Claude is Anthropic's family of artificial intelligence models, designed to be more controllable and more reliable than its competitors. The current lineup includes Opus 5.5, released on September 22, 2026 and recommended by default, Fable 5.1 for the most demanding reasoning, Sonnet 5 and Haiku 4.5. Complete ecosystem with Claude Code, Cowork, Design, the Microsoft 365 add-ins and the Webflow MCP connector.
Claude is built by Anthropic, an American company based in San Francisco and founded in 2021 by former OpenAI researchers including Dario Amodei and Daniela Amodei. The company was built around a simple idea: a more controllable model beats a more spectacular one.
That approach has a name, Constitutional AI, documented in this Anthropic research publication. Claude follows an explicit set of principles and self-corrects rather than relying solely on human filtering after the fact. In practice, the model is more willing to admit it does not know.
Who is it for, exactly? Claude is aimed primarily at:
Claude has been the primary assistant at our Webflow agency for two years. We use it for writing SEO and AEO articles of 2,000 words and up, analyzing client briefs that sometimes run to 150 pages and generating custom Webflow code. Opus 5.5 handles our day-to-day work, Fable 5.1 takes the heaviest tasks. The team runs on two Max 5x subscriptions, one for development and one for agency management and growth. The API costs us between $50 and $100 a month to run our internal tools.
Anthropic maintains four reference models in production. Opus 5.5, released on September 22, 2026, opens the 5.5 family and replaces Opus 5 as the recommended default. Fable 5.1 replaced Fable 5 on September 1, at the same price. Anthropic has announced Sonnet 5.5 and Haiku 5.5 for the coming weeks, with no date or pricing yet. Earlier versions remain available through the API for integrations that need a pinned version.
| Model | Released | API pricing | Context | Max output | Best for |
|---|---|---|---|---|---|
| Claude Opus 5.5 Gen 5.5 | Sept. 22, 2026 | $4 / $20 | 1M | 128K | The default model: writing, code, agents, document analysis |
| Claude Fable 5.1 Fifth gen | Sept. 1, 2026 | $10 / $50 | 1M | 128K | The most demanding reasoning, agentic sessions running for hours |
| Claude Sonnet 5 Fifth gen | June 30, 2026 | $2 / $10 | 1M | 64K | High-volume integrations, cost-controlled agents |
| Claude Haiku 4.5 | Oct. 2025 | $1 / $5 | 200K | 64K | Real-time chatbots, classification, high volume |
| Fable 5, Opus 5, Opus 4.5 to 4.8, Sonnet 4.5 and 4.6 API only | July 2026 and earlier | $3 to $10 input | 200K to 1M | 64K to 128K | Integrations that need a pinned version |
Prices in US dollars per million tokens, input / output. Source: Anthropic documentation, September 2026.
Opus 5.5 by default. Anthropic places it "at the level of Claude Fable 5.1 on most work" for 40% of the price, at $4/$20. Its system card backs that up: 89.9% on SWE-bench Pro against 81.2% for Fable 5.1. On Terminal-Bench 4.0, the gap reaches 66.4% against 55.8%. It also costs 20% less than Opus 5, which it replaces. We use it every day for writing, analysis and code. Details in the official Claude Opus 5.5 announcement.
Fable 5.1 for the heaviest tasks. Anthropic reserves it for the most demanding reasoning and agentic sessions lasting several hours. At $10/$50, it costs two and a half times as much as Opus 5.5. We give it full client specifications, whole-site migrations and long autonomous sessions. Source: official Anthropic documentation.
Sonnet 5 for volume. At $2/$10, it keeps the best ratio between cost and capability in the range for integrations that process large volumes. Anthropic dropped the increase to $3/$15 that had been scheduled for September 1: the launch price becomes the standard price. Haiku 4.5 stays reserved for simple, very high-volume tasks such as classification or real-time chatbots.
One detail changes every budget calculation and almost no one talks about it. Since Opus 4.7, Anthropic uses a new tokenizer that produces roughly 30% more tokens for the same text.
The consequence is immediate. Moving from Sonnet 4.6 ($3/$15, old tokenizer) to Sonnet 5 ($2/$10, new tokenizer) does not save 33%, because the same prompt consumes more tokens. The real saving lands closer to 10 to 15% depending on the type of content.
Reasoning effort also blurs comparisons. Opus 5.5 runs at medium effort by default, against high for Opus 5. Anthropic presents the result as a model 40% cheaper to run. Artificial Analysis nonetheless measures an almost identical cost per task at maximum effort, $5.98 for Opus 5.5 against $5.86 for Opus 5, because the new model consumes more tokens.
Before locking an API budget, re-measure with the token counting endpoint at the effort level planned for production, rather than applying simple arithmetic to the listed rates. The detail is in the Anthropic pricing documentation.
Claude Code runs where the code lives, not in a browser tab. It reads the codebase, runs commands, edits files and handles Git workflows in plain language.
The tool is no longer limited to the terminal. It ships as a desktop app for macOS and Windows, on the web at claude.ai/code and as an extension for VS Code and the JetBrains IDEs. The terminal remains the most complete surface.
Technical requirements:
Anthropic now recommends its native installer: curl -fsSL https://claude.ai/install.sh | bash on macOS, Linux and WSL, or irm https://claude.ai/install.ps1 | iex in PowerShell. Homebrew, WinGet and npm also work. Every method ends with browser authentication. Details in the Claude Code setup documentation.
How it works: on startup, Claude Code indexes the project and reads the CLAUDE.md or AGENTS.md file describing the technical stack, code conventions and testing rules. It asks permission before every file change and every command it runs.
Claude Code has transformed our Webflow workflows. Use cases in production:
The latest releases point toward team usage: a restricted mode that scopes what the agent can reach, self-hosted environments for Team and Enterprise, and skills and plugins synced from the claude.ai account. Since September 22, 2026, Opus 5.5 is the default model on the Pro and Team standard plans, in place of Sonnet. Full details: Claude Code official changelog.
Since September 16, 2026, chat and Cowork form a single application. One conversation can launch a long task that keeps running in the cloud after the laptop is closed. The rollout starts with the Pro and Max plans, then Team and Free, with 30 days' notice for Enterprise accounts. Details in the Anthropic announcement.
Claude Cowork remains the agent aimed at non-developers. It reaches local files and authorized applications, then runs multi-step tasks end to end. It is available on desktop for macOS and Windows, still in beta on web and mobile. Typical use cases:
Claude Design generates designs, interactive prototypes, slides and one-pagers from written instructions. It reads your design system from your codebase or your Figma files to apply your brand automatically. It now works directly inside conversations, alongside Claude Docs and Claude Slides, in beta on paid plans. Export to Canva, PDF, PPTX, HTML, or hand off directly to Claude Code for implementation.
Browser and office apps: Claude in Chrome has been open to all paid plans since August 26, 2026. The extension acts autonomously in the browser, under a classifier that checks every action. On the Microsoft 365 side, Claude integrates natively into Excel, PowerPoint and Word, with Outlook still in beta. Context carries over between applications: work started in Excel can finish in a PowerPoint deck without explaining everything again.
Skills and plugins: bundles combining instructions, connectors and sub-agents that turn Claude into a domain specialist. Anthropic runs a public marketplace. Enterprise administrators can create private ones for their teams. For developers, the Skills API has been generally available since August 20, 2026.
Six frontier models shipped in three weeks: Fable 5.1 on September 1, Gemini 3.8 Flash on the 2nd, GPT-6 Astra on the 3rd, then Opus 5.5, GPT-6 Sol and GPT-6 Luna on the 22nd. None of them wins every benchmark. Each vendor publishes the results that favor it. The independent leaderboards contradict each other too.
| Model | Vendor | Released | API pricing | Main strength |
|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic | Sept. 22, 2026 | $4 / $20 | No. 1 on the Artificial Analysis Intelligence Index at launch (58), 20% cheaper than Opus 5 |
| Claude Fable 5.1 | Anthropic | Sept. 1, 2026 | $10 / $50 | No. 1 on Agent Arena, No. 2 on WebDev Arena behind GPT-6 Astra |
| Claude Sonnet 5 | Anthropic | June 30, 2026 | $2 / $10 | 1M tokens of context, launch price now permanent |
| GPT-6 Astra | OpenAI | Sept. 3, 2026 | $10 / $50 | No. 1 on WebDev Arena, 97.6% on FrontierMath Tier 4 according to OpenAI |
| GPT-6 Sol | OpenAI | Sept. 22, 2026 | $2 / $10 | Direct competitor to Sonnet 5, at the same price |
| Gemini 3.8 Flash | Sept. 2, 2026 | $0.75 / $3.75 | The cheapest and fastest in the comparison, natively multimodal |
GPT-6 Astra charges double beyond 272K input tokens. Gemini 3.8 Flash at introductory pricing through December 31, 2026, then $1.50 / $7.50. Artificial Analysis and Arena rankings recorded on September 23, 2026, before Opus 5.5 appeared on WebDev Arena.
A vendor publishes the benchmarks it chose to measure, under the conditions it chose. Terminal-Bench 4.0 is a clear example. Anthropic reports 66.4% for Opus 5.5 and OpenAI 57.9% for GPT-6 Astra. Vals.ai measures 61.6% against 57.1% on the same test. Artificial Analysis puts the two models level, at 59.6%.
The listed price misleads just as much. GPT-6 Astra and Fable 5.1 both cost $10/$50 per million tokens. Artificial Analysis nonetheless records $3.26 per task for Astra against $7.63 for Fable 5.1 on its index, because the two models do not consume the same volume of tokens. The only measurement that counts for a team remains a test on its own use cases, with its own prompts.
We gave both models the same brief: the landing page for ARC, a fictional customer intelligence SaaS, with an instruction to be original and no skill, template or design file. GPT-6 Astra ran in Codex at low effort, Fable 5.1 in Claude Code at medium effort. The test ran in three stages: a first draft, an analysis drawing on Mobbin's 600,000 UX and UI references through MCP, then a redesign based on that report.
On the first draft, Astra takes a clear lead. In 11 minutes 34 seconds it delivers interactive dashboards and a polished integrations section. Fable 5.1 takes around fifteen minutes for a less finished result. Both keep the generic typography of sites generated with Claude Code or Lovable. Both also get the third-party tool logos wrong.
The gap narrows after the Mobbin analysis. Both models flag the same flaws: an H1 that fails to state the promise and no button to book a demo. Astra's report goes further. In the final version, Astra produces interactive assets and a more restrained art direction. Fable 5.1 handles its orange and black palette well and draws its own illustrations, without native image generation.
Our verdict leans toward Astra. The independent leaderboards point the same way without settling it. Astra leads WebDev Arena with 1,793 points against 1,753 for Fable 5.1. Design Arena nonetheless ranks Fable 5.1 6th in the Website category, well ahead of Astra in 16th. The prompt and the context weigh more than the choice of model: a design system turned into a design.md file plus targeted references bring the two results much closer.
Opus 5.5 came out after this test. Anthropic places it at the level of Fable 5.1 on most tasks for 40% of the price. It is not yet ranked on WebDev Arena. Design Arena still has too few matchups to call it.
Claude carries the bulk of our production. Opus 5.5 is our everyday model for writing, analysis and code. Fable 5.1 takes over on the heavy tasks: full client specifications, migrations, long agentic sessions. We also drive the Webflow CMS from Claude through the MCP connector.
GPT-6 Astra is our design model. On landing page mockups, it produces more polished interfaces on the first draft, as our test showed. Routing by task has replaced picking a single model. Settling that routing is part of our agentic web strategy engagements: which model on which task, at what cost, with what level of human supervision.
Two grids coexist: subscriptions billed per user and the API billed per token. The free plan gives access to Sonnet and Haiku only, with no Opus, Fable or Claude Code. Fable keeps a separate status: on Pro and Team standard, it can only be used with paid usage credits. Max and Team premium can spend up to half of their weekly limits on it. Sources: claude.com/pricing and the Anthropic help page on Fable access.
| Plan | Price | Models | Included |
|---|---|---|---|
| Free | $0 | Sonnet and Haiku, with quotas | Web, iOS, Android and desktop, web search, memory, file creation, Slack and Google Workspace connectors |
| Pro | $17 billed annually $20 billed monthly |
Opus, Sonnet and Haiku. Fable with paid credits | Everything in Free, plus Claude Code, Cowork, Design, Docs and Slides, Science, Research, Claude in Chrome, unlimited projects and Microsoft 365 integration |
| Max | From $100 | The full range, Fable up to 50% of weekly limits | Everything in Pro, with 5x or 20x higher usage, higher output limits, early access to new features and priority during peak hours |
| Team | $20 / $25 standard seat $100 / $125 premium seat |
Standard: same as Pro. Premium: same as Max | Enterprise search, central billing, SSO, admin controls, no model training on your content by default |
| Enterprise | $20 per seat + API usage |
The full range, Fable depending on the organization's settings | Everything in Team, plus SCIM, audit logs, Compliance API, custom retention, IP allowlisting, HIPAA-ready option and Claude Security in beta |
Prices per user per month. The 5-hour usage limits were raised on September 22, 2026 on paid plans, with no figure published. Source: claude.com/pricing, September 2026.
On the API side, billing is per million tokens, with two optimization levers that often weigh more than the choice of model itself.
| Model | Input | Output | 5 min cache write | Cache read | Batch |
|---|---|---|---|---|---|
| Haiku 4.5 | $1 | $5 | $1.25 | $0.10 | $0.50 / $2.50 |
| Sonnet 5 | $2 | $10 | $2.50 | $0.20 | $1 / $5 |
| Opus 5.5 | $4 | $20 | $5 | $0.20 | $2 / $10 |
| Fable 5.1 | $10 | $50 | $12.50 | $0.25 | $5 / $25 |
Prices per million tokens. A cache read costs 10% of the input price, except on Opus 5.5 (5%) and Fable 5.1 (2.5%). Batch processing applies a 50% discount. Source: Anthropic pricing documentation, September 2026.
Caching changes the economics of an agent that resends the same system context on every turn. On Fable 5.1, a cache read costs $0.25 per million tokens, forty times less than standard input. The Batch API applies a 50% discount on both input and output, for any processing that does not need an immediate answer. A well-built agent using caching and batching can cost several times less than a naive implementation on the same model.
The math is done in minutes. At $20 a month, Pro pays for itself as soon as it saves about twenty minutes a month for someone billing $60 an hour. The real argument is access: without a paid subscription, there is no Opus, no Fable and no Claude Code. For very heavy usage, the Max plan from $100 per month becomes relevant once usage limits, rather than price, are the blocking factor. It is also the only individual plan where Fable runs without extra credits.
Our own usage gives an order of magnitude. Over the last six weeks, the account dedicated to agency management and growth ran 91 Claude Code sessions and generated 5.9 million tokens, more than half of them with Fable 5.1. At API rates, that usage would have cost just over $1,000, against a Max 5x subscription billed at $100 a month.
For a team, the Team plan generally makes sense from three regular users onwards, mainly for central management, SSO and the native Microsoft 365 and Slack connectors.
Anthropic's free plan gives access to Sonnet and Haiku with usage quotas, enough for a first hands-on look. Four alternatives round out the picture.
Gives access to several Claude models with daily limits. The conversational interface is clean, history is saved and conversation sharing is built in. Good for testing before a paid subscription.
Embeds Claude in its augmented web search. Perplexity combines a search engine and an LLM. Access to Claude stays indirect, so limited, but it works for occasional discovery.
Embeds Claude natively in the Notion workspace, subject to a paid Notion subscription. Excellent for document summaries, collaborative writing and Notion database automation.
Lets you install a Claude bot in a Slack workspace. Limits are those of the standard API. Handy for internal support, automated FAQs and developer assistance.
We produce articles of 2,000 to 3,000 words optimized for both Google and LLMs (ChatGPT, Perplexity, Claude). Our six-step process:
E-E-A-T is non-negotiable: field experience cited, expertise demonstrated through hard numbers, traceable sources.
The gain sits in the middle of the process. Shaping the ideas, writing and loading the article into the CMS used to take at least two hours per article. Claude now gets it done in 15 to 20 minutes, then pushes the article as a draft into Webflow through MCP. Framing the topic upstream and reviewing downstream stay in our hands, with several rounds before the angle is signed off.
We regularly handle documents of 50 to 150 pages: strategic briefs, wireframes, functional specifications, technical constraints. Claude extracts for us:
The PDF goes straight into the interface. These files are among the heavy tasks we give to Fable 5.1: the whole document fits in its 1M token context. An analysis that used to take weeks now wraps up in a few hours, commercial proposal included. We write around three proposals a month this way, 50 since May 2025. The largest file came from an industrial company.
On our client projects, these workflows have already been used to migrate and map several hundred pages. Priority use cases tested in production:
The step-by-step workflow is in our guide to the Claude x Webflow MCP connector for automating the CMS, with the exact permissions to configure and the use cases validated in production.
Advanced scroll animations, dynamic CMS filters, third-party API integrations, JavaScript performance work. Claude produces clean, commented code that respects Webflow conventions and handles errors properly.
The standard workflow is simple: functional brief, initial code generation, cross-browser testing, fix iterations, production deployment.
On the upstream design phase, we also document our Claude + Skills workflow to generate a homepage look and feel ready for Figma in minutes. Skills combined with a structured brief let us compress the blank-page phase without losing coherence with the client's design system. For first landing page mockups, GPT-6 Astra now complements this setup.
This kind of automation sits at the heart of our AI and automation offering.
Opus 5.5 for the vast majority of use cases. Anthropic now recommends it as the starting point: 1M tokens of context at $4/$20 per million tokens, at the level of Fable 5.1 on most tasks according to the vendor.
Fable 5.1 for the most demanding reasoning and long agentic sessions, at $10/$50. Sonnet 5 at $2/$10 for high-volume integrations. Haiku 4.5 remains the choice for high volume and low stakes, with 200K tokens of context at $1/$5.
Fable 5.1 shipped on September 1, 2026 as Anthropic's most capable model. Opus 5.5 followed on September 22 at $4/$20, 40% of Fable 5.1's price. On the benchmarks in its system card, Opus 5.5 even beats Fable 5.1 on most coding and agentic tasks.
Both models share always-on reasoning, controlled through the effort parameter. Both also carry safety classifiers that can decline a request, returning an HTTP 200 response with a refusal reason. Any integration needs to handle that case and fall back to another model, which Anthropic offers to automate toward Opus on Fable 5.1.
The deciding difference for IT leadership is retention. Fable 5.1 requires 30-day data retention, while Opus 5.5 remains eligible for zero data retention.
Mythos 5.1 is the same model as Fable 5.1, shipped without some of its safeguards. Anthropic reserves it for Project Glasswing and its verification programs in cybersecurity and life sciences, for now with US organizations only. It is not offered in claude.ai subscriptions. Details in the Anthropic announcement of Fable and Mythos 5.1.
Anthropic pulled access to Fable 5 on June 12, 2026, three days after launch, following export controls applied by the US government. The company explained it had no reliable way to verify user nationality in real time. It chose to suspend access for everyone rather than filter approximately.
Access was restored on July 1, 2026 on the Claude Platform, claude.ai, Claude Code and Cowork, then progressively on AWS, Google Cloud and Microsoft Foundry. Details in the Anthropic statement on redeploying Fable 5.
The episode is worth remembering when building an architecture on a single model. Depending on a frontier model carries regulatory risk, independent of the vendor's technical quality.
Each model excels in different areas. The published figures mostly come from the vendors themselves.
Claude excels at:
ChatGPT (GPT-6 Astra) excels at:
We use both depending on the task: Claude for production, GPT-6 Astra for mockups. Intelligent routing beats an exclusive choice.
Partly. The Free plan gives access to Sonnet and Haiku on web, mobile and desktop, with usage quotas. Opus, Fable and Claude Code require a paid subscription, starting at $17 a month billed annually with the Pro plan.
The listed price is not enough to build a budget. Four factors move the real bill.
The tokenizer first. Since Opus 4.7, the same text consumes roughly 30% more tokens than with earlier models.
Prompt caching next. A cache read costs 10% of the standard input price, 5% on Opus 5.5 and 2.5% on Fable 5.1. An agent resending the same context on every turn benefits on every call. Be careful though: caching works on prefix matching, so any variable element at the start of the prompt, such as a timestamp, drops the cache rate to zero with no error message at all.
The Batch API applies a 50% discount on input and output for anything that is not real time.
Reasoning effort last. A higher effort level multiplies output tokens. The cost of the same model therefore varies widely with the setting you choose.
Claude Code is included in every paid subscription from the Pro plan up, at $17 a month billed annually or $20 monthly. The free plan does not include it. Developers who hit the limits move to Max, from $100 a month.
Through the API, Anthropic measures an average cost of about $13 per developer per active day, under $30 for 90% of users. Source: Anthropic blog on task costs with Opus 5.5. Claude Design and Cowork are also included from the Pro plan, at no extra cost.
Anthropic introduced Claude Code on February 24, 2025 as a research preview, alongside Claude 3.7 Sonnet. The tool became generally available on May 22, 2025, with the Claude 4 generation.
Claude Code requires an install (native installer, Homebrew or npm) and a connection to a paid Claude account or the API. Used from the command line, the desktop app or an IDE extension, for experienced developers. Local file manipulation, native Git workflows, full codebase debugging. Ideal for advanced automation.
The Webflow MCP connector activates in two clicks from Claude.ai, with no technical setup. Accessible to non-developers. Operations on the Webflow CMS, pages and metadata through the Designer API and the Data API. A no-code approach suited to marketing, content and SEO teams.
Both use the Webflow APIs, but with different paradigms: a terminal for developers against a graphical interface for business teams.
No, and Anthropic does not claim otherwise. Claude augments developers, it does not replace them.
Tasks Claude handles excellently:
Critical limits:
Realistic position: Claude is the equivalent of a highly productive junior to mid-level developer who needs senior supervision.
Anthropic makes solid contractual commitments on data protection. The main guarantees:
inference_geo parameter, with a 10% premium for US-only inferenceEnterprise plans add SAML and OIDC SSO, custom DPAs, audit logs and a Compliance API. Inference hooks, in Enterprise beta, inspect prompts and tool calls before they reach the model.
For the most sensitive data, Anthropic offers deployment through AWS Bedrock or Google Cloud inside a private client VPC. Fable 5.1 requires 30-day retention, which rules it out of zero-retention environments. Since September 1, 2026, Enterprise Frontier Safeguards keep those logs in the client's own cloud, with zero retention granted to eligible customers during the rollout. Opus 5.5 remains eligible for zero data retention.
No. Claude is built by Anthropic, an American company based in San Francisco. The confusion often comes from Mistral AI, the French company behind the Le Chat assistant. Anthropic announced a Paris office in November 2025. Claude answers in French at a quality close to English.
Yes. Claude supports more than 15 languages including French, at a quality close to English on our workloads, with no literal-translation feel. We write most of our French editorial output with it.
Multilingual performance we observe:
One limit to know: public benchmarks are overwhelmingly English, so quantitative comparisons in other languages remain poorly documented. Practical recommendation: run A/B tests on your own non-English use cases before deploying to production.
.svg.png){kind=logo}
