Anthropic: Claude and Safety-Focused AI

Learning Objectives

• Explain Anthropic's founding story and how its safety focus shapes its products
• Compare the three Claude models — Opus 4.7, Sonnet 4.6, and Haiku 4.5 — and identify when to use each
• Understand Constitutional AI and why Anthropic's alignment approach is technically distinctive
• Explain Claude Mythos Preview, Project Glasswing, and why Anthropic chose a limited release
• Describe Claude Cowork, Managed Agents, and Claude Design as Anthropic's expansion beyond the API

Anthropic's Origin: A Safety-Motivated Departure

Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei, and a team of researchers who departed OpenAI. The stated reason: concerns about OpenAI's increasing commercial orientation and insufficient focus on AI safety research.

Dario Amodei had been VP of Research at OpenAI. Daniela Amodei had been VP of Operations. Among the co-founders: Tom Brown, the lead researcher on GPT-3; Chris Olah, one of the world's leading neural network interpretability researchers; and several others.

The founding premise was explicit: AI poses genuine risks, those risks require serious technical research, and a company specifically organized around safety could both advance AI capability and advance understanding of how to make AI safe. Anthropic describes itself as an AI safety company that also builds AI products to fund that safety research.

Funding: Anthropic is one of the best-funded AI companies in the world, backed by both major US cloud providers as strategic investors with deeply intertwined cloud-and-compute relationships:

• Amazon — $8 billion committed (doubling its initial $4 billion with an additional $4 billion in November 2024), making AWS the primary cloud partner. In early April 2026 Amazon announced an additional $5 billion + up to $20 billion further commitment, deepening the relationship.
• Google — On April 24, 2026, Google announced a major expansion of its prior ~$3 billion stake (~14% of Anthropic) to up to $40 billion: $10 billion immediate cash at a $350 billion valuation, plus $30 billion conditional on Anthropic hitting performance targets. The deal includes 5 GW of TPU compute capacity (on top of Anthropic's earlier 3.5 GW Broadcom chip deal) plus information rights, board observation, and preferred compute pricing.
• SpaceX — On May 6, 2026, Anthropic announced a compute deal giving it access to over 300 megawatts and more than 220,000 NVIDIA GPUs at SpaceX's Colossus 1 data center, deployable within one month. This is the first time SpaceX has been named as a frontier-lab compute partner and adds a third hyperscaler-grade compute supplier to the Amazon Trainium and Google TPU mix. The same announcement doubled Claude Code's five-hour rate limits across Pro, Max, Team, and Enterprise plans and removed the peak-hours throttle for Pro and Max.
• Series G — In February 2026, Anthropic closed a $30 billion Series G round at a $380 billion valuation — led by GIC and Coatue — after a $13 billion Series F in September 2025 at $183 billion.
• Series H (closed) — On May 28, 2026, Anthropic closed a $65 billion Series H at a post-money valuation of $965 billion — co-led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, with Capital Group, Coatue, D1 Capital Partners, and institutional backers Baillie Gifford, Blackstone, Brookfield, and Fidelity Management & Research also participating. Of the $65 billion raise, $15 billion came from hyperscalers including $5 billion from Amazon, and strategic semiconductor partners Samsung, SK Hynix, and Micron joined the round. The raise put Anthropic past the $852 billion that OpenAI hit in March, making Anthropic the world's most valuable AI startup. Anthropic expects its first profitable quarter ahead, with annualized revenue at $47 billion and growth of 130 percent year-over-year.

The Amazon-Google-SpaceX three-way compute structure is unusual in the industry: rather than aligning to a single hyperscaler (as OpenAI did with Microsoft), Anthropic has chosen to be the supplier-of-choice to multiple, while staying neutral on cloud infrastructure.

The self-aware tension: Anthropic acknowledges an inherent tension in its model — building frontier AI while believing frontier AI is potentially dangerous. The reasoning: if powerful AI is coming regardless, it is better to have safety-focused labs at the frontier than to cede it to less safety-focused developers.

The Claude Model Family

Claude Fable 5 — The Flagship

Claude Fable 5 is Anthropic's most capable publicly available model, released on June 9, 2026 — the public, safeguarded version of its Mythos-class frontier model. It is state-of-the-art on nearly every capability benchmark Anthropic tested, works autonomously across millions of tokens of memory, and runs safety classifiers that fall back to Opus 4.8 on sensitive cybersecurity and biology requests. Pricing is $10 input / $50 output per million tokens, with subscription plans getting free access through June 22, 2026.

Claude Opus 4.8 — The Workhorse

Claude Opus 4.8 was Anthropic's flagship from late May 2026 and, since Fable 5's June 9 launch, now sits just below it as the economical default for routine work — and the model Fable 5 falls back to. Released on May 28, 2026 — forty-one days after Opus 4.7 and coinciding with Anthropic's Series H close — it brought sharper uncertainty handling, improved agentic judgment, and a new coordination layer called Dynamic Workflows.

Key behaviors and capabilities:

• Sharper uncertainty handling — Anthropic reports Opus 4.8 is four times less likely to overlook code flaws compared to Opus 4.7, and Bridgewater's evaluation noted the model proactively raises concerns about analytical inputs and outputs rather than asserting unsupported claims
• Dynamic Workflows (research preview) — a coordination layer that lets a single Opus session orchestrate hundreds of parallel subagents for codebase-scale work; multi-hundred-thousand-line migrations from kickoff to merge with the existing test suite as the success bar
• Online-Mind2Web: 84% — leading agentic browser-use performance
• Legal Agent Benchmark — first model to break 10% on the all-pass standard
• Super-Agent benchmark — only model completing every case end-to-end
• CursorBench — exceeds prior Opus models across every effort level
• 1 million token context window — full-codebase and full-document analysis with no per-token price penalty
• Improved alignment metrics — prosocial traits up, misaligned-behavior rates down relative to Opus 4.7
• Carries forward from Opus 4.7: 3.75 megapixel vision, xhigh effort level, task budgets, exceptional coding and agentic workflows, precise instruction following

Pricing (API):

• Standard mode: $5 per million input tokens, $25 per million output tokens — unchanged from Opus 4.7, making migration a model-ID swap rather than a cost rebudget
• Fast mode: $10 per million input tokens, $50 per million output tokens — 2.5-times speed at twice the per-token cost; three times cheaper than Fast mode was on previous Claude models

Available on Claude Platform, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Model ID: claude-opus-4-8.

When to use: Long-horizon agentic work where confident-but-wrong answers compound across hours of automated reasoning, codebase-scale migrations (with Dynamic Workflows), complex reasoning that requires the model's full capability, vision-heavy tasks requiring high-resolution image understanding, tasks where quality significantly outweighs cost concerns.

Claude Opus 4.8

Anthropic

Closed

Strengths

Flagship model; sharper uncertainty handling; Dynamic Workflows for parallel subagents; 84% Online-Mind2Web; 1 million context; 3.75MP vision; xhigh effort; Fast mode

Context Window

1 million tokens

Pricing

$5/$25 per million tokens (standard); $10/$50 (Fast mode)

Anthropic →

Opus 4.7 remains available via API for teams that have not migrated, but Opus 4.8 is the new flagship.

Claude Sonnet 4.6 — The Balanced Workhorse

Claude Sonnet 4.6 is the model most professionals should use as their default — a balance of capability, speed, and cost that makes it appropriate for the vast majority of tasks.

Key characteristics:

• Full 1 million token context window (same as Opus, GA March 13, 2026)
• Strong reasoning, coding, and analysis capability — meaningfully behind Opus only on the most demanding tasks
• $3 per million input tokens, $15 per million output tokens — significantly cheaper than Opus
• Excellent for production API applications where cost scales with volume
• Powers the Claude.ai web interface and most Claude integrations

When to use: The default choice for nearly all professional use cases — writing, research, coding assistance, document analysis, content generation, API integration. Only escalate to Opus when Sonnet's output is materially insufficient for a specific task.

Claude Sonnet 4.6

Anthropic

Closed

Strengths

Balanced capability/cost; 1 million context; strong reasoning and coding; ideal for production API use

Context Window

1 million tokens

Pricing

$3/$15 per million tokens

Anthropic →

Claude Haiku 4.5 — Speed and Efficiency

Claude Haiku 4.5 is Anthropic's fastest and most cost-efficient model, designed for high-volume real-time applications where speed and cost matter more than maximum capability.

Key characteristics:

• Sub-200ms response times for typical queries
• $0.80 per million input tokens, $4 per million output tokens — the most cost-efficient Claude model
• Appropriate for: real-time chat interfaces, high-volume API calls, mobile applications, edge cases where latency matters
• Capable model at smaller scale — handles most customer service, Q&A, and classification tasks well

When to use: Building products with AI at the user-facing layer (chatbots, customer service), applications where thousands or millions of API calls per day make cost a primary concern, tasks where response speed is directly user-visible.

Claude Haiku 4.5

Anthropic

Closed

Strengths

Fastest Claude; most cost-efficient; real-time applications; high-volume production use

Context Window

200K tokens

Pricing

$0.80/$4 per million tokens

Anthropic →

Claude Mythos Preview: The Cybersecurity Breakthrough (April 2026)

On April 7, 2026, Anthropic announced Claude Mythos Preview — its most powerful model ever built and what the company calls "a step change" above even the latest Opus. Mythos leads 17 of 18 benchmarks Anthropic measured, including 93.9% on SWE-bench Verified (vs. Opus 4.7's 87.6%) and 82% on Terminal-Bench 2.0 (vs. 65.4%). In June 2026, the Mythos line reached the public: Claude Fable 5 ships the safeguarded model to everyone, while Claude Mythos 5 keeps the unrestricted version for vetted Project Glasswing and biomedical partners.

Why it is not publicly available: Anthropic determined that Mythos's cybersecurity capabilities were too dangerous for a general release. The model can autonomously discover and chain zero-day exploits across major operating systems and browsers — finding thousands of vulnerabilities during testing, including a 27-year-old bug in OpenBSD and a 16-year-old bug in FFmpeg.

Instead of a public release, Anthropic launched Project Glasswing — a $100 million cybersecurity initiative providing Mythos Preview access to approximately 50 organizations (including Amazon, Apple, Microsoft, Google, CrowdStrike, and the Linux Foundation) exclusively for defensive security work.

Beyond the API: Claude Cowork, Managed Agents, and Claude Design

Claude Cowork (GA February 2026)

Claude Cowork became generally available on February 24, 2026 — a desktop agent for non-technical knowledge work. Unlike the Claude API and claude.ai chat interface, Cowork runs on your desktop and can autonomously navigate local files, folders, and applications.

Key capabilities:

• Local file access: Read, create, and edit files on your computer autonomously
• Cross-application workflows: Move between local applications and synthesize information across sources
• Enterprise plugins: Integrations with Google Drive, Gmail, DocuSign, FactSet, and more
• Role-based access controls: SCIM identity management for enterprise deployment

Anthropic describes Cowork as "Claude Code for the rest of your work" — bringing the agentic desktop experience to non-developers.

Claude Managed Agents (Public Beta)

Claude Managed Agents is a public beta offering composable APIs for building and deploying cloud-hosted agents at scale. Where Claude Code and Cowork are desktop agents, Managed Agents allows organizations to deploy persistent, cloud-based AI agents through Anthropic's infrastructure — handling orchestration, scaling, and monitoring.

Claude Design (Research Preview, April 2026)

Claude Design launched on April 17, 2026 as a new Anthropic Labs product for AI-powered visual creation. It generates prototypes, presentation slides, one-pagers, and UI mockups from natural language — competing with Figma, Canva, and Gamma for rapid ideation workflows.

The distinctive feature is automatic brand system integration: during onboarding, Claude Design reads your team's codebase and design files to learn your typography, color palette, spacing tokens, and component patterns. Every subsequent project applies these brand guidelines, so output looks like it came from your design team rather than a generic AI.

Built on Claude Opus 4.7, Claude Design inherits the model's 3.75-megapixel vision capability for reading reference materials and screenshots accurately. Finished projects export to Canva (fully editable), PDF, PPTX, HTML, standalone URLs, or Claude Code for development handoff.

Availability: Research preview for Claude Pro, Max, Team, and Enterprise subscribers (not free tier).

Together, Cowork, Managed Agents, and Claude Design represent Anthropic's expansion from a pure API company into consumer and enterprise product surfaces — each targeting a different type of knowledge work.

Claude's Distinctive Characteristics

What Claude Does Well

Following complex instructions: Claude models have a reputation for accurately executing multi-part, nuanced instructions. For enterprise applications where output format, tone, and constraints are precisely specified, this consistency is valuable.

Thoughtful and calibrated responses: Claude models tend to acknowledge uncertainty, avoid overconfident claims, and flag when a question is outside their knowledge. For professional contexts where accuracy matters, this calibration is a feature.

Coding and agentic tasks: Claude's coding capability is particularly strong for complex, multi-file, long-horizon tasks — the kind that go beyond simple autocompletion. Claude Code (the CLI and API product) is designed around Opus 4.7's agentic capabilities, with the new xhigh effort level as the default.

Long context: The 1 million token context window (GA March 2026) matches GPT-5.5 and Gemini 3 Pro, enabling massive document analysis and code review tasks that would require chunking with smaller context models. Claude scores highest among frontier models on MRCR v2 retrieval accuracy at this context length.

What Claude Does Not Do

Image generation: Anthropic does not offer an image generation product. If you need AI image generation, use GPT Image 2 (OpenAI's April 2026 flagship, branded ChatGPT Images 2.0), Gemini's Nano Banana Pro, or Midjourney.

Real-time web search (default): Standard Claude API calls do not include real-time web search. Claude has a knowledge cutoff date. For current information, use Claude with web search integration or use Perplexity.

Anthropic's Research Focus

Beyond its products, Anthropic is doing research that the broader field uses:

• Interpretability research: Chris Olah's team has made significant progress on understanding what transformer circuits are actually computing — one of the most important research directions in AI safety
• Constitutional AI: Published and influential technique for training value-aligned models
• Evals and safety testing: Anthropic publishes research on how to evaluate AI safety properties and runs extensive red-teaming before model releases
• Claude Character: Anthropic publishes the principles that shape Claude's character and behavior — an unusual level of transparency for a commercial AI lab

Key Takeaways

• Anthropic was founded in 2021 by former OpenAI researchers with a specific safety-focused mission, backed by Amazon (committed over $13 billion across rounds), Google (committed up to $40 billion in cash + TPU compute, announced April 24, 2026), and SpaceX (300 megawatts and 220,000 NVIDIA GPUs at Colossus 1, announced May 6, 2026); closed a $65 billion Series H on May 28, 2026, at a post-money valuation of $965 billion, making Anthropic the world's most valuable AI startup ahead of OpenAI, with annualized revenue at $47 billion and growth of 130 percent year-over-year
• Constitutional AI (CAI) is Anthropic's distinctive alignment technique — training models to critique their own outputs against a set of principles
• Claude Opus 4.8 is the flagship (released May 28, 2026, sharper uncertainty handling, Dynamic Workflows for parallel subagents, 1 million context, 3.75MP vision, $5/$25 per million tokens standard, $10/$50 Fast mode) — use for complex, high-stakes long-horizon agentic work where confident-but-wrong answers compound
• Claude Sonnet 4.6 is the default professional choice (1 million context, $3/$15 per million tokens) — best balance of capability and cost
• Claude Haiku 4.5 is for high-volume, real-time applications (fastest, $0.80/$4 per million tokens)
• The Mythos-class model reached the public in June 2026 as Claude Fable 5 (safeguarded, the new flagship) and Claude Mythos 5 (unrestricted, invitation-only through Project Glasswing — a $100 million cybersecurity defense initiative with approximately 50 partner organizations); Mythos Preview (April 2026) was the original limited release
• Claude Cowork (GA February 2026) is a desktop agent for non-technical knowledge work; Claude Managed Agents enables cloud-hosted agent deployment at scale; Claude Design (research preview, April 2026) brings AI-powered visual creation with automatic brand system integration
• Anthropic does not offer image generation; Claude has stronger content moderation defaults than competing models