What Is Claude AI? Anthropic's Models and Pricing Explained

What Is Claude AI?

Claude AI is Anthropic's family of large language models, designed to work as a conversational assistant, a coding partner, and an API-level reasoning engine for other software. Anthropic was founded in 2021 by Dario Amodei and Daniela Amodei, both former OpenAI researchers, specifically to build AI systems that are "helpful, honest, and harmless."

That third word, harmless, is the part that shapes how Claude is built. Anthropic trains Claude using an approach it calls Constitutional AI: instead of relying only on human thumbs-up/thumbs-down feedback, the model is trained against an explicit written "constitution" of principles and values. That constitution has grown enormously. In January 2026, Anthropic published a new version running to roughly 23,000 words, up from about 2,700 words in 2023, shifting toward reason-based explanations of ethical principles rather than blanket rules (Anthropic, "Claude's new constitution").

Claude is sold in two main forms. The consumer assistant runs as web, desktop, and mobile apps at claude.ai, with Free, Pro, Max, Team, and Enterprise plans. The developer platform is the Claude API, available directly from Anthropic and through Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, billed per million tokens.

By mid-2026, Anthropic describes Claude as moving from a "chat tool" toward a "platform layer": something that plugs into a company's existing infrastructure (self-hosted sandboxes, Microsoft 365, legal and workflow connectors) rather than a window you type questions into.

The short version

The Claude Model Lineup in 2026

As of June 2026, Anthropic runs four active model tiers, and the gap between the cheapest and most expensive is larger than most people expect: a 10x difference in per-token pricing between Claude Haiku 4.5 and Claude Fable 5.

Pricing is from Anthropic's own model pages and launch coverage (Anthropic, 2026). Prompt caching cuts input costs by up to 90% and batch processing cuts costs by up to 50% on Opus-tier models, according to Anthropic's pricing pages.

Claude Haiku 4.5: the fast, cheap tier

Released October 15, 2025, Haiku 4.5 is Anthropic's small model, and "small" is relative. It scored 73.3% on SWE-bench Verified, which Anthropic and partners describe as making it "one of the world's best coding models" at that price point, roughly matching Claude Sonnet 4 on coding tasks at one-third the cost.

Haiku 4.5 is the model behind most "instant" Claude features: quick chat replies, high-volume API calls, and any workflow where you're running thousands of requests and the per-token cost adds up fast.

Claude Sonnet 4.6: the everyday default

Sonnet 4.6 launched February 17, 2026 with a 1 million token context window in beta, and Anthropic's own messaging leans hard on one idea: "Tasks that previously required an Opus-class model... are now achievable [with] Sonnet" (CNBC, citing Anthropic, Feb 17, 2026).

Sonnet 4.6 is the model most free-tier and Pro users interact with by default. It handles agentic coding, code review, frontend UI generation, and long-document analysis without the per-token cost of Opus or Fable.

Claude Opus 4.8: the pre-Fable flagship

Before Fable 5 shipped, Opus 4.8 was Anthropic's most capable public model, and it remains the practical choice for teams that don't need Fable 5's multi-day autonomy. Anthropic's alignment team reported that Opus 4.8 is "around four times less likely" than its predecessor (Opus 4.7) to let coding flaws slip through unremarked, with overall misalignment rates "substantially lower" than Opus 4.7 and close to the unreleased Claude Mythos Preview (VentureBeat, 2026).

Opus 4.8 introduced "dynamic workflows for parallel agentic coding," letting it spawn multiple subagents to work across a codebase at once, plus an "effort control" setting and a Fast Mode that cuts latency at $10/$50 per million tokens.

Claude Fable 5 (and Mythos 5): the new flagship

Claude Fable 5 is the headline release of June 2026. Anthropic's announcement describes it as a "Mythos-class model that we've made safe for general use" (Anthropic, "Claude Fable 5 and Claude Mythos 5"). Fable 5 and Mythos 5 share the same underlying architecture and capabilities, but Fable 5 wraps that architecture in additional safety classifiers and "dynamic routing": when a request touches high-risk domains like cybersecurity, advanced biology, or chemistry, Fable 5 can automatically hand the request to a more constrained model (typically Opus 4.8), reportedly affecting under 5% of sessions.

Mythos 5 itself, the unrestricted version, is not generally available. It's limited to approved US government and critical-infrastructure partners under what Anthropic calls "Project Glasswing." For everyone else, Fable 5 is "the same performance as Claude Mythos 5, except with much more strict guardrails in place to prevent it being used for harmful things," per Anthropic.

Fable 5 is available as `claude-fable-5` on the Claude API and claude.ai, inside Claude Code and Claude Cowork, and through Amazon Bedrock, Google Vertex AI, Microsoft Foundry, and GitHub Copilot. For a deeper look at how developers are actually using Claude's coding tools day to day, see our Claude Code Reddit roundup, and for ready-made automation add-ons, see our guide to Claude Code skills.

How Much Does Claude AI Cost?

Claude pricing splits into two systems: subscription plans for the chat app, and per-token API pricing for developers. The two are easy to mix up, and the gap between Anthropic's cheapest and most expensive models is the part most pricing pages bury.

Subscription plans (claude.ai)

Through June 22, 2026, Claude Fable 5 is included at no extra cost across Pro, Max, Team, and seat-based Enterprise plans. From June 23, 2026, Anthropic is shifting Fable 5 access on those plans to metered usage credits, with an intent to restore it as a standard included feature once capacity allows, according to Anthropic's own documentation.

API pricing per million tokens

The Number Most Guides Don't Show

Look at that table again: Fable 5 costs 10 times more per token than Haiku 4.5, on both input and output. Run the same moderate workload, say 10 million input tokens and 2 million output tokens in a week of heavy agentic coding, through each model and the gap becomes concrete: that's $100 + $100 = $200 on Fable 5, versus $10 + $10 = $20 on Haiku 4.5. Same task, same token counts, a $180 difference based purely on which model name you typed into the API call.

That's not an argument for always choosing the cheapest model. SWE-bench Pro shows an 11-point capability gap between Fable 5 (80.3%) and Opus 4.8 (69.2%), and a much larger one against Haiku. But it does mean the default choice of "just use the most capable model for everything" gets expensive fast at scale, and most production workloads are better served by routing routine tasks to Haiku or Sonnet and reserving Fable 5 for the genuinely hard, multi-day agentic jobs it was built for.

For a broader look at how token-based pricing works across the industry, our explainer on what an LLM is covers the training and inference cost basics that drive these numbers.

How Claude Compares to ChatGPT and Gemini

On raw capability, the clearest data point in mid-2026 is SWE-bench Pro: Claude Fable 5 scored 80.3%, compared with roughly 58.6% for GPT-5.5, a 22-point gap on a benchmark specifically designed to measure real-world software engineering ability. That single number is why Fable 5's launch coverage focused so heavily on coding and agentic tasks rather than general chat quality.

Outside of benchmarks, the practical differences come down to positioning:

Reddit's developer community has been tracking this shift closely. The consensus in our Claude vs ChatGPT Reddit roundup is that Claude leads on coding quality and writing style, while ChatGPT remains the more versatile all-rounder with a larger plugin and integration ecosystem. For a side-by-side feature and pricing breakdown across all four major assistants, see our AI tools comparison guide.

The honest answer for most people is that the "best" assistant depends on the task: Claude's advantage shows up most clearly in long documents, large codebases, and agentic workflows that run for hours or days, which is exactly the niche Fable 5 was built to extend.

What Can You Do With Claude AI?

Claude's capabilities fall into four broad categories, and which Claude model you should reach for depends heavily on which of these you're doing.

Coding and software engineering

This is where Claude has built its reputation in 2026. Claude Code (the CLI and web-based coding agent) runs on Sonnet 4.6 by default and Fable 5 or Opus 4.8 for harder tasks. Opus 4.8's "dynamic workflows for parallel agentic coding" let it spawn multiple subagents to work across a large codebase simultaneously. Our Claude Code Reddit roundup tracks what developers are actually building with it, from multi-agent pipelines to full production deployments.

Long document and research work

With a 1 million token context window on Sonnet 4.6 and Fable 5, Claude can ingest entire codebases, legal contracts, or research papers in a single request without chunking. This is the use case Anthropic's original "200K context" Claude built its early reputation on, now extended 5x.

Computer use and agents

Sonnet 4.6 scored 72.5% on OSWorld, a benchmark for controlling a desktop environment (clicking, typing, navigating applications) the way a human would. Combined with Fable 5's multi-day autonomous operation, this is the foundation of Anthropic's "platform layer" pitch: Claude as a digital employee that can run a project over days, not just answer one prompt.

Extending Claude with skills

Claude Code supports "skills," reusable instruction files that add a specific capability behind a single slash command, so you don't retype a long prompt every time. Our Claude Code skills guide covers the ones that hold up in real workflows after testing 40+ of them.

Where the safety layer shows up

Anthropic's Constitutional AI approach is most visible in Fable 5's dynamic routing: requests touching cybersecurity, advanced biology, or chemistry can be automatically redirected to a more constrained model. For everyday coding, writing, and research tasks, this layer is invisible. It mainly matters if you're working in a regulated or dual-use-sensitive domain.

Common Misconceptions About Claude AI

"Claude is just a ChatGPT clone"

Claude and ChatGPT are both large language models, but they're built on different training philosophies. Claude's Constitutional AI approach trains the model against an explicit, published set of principles (now around 23,000 words) rather than relying primarily on human preference ratings. The practical effect shows up in tone, refusal behavior, and how each model handles long, structured tasks.

"Fable 5 and Mythos 5 are the same thing you can use"

They share the same core architecture, but only Fable 5 is publicly available. Mythos 5 itself is restricted to approved US government and critical-infrastructure partners under Anthropic's "Project Glasswing." When people refer to "Claude Mythos," they're almost always describing what Fable 5 can do, since that's the version anyone outside that program can actually access.

"All Claude models cost about the same"

The spread is large: $1/$5 per million tokens for Haiku 4.5 versus $10/$50 for Fable 5, a 10x difference on both input and output pricing. Picking the wrong model for a high-volume task can multiply your API bill without a proportional gain in output quality.

"Claude can browse the web and use any app by default"

Computer use and tool access depend on the surface you're using and what's been connected. Claude's web and desktop apps, the API, and Claude Code each expose different levels of tool access, and OSWorld's 72.5% score for Sonnet 4.6 reflects performance under test conditions with tools enabled, not a default always-on capability for every Claude interaction.

Where Claude AI Is Headed

Two threads from 2026 point at where Claude goes next.

The first is the platform shift. Anthropic has been explicit that it sees Claude moving from a chat window to infrastructure: self-hosted sandboxes, Microsoft 365 integrations, and workflow connectors that let Claude operate inside a company's existing systems rather than alongside them. Fable 5's multi-day autonomous operation is the technical piece that makes this possible, since a model that needs constant prompting can't run a project unsupervised over days.

The second is the safety-capability split that Fable 5 and Mythos 5 represent. Anthropic now ships two versions of its frontier architecture: a public one wrapped in safety classifiers and dynamic routing, and a restricted one for vetted government and infrastructure partners. If that pattern holds, future "Mythos-class" releases will likely follow the same template: a Fable-style public release with guardrails arriving alongside, or shortly after, a more capable restricted version.

For now, the practical takeaway is that Claude's lineup spans a genuinely wide range, from a $1/$5 model that handles routine tasks at scale to a $10/$50 model posting the best public SWE-bench Pro score on record. Most users and teams will get the most value picking the model that matches the task rather than defaulting to whichever one launched most recently.

Frequently Asked Questions

Related Articles