Claude Models Explained: Opus, Sonnet, Haiku, and the New Mythos Tier (2026 Guide)

If you have opened Claude.ai or the Claude API in the last few months, you have run into a small but consequential decision: which model do you actually want answering you? That choice between the built-in models Anthropic ships under the Claude name is the difference between a fast, cheap reply and a slow, expensive, far more capable one. This guide explains every Claude-built-in model available in 2026, how the lineup is organized, what each model is good at, and how to pick the right one without overpaying or under-powering your work.

Whether you are a Claude.ai subscriber tapping the model dropdown or a developer hard-coding a model ID into a production app, the same family of models sits underneath. Understanding that family is the single highest-leverage thing you can learn about the platform.

What “Claude models” actually means

Claude is not one model. It is a family of large language models built by Anthropic, exposed through a consistent set of interfaces. When people say “Claude’s built-in models,” they usually mean one of two things, and in practice, both point at the same underlying lineup:

• In the Claude apps (claude.ai, the desktop app, the mobile app), the built-in models are the options you select from the model picker, the ones the product offers you out of the box.
• On the Claude API and partner platforms, the built-in models are the published model IDs you call programmatically, such as claude-opus-4-8 or claude-sonnet-4-6.

The important thing to internalize early: these are the same models. The model that powers a thoughtful answer in the chat window is the same model a developer invokes by ID in code. The branding, the capabilities, and the price tiers are unified across surfaces. Once you understand the family, you understand both the consumer product and the developer platform at the same time.

The three-tier architecture: Opus, Sonnet, Haiku

Since the Claude 3 generation launched in March 2024, Anthropic has organized its models into three named size tiers. The naming is deliberately poetic; all three are forms of writing, but the hierarchy is strictly about the trade-off between capability, speed, and cost:

• Opus is the most capable tier. It is built for the hardest work: complex multi-step reasoning, long-horizon agentic coding, architecture decisions, and high-autonomy tasks where getting the right answer on the first try matters more than how long it takes or what it costs.
• Sonnet is the balanced tier. It pairs strong reasoning and coding ability with much faster output and a lower price, which makes it the default workhorse for most everyday and production use.
• Haiku is the speed-and-cost tier. It is the fastest and cheapest of the three, tuned for high-volume, latency-sensitive jobs like classification, routing, extraction, quick edits, and large batch processing, while still retaining near-frontier intelligence.

A rule of thumb that holds up well: start with Sonnet, reach for Opus when the task is genuinely hard, and drop to Haiku when speed and volume matter more than peak nuance. Most users overestimate how often they need the top tier.

Two naming details trip people up. First, the official format is “Claude Haiku 4.5,” not “Claude 4.5 Haiku”: the tier name comes after the word Claude and before the version number. Second, the version numbers (4.5, 4.6, 4.8) advance independently across tiers, so the newest Opus and the newest Sonnet do not always carry the same number.

In 2026, the family gained a fourth, higher-tier sitting above Opus, which is the Mythos-class tier. It is covered in its own section below because its availability situation is unusual.

The current model lineup (mid-2026)

Here is the lineup of generally available built-in Claude models as of mid-2026. These are the models you will actually find in the app picker and on the API today.

Claude Opus 4.8 — the flagship

Claude Opus 4.8 is Anthropic’s most capable Opus-tier model and the recommended starting point for the most demanding work. It is purpose-built for complex reasoning, long-horizon agentic coding, and high-autonomy workflows where the model needs to plan, execute, and verify across many steps.

• API model ID: claude-opus-4-8
• Context window: 1 million tokens
• Maximum output: 128K tokens
• Pricing: $5 per million input tokens, $25 per million output tokens
• Thinking: adaptive thinking, with the effort parameter defaulting to high on all surfaces

Opus 4.8 is the choice when a single failed task is expensive — large cross-file refactors, system architecture, scientific reasoning, or agent runs that have to stay reliable over long sessions. Anthropic has emphasized reliability gains in this release, including a substantial reduction in the model letting flaws in its own generated code slip past unremarked, compared with the previous Opus.

Claude Sonnet 4.6 — the balanced workhorse

Claude Sonnet 4.6 offers the best combination of speed and intelligence in the lineup, and it is the model most teams should reach for first. It delivers the large majority of Opus-level quality on everyday coding, writing, and analysis tasks at a meaningfully lower price and faster output, which is exactly why it tends to be the default in production applications and in tools like Claude Code.

• API model ID: claude-sonnet-4-6
• Context window: 1 million tokens
• Pricing: $3 per million input tokens, $15 per million output tokens
• Thinking: adaptive thinking

Sonnet 4.6 was a notable release because it was the point at which a Sonnet-tier model began winning coding preference comparisons against the previous generation’s Opus, a reminder that “bigger” stopped automatically meaning “better” some time ago. For most development, writing, document analysis, and enterprise API scenarios, Sonnet 4.6 is the safest place to begin.

Claude Haiku 4.5 — the fast, cheap tier

Claude Haiku 4.5 is the fastest model in the family, with near-frontier intelligence at the lowest price point. It is built for high-throughput, latency-sensitive workloads where cost per call matters and the task does not demand the deepest reasoning.

• API model ID: claude-haiku-4-5-20251001 (alias claude-haiku-4-5)
• Context window: 200K tokens
• Pricing: $1 per million input tokens, $5 per million output tokens
• Thinking: adaptive thinking

Haiku is not a “weak” model; it is positioned as fast and inexpensive while keeping capabilities close to the frontier. It shines in real-time chat and support bots, large-scale short-text classification, simple code edits, rapid prototypes, and as the cheap worker tier in multi-agent architectures where small subagents handle routing and the heavy model is reserved for hard subtasks. Notably, the apps can route simple requests to Haiku automatically through smart model switching, so casual users often benefit from it without picking it explicitly.

The Mythos-class tier: Fable 5 and Mythos 5

In June 2026, Anthropic introduced a tier that sits above Opus, called the Mythos class. It comprises two models that share the same underlying capability but differ in their safety configuration:

• Claude Fable 5 (claude-fable-5) — described as Anthropic’s most capable, widely released model, built for the most demanding reasoning and long-horizon agentic work. It runs with always-on adaptive thinking, a 1-million-token context window, and 128K output tokens. It’s priced at $10 per million input tokens and $50 per million output tokens, which is double that of Opus 4.8. Fable 5 includes a safety classifier system that automatically routes flagged high-risk requests (in domains such as cybersecurity, biology, and chemistry) to Opus 4.8 instead, reportedly triggering in under 5% of sessions.
• Claude Mythos 5 (claude-mythos-5) — the same model without those safety classifiers, offered only in limited availability to approved customers through Anthropic’s Project Glasswing program.

An important and time-sensitive caveat: on June 12, 2026, just days after launch, Anthropic announced it had received a US government export-control directive requiring it to suspend access to both Fable 5 and Mythos 5. As a result, neither Mythos-class model is currently usable by general customers. Access to every other Claude model, like Opus, Sonnet, and Haiku, is unaffected and remains fully available.

There is also a separate, invitation-only Claude Mythos Preview model offered as a research preview for defensive cybersecurity workflows within Project Glasswing, with no self-serve sign-up.

The practical takeaway: if you are choosing a built-in model to build on today, plan around Opus 4.8 as your top-tier option. Treat the Mythos class as a frontier capability whose availability is in flux, and check Anthropic’s official announcements before designing anything around it.

Capabilities shared across Claude’s models

One of the strengths of the Claude family is consistency. The current built-in models share a common feature set, so you can move work between tiers without re-architecting. Key shared capabilities include:

Multimodal input with vision. All current Claude models accept both text and image input and produce text output, with genuine vision understanding. Recent generations sharpened image handling considerably — the Opus 4.7 release tripled image resolution, making vision reliable enough for professional document and screenshot work.

Large context windows. Opus 4.8, Sonnet 4.6, and Fable 5 all support a 1-million-token context window at standard per-token pricing, while Haiku 4.5 carries a 200K window. A million tokens is enough to load entire codebases, long contract sets, or book-length documents into a single request — a major advantage for long-document analysis.

High output capacity. The flagship models support large maximum outputs (128K tokens on Opus 4.8 and Fable 5), and the Message Batches API can push selected models up to 300K output tokens with a beta header. Large single-pass output matters for long-form generation — technical documentation, reports, and long code files — where chopping output into multiple calls risks consistency errors.

Adaptive thinking. Earlier Claude models exposed an “extended thinking” mode you toggled manually. Starting with the 4.6 generation, extended thinking was deprecated and then removed (it is gone in Opus 4.7 and later) in favor of adaptive thinking, where the model decides how much to reason internally based on the difficulty of the request. On Opus 4.8, an effort parameter (defaulting to high) lets you tune how much reasoning the model invests.

Tool use and agentic features. The built-in models support tool use (function calling), system prompts, structured outputs, the Files API, PDF support, and the full set of server-side and client-side tools that power agentic applications and computer use.

Prompt caching and batch processing. For cost control, prompt caching can cut the cost of repeated context dramatically, and the Message Batches API offers a 50% discount on asynchronous workloads. Combining the two is the standard way high-volume teams keep their bills manageable.

How to choose the right Claude model

The best model is the cheapest one that reliably does the job. Here is a task-oriented decision guide:

Choose Claude Opus 4.8 when:

• You are doing large code changes across many files or complex system refactoring.
• The task involves long-chain agent planning where reliability over many steps matters.
• A wrong answer is costly — high-stakes analysis, architecture, or scientific reasoning.
• You want the model to spend more time understanding context before acting.

Choose Claude Sonnet 4.6 when:

• You are doing everyday coding, writing, knowledge work, or long-document analysis.
• You are building a production API application and want a strong default.
• You need a balance of quality, speed, and cost — which is most of the time.

Choose Claude Haiku 4.5 when:

• You are running high-volume, latency-sensitive work: classification, routing, extraction, support bots.
• You need cheap subagents inside a multi-agent system.
• The task is simple and well-defined, and throughput matters more than nuance.

A practical workflow many developers adopt is tiered model selection: prototype and route with Haiku, run the bulk of work on Sonnet, and escalate to Opus only for the subtasks that clearly justify the premium. You do not need to chase the strongest model from the start — you need to match the model to the cost of being wrong.

Models in Claude.ai vs. the API

Although the models are the same, how you access them differs between the consumer app and the developer platform — and a couple of details are worth knowing.

In the Claude apps

In claude.ai and the desktop and mobile apps, the built-in models appear in a model picker. Subscribers on paid plans can select among the available tiers, while the free tier provides access to Sonnet with daily usage limits. The apps also feature smart model switching, which can automatically route simpler requests to a cheaper, faster model so that casual interactions stay snappy and economical without the user thinking about it.

On the API: model IDs and versioning

On the API, you specify a model by its ID. Two versioning rules matter:

• Every Claude model ID is a pinned snapshot. Calling a given ID gets you that exact, fixed release — it will not silently change underneath you.
• Starting with the 4.6 generation, model IDs use a dateless format (for example, claude-opus-4-8 and claude-sonnet-4-6) that is still a pinned snapshot, not an evergreen pointer to “the latest.” Older models used a dated format like claude-haiku-4-5-20251001, sometimes with a convenience alias that resolves to the dated ID.

This pinning matters for production stability: you decide when to upgrade, and migrations between versions are documented so you can check for breaking changes (such as changes to sampling parameters or the removal of manual extended thinking in newer models).

Where Claude’s models run

The same built-in models are available across multiple deployment surfaces, which is a meaningful advantage for teams with existing cloud commitments. Current Claude models can be accessed through:

• The Claude API (Anthropic’s first-party platform)
• Claude Platform on AWS, which uses the same model IDs as the Claude API and follows Anthropic’s own model lifecycle rather than Bedrock’s
• Amazon Bedrock
• Google Cloud’s Vertex AI
• Microsoft Foundry (note that on Foundry, Opus 4.8 carries a 200K context window rather than 1M)

On Bedrock and Vertex, newer models offer both global endpoints (dynamic routing for maximum availability) and regional endpoints (guaranteed data routing through specific geographies), which helps with data-residency requirements. Because availability, context limits, and regional support can vary by platform, it is always worth confirming the specifics on the official documentation for your chosen surface before you build.

How Claude’s models have evolved

A brief look at the trajectory makes the current lineup easier to understand. The three-tier Opus/Sonnet/Haiku structure arrived with Claude 3 in March 2024, alongside vision. Claude 3.5 Sonnet (mid-2024) showed a mid-tier model could outperform the prior flagship, and its updated version introduced computer use — direct control of a computer interface. Claude 3.7 Sonnet (early 2025) brought extended, step-by-step thinking. The Claude 4 generation made professional-grade coding a daily reality and powered Claude Code as a serious development tool.

Through 2026, the cadence accelerated: Sonnet 4.6 closed much of the gap with the previous Opus; Opus 4.7 sharpened vision and added self-verification; Opus 4.8 emphasized reliability and a faster mode; and the Mythos-class Fable 5 and Mythos 5 pushed the frontier — before the export-control suspension paused public access to that top tier. The throughline is consistent: each generation pushes capability up while pushing the price of a given capability level down.

Frequently asked questions about Claude’s models

Conclusion: pick the model that fits the job

Claude’s models give you a clean, three-tier menu — Opus for the hardest work, Sonnet for the everyday default, and Haiku for fast, high-volume tasks — plus a frontier Mythos tier whose availability is currently restricted. The smartest approach is rarely “always use the biggest model.” It is to match the model to the cost of being wrong: default to Sonnet, escalate to Opus when reliability and complexity demand it, and lean on Haiku when speed and volume dominate.

Because the lineup, pricing, and availability change quickly — as the 2026 Mythos suspension makes clear — treat any single guide as a snapshot and confirm current details on Anthropic’s official model documentation before you commit a production decision to a specific model.

Written by Dhanushri Devi Kannan, Senior at University of Illinois Chicago