MODEL DOSSIER — CLAUDE (Anthropic)

What it is

Claude is Anthropic's family of large language models — the product of a lab founded in 2021 by former OpenAI researchers who left over disagreements about how fast to scale and how carefully. That origin story is not trivia; it shapes the product. Claude is built around a single wager: that the most valuable AI is one you can hand a long, messy, multi-step task and trust to work on it unsupervised. Everything in the lineup — the million-token context window, the agentic tooling, the "constitutional AI" training approach, Claude Code — serves that wager.

The practical result: Claude is the model professionals reach for when the work is code, documents, analysis, or any task where the model has to hold a lot of context and act on it across many steps. It is not the broadest consumer product on the market. It is the deepest working tool.

---

The family

Anthropic ships models in tiers, named by capability. As of mid-2026 the active lineup: ⚑ unverified

Older versions (Opus 4.7, 4.6, Sonnet 4.6) remain served for teams pinned to them, but the four above are the ones to know.

How the lineup got here

The generational arc matters because it explains what Anthropic optimizes for. Claude 1 and 2 (2023) were cautious, text-only, and commercially marginal. Claude 3 (March 2024) introduced the three-tier Opus/Sonnet/Haiku structure and vision. The inflection point was Claude 3.5 Sonnet (mid-2024), which beat the then-flagship Opus at a third of the price and made Claude the developer community's coding default — a position Anthropic has defended through every release since. The Claude 4 series (2025) added extended thinking, computer use, and the long agentic runs; the 4.5–4.8 cadence through late 2025 and 2026 iterated fast on autonomy, memory, and instruction-following. The Claude 5 generation (2026) — Fable 5 at the top, Sonnet 5 in volume — is the first built around always-available adaptive reasoning rather than reasoning as a mode you switch on. ⚑ unverified

The pattern across all of it: Anthropic ships fewer, denser releases than OpenAI, keeps the tier structure stable, and concentrates gains on coding, agentic reliability, and long-context work rather than consumer surface area.

Three things about this table that matter in practice:

The context window is a million tokens across the working tiers. That is roughly 750,000 words — an entire codebase, a legal case file, a year of correspondence — held in a single request, at standard pricing with no long-context surcharge. ⚑ unverified This is the single most under-appreciated spec in the lineup. Most workflows that people solve with retrieval pipelines and vector databases on other platforms fit directly into Claude's context.

Thinking is now adaptive, not budgeted. Current Claude models decide for themselves when and how deeply to reason before answering, tuned by an effort parameter (low through max) rather than a fixed token budget. On Fable 5, thinking is always on and cannot be disabled. The upshot for buyers: you pay for reasoning depth roughly in proportion to task difficulty, and you control the dial.

Fable 5 is a different animal. It runs safety classifiers that can refuse requests in biosecurity and most offensive-cybersecurity territory, requires 30-day data retention (no zero-retention option), and single requests on hard tasks can run many minutes. ⚑ unverified It is priced and built for the hardest work, not as a default. For most teams, Opus 4.8 or Sonnet 5 is the right ceiling.

---

What distinguishes Claude

1. Long-horizon agentic work

The current generation's defining capability is not answering questions — it is executing. Opus 4.8 and Fable 5 are trained and measured on multi-hour autonomous runs: complex refactors, overnight coding sessions, end-to-end deliverables (analysis, spreadsheets, documents) that complete without a human correcting course. The models delegate to parallel sub-agents, maintain file-based memory across sessions, and self-verify against specifications. Competing models can do individual steps as well or better; Claude's edge is coherence across hundreds of steps.

2. The constitutional approach

Anthropic trains Claude against an explicit written constitution — a document of principles the model is optimized to follow — rather than relying purely on human raters labeling good and bad outputs. Whether this produces "safer" AI is a live debate. What it observably produces is a model with a consistent character: fewer sycophantic reversals under pressure, more willingness to push back on a bad premise, and refusals that are (usually) reasoned rather than reflexive. For professional use, the practical benefit is predictability. The practical cost is that Claude occasionally declines legitimate edge-case work, particularly in security research — a real limitation if that is your field.

3. Claude Code

Claude Code started in 2025 as a terminal tool and has become Anthropic's second product line in all but name. It is an agentic coding environment: you describe the outcome, and the agent reads your codebase, plans, edits files, runs tests, and iterates until the work passes.

It now ships on four surfaces: ⚑ unverified

• CLI — the original. Runs in any terminal, scriptable, the power-user surface.
• Desktop app — redesigned in 2026 around parallel sessions: multiple agents working simultaneously on git-isolated copies of your repo, drag-and-drop pane layout, an integrated terminal and editor, live in-app preview of the running application (the agent watches console errors and fixes them), inline code review before you push, and CI/PR monitoring after.
• Web and mobile — cloud sessions at claude.ai/code, monitorable from the iOS app; a local desktop session can be handed off to the cloud mid-task and picked up from a phone.
• IDE integrations — VS Code and JetBrains extensions that embed the agent in an existing editor workflow.

Why it matters beyond the demo appeal: Claude Code is the reference implementation of the agentic pattern the whole industry is converging on — an AI that owns a task rather than autocompleting a line. It is also the anchor of Anthropic's subscription economics: a Pro or Max plan bundles what would otherwise be substantial API spend into a flat monthly rate, which is why Claude Code usage has migrated heavily onto subscriptions. If you write software and evaluate only one thing on this page, evaluate this.

4. MCP — the integration standard Anthropic started

The Model Context Protocol is Anthropic's most consequential export. Released in November 2024, MCP is an open standard for connecting AI models to external tools and data — a universal adapter, so that a Slack integration or a database connector is written once and works with any MCP-capable model instead of being rebuilt per vendor.

It won. OpenAI adopted it in March 2025; Google, Microsoft, and Salesforce followed within the year. In December 2025 Anthropic donated the protocol to the Linux Foundation under a new Agentic AI Foundation, with OpenAI, Google, and Microsoft as co-sponsors — the moment it stopped being Anthropic's protocol and became industry plumbing. ⚑ unverified By mid-2026 the ecosystem counts tens of thousands of public MCP servers and roughly 97 million monthly SDK downloads, and a large majority of enterprise AI teams report MCP-backed agents in production. ⚑ unverified

For a buyer, MCP matters two ways. First, anything you connect to Claude — internal databases, SaaS tools, file systems — rides a standard that every major vendor now supports, so the integration work is portable rather than lock-in. Second, it signals how Anthropic competes: by making the agent layer open and betting it has the best model to run on top of it.

5. The developer platform

The API is a single well-designed surface (platform.claude.com) with official SDKs in Python, TypeScript, Java, Go, Ruby, C#, and PHP, plus a full-featured CLI (ant). The features that change the economics:

• Prompt caching — repeated context (a system prompt, a document, a codebase) is cached and re-read at roughly a tenth of the input price. For agentic workloads that resend history every turn, this routinely cuts costs 50–90%. ⚑ unverified
• Batch API — asynchronous processing at 50% off for anything that can wait up to a day. ⚑ unverified
• Structured outputs — schema-guaranteed JSON, so extraction pipelines don't need retry-and-parse loops.
• Managed Agents — a newer hosted surface where Anthropic runs the agent loop and a sandboxed container per session; you define the agent, it executes with files, bash, and MCP tools server-side.

Claude is also available through Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, plus an Anthropic-operated "Claude Platform on AWS" with same-day feature parity — relevant if your procurement or compliance runs through a cloud vendor. ⚑ unverified

One more platform note that separates Claude in practice: the Claude Agent SDK packages the entire Claude Code harness — the agent loop, file tools, bash, search, permissions, sub-agents — as a library you can embed in your own product. Teams that want "a coding agent inside our app" without building the orchestration layer from scratch start here, and it is a large part of why Claude sits underneath so many third-party AI developer tools.

---

Pricing — plans and API

Consumer and team plans ⚑ unverified

Annual billing runs roughly 15% below the monthly rates. ⚑ unverified

The plan structure has one honest wrinkle: usage limits are opaque. "5x" and "20x" are multiples of a Pro baseline that Anthropic does not publish as a hard number, and heavy Claude Code users do hit the ceilings. The rule of thumb that holds up: casual use fits in Pro; a professional developer using Claude Code daily belongs on Max 5x; only sustained all-day agentic work justifies Max 20x.

API ⚑ unverified

Pay-per-token at the rates in the family table above; no subscription required, billed per request through the developer console. Sonnet 5's introductory pricing ($2/$10 per million tokens through August 31, 2026) makes it the aggressive default for production workloads this year. With prompt caching and batching applied, real-world effective rates land well below the sticker.

---

Honest limits

It is not a full multimodal suite. Claude reads images and PDFs well, but it does not generate images, video, or audio. If your workflow includes creative asset generation, you will pair Claude with something else. OpenAI and Google both ship broader modality coverage in one subscription.

Frontier pricing is frontier pricing. Fable 5 at $10/$50 per million tokens is among the most expensive mainstream API models available, and Opus-tier agentic runs consume tokens fast — a single long autonomous session can burn dollars, not cents. The mitigations (caching, batching, Sonnet 5 for volume) are real, but budget-sensitive high-volume work goes to Haiku or elsewhere.

Refusals are a working constraint at the top tier. Fable 5's safety classifiers target biology and offensive cybersecurity, and false positives on adjacent legitimate work (security tooling, life-sciences research) happen. Anthropic ships an automatic fallback mechanism to Opus 4.8, but if you work in those fields, the friction is real and you should test before committing. ⚑ unverified

Usage limits bite on subscriptions. The flat-rate plans meter usage in ways that are deliberately fuzzy, and the ceilings have tightened more than once as Claude Code demand grew. If your workload is genuinely unbounded, the API is the honest pricing surface.

Consumer polish trails. The claude.ai app is capable — Projects, Artifacts, connectors, voice — but ChatGPT remains ahead on breadth of consumer features, custom bots, and third-party plugin ecosystem. Claude's center of gravity is professional work, and the product shows it.

---

When to choose Claude

Choose Claude when:

• The work is software. Claude Code plus Opus-tier models is the strongest agentic coding stack shipping today, and the subscription bundling makes it the best value for a working developer.
• The work is long documents. Contracts, case files, codebases, research corpora — the 1M-token window plus strong retrieval-free reading is the differentiator.
• You are building agents. The API's agent surface (tool use, MCP, prompt caching, Managed Agents) is the most complete, and MCP-based integrations are portable industry infrastructure, not vendor lock-in.
• Consistency and steerability matter more than breadth. Claude follows detailed instructions literally and holds character across long sessions — valuable in production pipelines and professional drafting.
• Your data posture is conservative. Anthropic's enterprise data terms, no-training-by-default posture on commercial tiers, and cloud-marketplace availability (Bedrock, Vertex, Foundry) fit cautious procurement. ⚑ unverified

Look elsewhere when:

• You need image, video, or audio generation in the same product.
• You want the largest consumer ecosystem — custom bots, integrations, a mobile-first assistant for everything.
• Your workload is high-volume and cost-dominated and a cheaper open-weight or Luna/Flash-class model clears your quality bar.

---

How to start

As a user (10 minutes):

1. Create a free account at claude.ai.
2. Give it a real task from your actual work — a document to analyze, a draft to rewrite — not a trivia question. Claude's gap over free-tier competition shows on work, not chat.
3. If it earns a place in your week, Pro at $20/mo unlocks Opus-tier models and Claude Code. ⚑ unverified

As a developer (30 minutes):

1. Install Claude Code: npm install -g @anthropic-ai/claude-code, then run claude in a repo and log in with your Pro/Max account. ⚑ unverified
2. Hand it a genuine bug or small feature and watch it work end-to-end — plan, edit, test, iterate. This is the fastest way to calibrate what current agents can and cannot do.
3. For API work: get a key at platform.claude.com, pip install anthropic (or the TypeScript SDK), and start with claude-sonnet-5 — the price/quality default — moving up to Opus 4.8 only where quality measurably demands it.
4. Turn on prompt caching before you scale anything. It is the single largest cost lever on the platform.

As a team: run a two-week pilot on Team Standard with your three heaviest document-or-code workflows, and measure time saved rather than vibes. Claude's economics justify themselves on work throughput or not at all.

---

Prices, plan structures, and model availability change frequently; every figure above marked ⚑ unverified should be checked against claude.com/pricing and platform.claude.com before publication.