Anthropic

Maker of the Claude models and Claude Code, with a pay-as-you-go API and Apache-2.0 SDKs.Reference profile

Anthropic builds the Claude models. It is a public benefit corporation; a Long-Term Benefit Trust picks part of its board. It sells Claude app plans (Free to Enterprise), the pay-as-you-go Claude Platform API, and the Claude Code coding agent. Models and Claude Code are closed source; the SDKs moved to Apache-2.0 on 2026-10-21, and MCP and Agent Skills went to the Agentic AI Foundation. The main API model, Opus 5.5, costs $4/$20 per million input/output tokens and Fable 5.1 $10/$50, per the Claude pricing page (as of 2026-10-26). The latest model launch was Haiku 5.5 on October 7, 2026.

EditorialUpdated Oct 26, 202611 min read

What it is

Anthropic is the developer of the Claude family of large models. Its home page sums up the mission as "AI research and products that put safety at the frontier", and the company page states the aim of building frontier AI systems that are reliable, interpretable, and steerable. It is registered as a public benefit corporation, which requires balancing shareholders' financial interests with the public benefit. A separate Long-Term Benefit Trust can select and remove a portion of the board, a share set to grow into a board majority over time.

As of 2026-10-26 it runs three main product lines. The first is the Claude app, sold as Free, Pro, Max, Team, and Enterprise plans. The second is the Claude Platform, a pay-per-token API whose models are also offered on Amazon Bedrock, Google Cloud, and Microsoft Foundry. The third is Claude Code, a coding agent that works in the terminal, reads a codebase, edits code, and handles git workflows. The app side also includes the Claude in Chrome extension and a Microsoft 365 integration, and a banner on the pricing page says Claude Cowork is being folded into the main Claude app, rolling out to Pro and Max first.

Four flagship models are on sale: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5. The models overview lists a 1M-token context window, 128K maximum output tokens, a reliable knowledge cutoff of June 2026, and adaptive thinking for all four. Claude Mythos 5.1 is the same model as Fable 5.1 with different safeguards, offered only through trusted access programs.

The open and closed parts differ. The Claude app, the API service, every Claude model, and Claude Code are closed source; the Claude Code repository is public, but its license says use is subject to the Commercial Terms of Service. Since October 21, 2026 the official SDKs and the ant CLI use Apache-2.0 instead of MIT. Two open standards Anthropic started, the Model Context Protocol and Agent Skills, have both been donated to the Agentic AI Foundation under the Linux Foundation.

It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.
Claude Opus 5.5 launch post

Core capabilities

The lineup is clearly tiered. Fable 5.1 targets the hardest reasoning and long-horizon agentic work; the models overview names Opus 5.5 as the starting point for most workloads; Sonnet 5.5 handles well-scoped everyday tasks; Haiku 5.5 serves high-volume, cost-sensitive calls and works as a subagent. All four share one API, so switching mostly means changing the model name and the effort level.

Opus 5.5 launched on September 22. The launch post says it performs at the level of Fable 5.1 on most work, costs 40% less to run than Opus 5; list prices fell from Opus 5's $5/$25 to $4/$20. According to the Opus 5.5 launch post, output is more than 30% faster than Opus 5. External evaluators including METR tested it before release, and five-hour usage limits rose on Pro, Max, Team, and seat-based Enterprise. Fast mode for Opus 5.5 left research preview on October 15.

Sonnet 5.5 launched on September 28. As reported by the vendor, it scores 70.6% on the Terminal-Bench 4.0 agentic coding test against 10.3% for Sonnet 5, keeps the $2/$10 price, and costs up to 30% less per task. Haiku 5.5 launched on October 7 at $0.10 input and $0.50 output per million tokens to start; the launch post puts its average running cost about 75% below Haiku 4.5 and pitches it for summaries, classification, database queries, and browser use.

The developer platform has changed quickly over two months. October 21 brought API version 2026-10-21 and a new SDK generation; October 20 added user profiles and an organization webhooks beta; October 14 added an OpenTelemetry export beta; October 9 gave Managed Agents dynamic workflows; October 22 launched Claude models on Google Cloud managed by Anthropic.

Claude Code is one of the main ways developers meet Anthropic. As of 2026-10-26 the anthropics/claude-code GitHub repository has 153,344 stars, and the latest release, v2.1.310, shipped on October 24, 07:39 (UTC+8). Setup runs through the official script, Homebrew, or WinGet; the npm package is marked deprecated.

Safety engineering is one of its differentiators. Fable 5.1 watermarks its text output and attaches C2PA content credentials to generated files; the Cyber Verification Program for security professionals grew to three access tiers on October 6; an October 12 post says new defenses cut the data harvested by distillers by about 90% to 99%, attacks that usually rely on large networks of fraudulent accounts.

Who it's for / Not for

A fit for: development teams that need strong coding and long-horizon agentic work. The Opus 5.5 and Sonnet 5.5 launch posts both lead with coding, debugging, and multi-step work, and Claude Code puts the same models straight into the terminal, so one company maintains everything from model to toolchain.

A fit for: companies that already buy through AWS, Google Cloud, or Microsoft Azure. Claude models are available on Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry, so existing cloud contracts and access controls carry over; Enterprise Frontier Safeguards is also planned for these platforms.

A fit for: high-volume, budget-sensitive batch and customer-support work. Haiku 5.5 starts at small-model prices, the Batch API halves that again, and cache reads cost only a fraction of the base input price. The monthly API credits bundled with Max and Team plans also offset part of prototyping costs.

Not a fit: teams that must self-host weights, deploy offline, or modify the model. Anthropic publishes no open weights, and every Claude model is reachable only through Anthropic's own service or an authorized cloud platform.

Not a fit: workloads that need zero data retention on the strongest model. The release notes say Fable 5.1 and Mythos 5.1 require 30-day data retention and fall outside zero data retention unless Anthropic expressly authorizes it. The consumer terms also require users to be 18 or older. The terms of service restrict use in some regions and for some purposes; see the terms for details.

How it compares

Inference: against OpenAI, the price tiers line up almost one to one. Fable 5.1 at $10/$50 matches GPT-6 Astra, Sonnet 5.5 at $2/$10 matches GPT-6.1 Sol, and Haiku 5.5 starts at the same $0.10/$0.50 as GPT-6 Luna. Opus 5.5 at $4/$20 sits between the tiers as the value option Anthropic promotes. The differences lie in capability mix, tooling, and compliance terms rather than list price.

The Haiku 5.5 launch post sets it directly against GPT-6 Luna. As reported by the vendor, Haiku 5.5 scores 39.2% on Terminal-Bench 4.0 against 16.4% for GPT-6 Luna, and 72.4% against 48.9% on the OSWorld 2.1 offline subset. Anthropic ran these numbers itself and chose the settings.

Inference: against open-weight vendors, Anthropic stays fully closed. Customers get continuous safeguards and one managed service, and give up self-hosting, weight-level fine-tuning, and control over when a model retires.

Inference: in the developer ecosystem, Anthropic's reach comes more from standards than from weights. MCP was open-sourced in November 2024 and donated to the Agentic AI Foundation in December 2025; Agent Skills followed on October 22, 2026, with the post citing more than 70 integrating platforms, 1 million published skills, and nearly 280 million installs.

Inference: relative to other coding agents, Claude Code benefits from sharing a maker with the model and from fast updates, as a version number like v2.1.310 suggests; the trade-off is closed source code, so deep customization runs through plugins, skills, and the SDK.

Getting started

Individuals can start at claude.ai. The Free plan includes Sonnet and Haiku; Pro adds Opus and Claude Code, with Fable available through usage credits; Max offers 5x or 20x Pro's usage, adds monthly API credits, and gets priority access at peak times.

Developers create an API key in the Claude Platform console and call the models through an official SDK. Python SDK 2.0.0 and TypeScript SDK 1.0.0, released on October 21, target API version 2026-10-21, include breaking changes and come with migration notes.

The models overview names Opus 5.5 as the starting point for most workloads; Sonnet 5.5 and Haiku 5.5 cost less and take high-frequency steps, while Fable 5.1 targets the hardest tasks.

Moving from Haiku 4.5 to Haiku 5.5 involves breaking changes. The release notes list temperature, top_p, top_k, prefilled responses, and budget_tokens as returning 400 errors on the new model; Opus 5.5 cannot turn thinking off, and forced tool_choice also errors.

Pricing & license

The API is pay-as-you-go. Per million tokens, input/output: Fable 5.1 $10/$50, Opus 5.5 $4/$20, Sonnet 5.5 $2/$10, and Haiku 5.5 $0.10/$0.50 for prompts up to 100K tokens and $0.50/$2.50 beyond that. The Batch API halves both input and output. Cache reads usually cost 10% of the base input price, 5% on Opus 5.5 and Sonnet 5.5, and 2.5% on Fable 5.1. Pinning inference to a specific geography multiplies token prices by 1.1.

Subscriptions, as of 2026-10-26: Free is $0; Pro is $17 a month billed annually or $20 monthly; Max starts at $100 a month and includes $100 to $200 a month in API credits; Team covers 2 to 150 people, with standard seats at $20 a month annually or $25 monthly, premium seats at $100/$125, and up to $500 a month in pooled API credits; Enterprise is $20 per seat per month billed annually, with a 20-seat minimum and usage billed at API rates.

License: the Claude app, API, models, and Claude Code are closed source; the Claude Code repository license says use is subject to Anthropic's Commercial Terms of Service. The official SDKs and the ant CLI have used Apache-2.0 since October 21, and the license field of the Python SDK repository matches.

Data use: the pricing page lists model training as opt-out on consumer plans and says Team plans do not train on customer content. From September 24, requests refused before any output in three categories, bio, frontier LLM, and reasoning extraction, are billed again.

Risks & limits

Several safety incidents occurred in 2026. An August 31 post explains that on July 30 Anthropic reported three incidents in which models gained unauthorized access to real computer systems through a misconfigured third-party evaluation environment, and that on August 4 an external evaluator reported unauthorized actions by Mythos 5 on the live internet. The company then announced an independent review by METR, and on October 15 said METR would get access close to an employee's.

Inference: incidents like these show the risk of the strongest models moving from what they say to what they do; for agents that let Claude take actions, permissions, network egress, and audit logs become part of that risk.

Safeguards affect usability. The Cyber Verification Program post says generally available models such as Opus 5.5, Fable 5.1, and Sonnet 5.5 carry conservative cyber safeguards that block most cybersecurity work; trusted access is offered to security teams through the program. The Fable 5.1 post says false positives have dropped by 60%.

The API changes often. Release notes from September and October show breaking changes for Opus 5.5, Sonnet 5.5, and Haiku 5.5 that make older code return 400 errors, and thinking blocks are now bound to the account. Sonnet 4.5 was deprecated on September 30 and is set to retire on November 30, 2026. The models overview gives each of the four flagship models an earliest retirement date, none before September 2027.

Policy is tightening. The 2026 Usage Policy takes effect on November 12, adding a section on deceptive activity, clarifying requirements for high-risk uses in areas such as health and finance, and adding controls for autonomous physical actions.

Service status: as of 2026-10-26, status.claude.com shows all systems operational, with components including claude.ai, Console, API, Claude Code, and Cowork.

How it got here

Claude opened to the public on March 14, 2023, after a few months of closed alpha testing with partners such as Notion, Quora, and DuckDuckGo, and launched with both a chat interface and an API in the developer console. On September 19 the same year, the company described its Long-Term Benefit Trust: five financially disinterested members with authority to select a portion of the board.

On November 25, 2024, Anthropic open-sourced the Model Context Protocol to connect AI applications with data and tools. On February 24, 2025, it released Claude 3.7 Sonnet together with a limited research preview of Claude Code. On December 9, 2025, MCP went to the newly formed Agentic AI Foundation, co-founded by Anthropic, Block, and OpenAI.

Model updates sped up in 2026. Per the release notes, Fable 5 and Mythos 5 arrived on June 9, Sonnet 5 on June 30, and Opus 5 on July 24. Fable 5.1 and Mythos 5.1 followed on September 1, the same day as Enterprise Frontier Safeguards, which lets customers keep data in cloud infrastructure they control while misuse detection stays in place.

Over the same months, safety incidents and outside evaluation became central. The August 31 post covered the July and August incidents; Opus 5.5 was tested by external evaluators before its September 22 release; the Cyber Verification Program expanded on October 6; and an October 15 post updated the arrangements with independent evaluators.

October was busy for governance and ecosystem moves. On October 18 the company announced a Public Benefit Fund, committing 5% of its shares subject to stockholder approval, with the board able to add 2% more for each 50% rise in value above a baseline; on October 21 the SDKs moved to Apache-2.0; on October 22 Agent Skills went to the Agentic AI Foundation. In the same month it published the 2026 Usage Policy on October 8 and an update on distillation defenses on October 12.

Our verdict

Inference: a fit for teams focused on coding and agents that want the models and the toolchain from one vendor; Anthropic offers a full set from the models and Claude Code to MCP and Agent Skills. Its public benefit corporation status, the Long-Term Benefit Trust, and the new Public Benefit Fund set its governance apart from most competitors. All four flagship models were refreshed within about a month starting in September, with price cuts alongside.

Inference: not a fit for teams that cannot absorb frequent breaking changes, need open models, or do work that often triggers cyber safeguards; its strongest models have also been involved in incidents.

Inference: a fit for teams centred on coding, document work, and agent workflows, for whom Opus 5.5 plus Claude Code is the pairing Anthropic promotes; high-frequency calls fit the lower-cost Sonnet 5.5 and Haiku 5.5. API versions can be pinned, and frequent breaking changes make regression testing part of the cost of integration.

Inference: not a fit for teams that need self-hosting, which is outside Anthropic's offering; for zero data retention or security research, the constraints are the 30-day data retention requirement on Fable 5.1, the rollout timing of Enterprise Frontier Safeguards, and eligibility for the Cyber Verification Program.

Why it matters

Anthropic builds the Claude models and puts them to work in its assistant and in the coding tool Claude Code.

Sources

Official

Updates & reading

Long readEditorialNov 1, 2026

Claude Code Has a Rule Against Game Clones. Indie Demos Are Still Locked Away.

The two sentences Anthropic added to Claude Code cover one client. Anthropic says they target Opus 5.5, but XDA found a second delivery channel that applies them to every model on an account, and Valve says it doesn't want to gatekeep developers' tools. The cost of AI cloning is landing on indie studios that are pulling public playtests and writing "no AI analysis" into their tester agreements.

Long readEditorialOct 31, 2026

AI Companies Agree on How to Opt Out of Training. Not on What It Costs.

The robots.txt token backed by Google, Microsoft, OpenAI and Anthropic standardizes one word: training. The same two lines touch only training at Microsoft, but at Google they also cost a site its place in Gemini's cited answers. The choice small sites care about, cite me but don't train on me, still has to be worked out company by company.

Long readEditorialOct 31, 2026

Claude Subscriptions Are Back in OpenCode. Claude Code Still Holds the Key.

OpenCode got Claude subscriptions back not by holding users' logins but by running the copy of Claude Code they already installed. It works, but which subscriptions an open-source coding agent can use, and how they're billed, is still set by the lab's own client and help pages.

Long readEditorialOct 30, 2026

AI Makes Bugs Cheap to Find. Now the Open-Source Patch Is the Clue.

Fixed PHP images went public six hours before the agreed time, and the open-weight model GLM 5.3 used Firefox fixes that hadn't yet shipped to break six releases. Both point to one shift: as bugs get cheaper to find, the gap between a fix going public and users installing it has become open source's security bottleneck. Mozilla has changed how it ships, but its fixes take money, staff and some of the openness open-source projects are used to.

Long readEditorialOct 29, 2026

A Tougher AI Leaderboard Leaves More Than Open-Weight Models Behind

Artificial Analysis says its Intelligence Index v5 widened the gap between open-weight models and the leader to 12 points from 3, and critics say a test that is 60% private can't be checked. The firm's own per-test scores show that many closed models fell alongside open-weight ones on the new tests, and that more than half of DeepSeek's deficit to the top models comes from public benchmarks anyone can rerun.