OpenAI catalog Gemini specifications

Use the coding model comparison for the model shortlist and the coding assistants comparison for client workflow choices. Follow model releases when reviewing deployment options.

FAQ

Which Claude model should I use for coding?

Evaluate Sonnet 5.5 for routine work and Opus 5.5 for difficult tasks. Escalate to Fable only when its additional cost resolves meaningful failures. Haiku is worth testing for focused, inexpensive steps.

What is Claude's context window in 2026?

Fable 5.1, Opus 5.5, and Sonnet 5.5 have 1M-token context windows and ordinary 128K output limits. Haiku 4.5 has 200K context and 64K output. Context documentation Haiku specifications

Does prompt caching store the model's answer?

Prompt caching reuses input prefixes. It does not turn a repeated task into a free cached answer; generated output still has its own billing. Caching documentation

Is Claude Code limited to the terminal?

No. Anthropic supports terminal, IDE, desktop, and browser surfaces. Choose the one that fits your review workflow and available account access. Claude Code overview

Should I use Projects or MCP?

Use Projects for shared work context and instructions. Use MCP when Claude needs a connection to external tools or data. They solve different parts of a workflow and can be complementary. Projects MCP

Sources

• Claude model overview
• Opus 5.5 specifications
• Sonnet 5.5 specifications
• Fable 5.1 specifications
• Haiku 4.5 specifications
• Sonnet 5.5 announcement
• Mythos 5.1 specifications and access
• Mythos 5 specifications and access
• Fable 5 specifications
• Opus 5 specifications
• Opus 4.8 specifications
• Opus 4.7 specifications
• Opus 4.6 specifications
• Opus 4.5 specifications
• Sonnet 5 specifications
• Sonnet 4.6 specifications
• Sonnet 4.5 specifications
• Model lifecycle and deprecations
• Claude API pricing
• Message Batches
• Messages API reference
• Official Python SDK README
• Tool-use documentation
• Claude Code installation and surfaces
• Claude Code model configuration
• Fable plan requirements
• Projects and beta availability
• Project management and knowledge
• Claude Code MCP connections
• Prompt caching
• Structured outputs
• Compaction
• Context windows
• OpenAI model catalog
• Gemini 3.8 Flash specifications

Last reviewed and updated: September 2026.