Anthropic company dossier 059
Anthropic: Company, Models, Products and Ecosystem
Company, research, models, products, developer infrastructure, distribution, compute, capital, governance and safety—and how the parts fit together.· information current to
01
Executive summary
Anthropic is a private AI research and product company whose Claude models power consumer applications, workplace agents, coding tools and an enterprise developer platform. It is a Delaware public benefit corporation governed by a conventional board and an independent Long-Term Benefit Trust with defined board-election rights.[1][2]
The operating chain is straightforward: external compute trains and serves Claude; research produces models and safety methods; Anthropic packages those models into direct products and the Claude Platform; cloud and services partners distribute them; customer usage funds further research, products and infrastructure. These layers reinforce one another, but they are not interchangeable.
Corporate formPBC[D] Private public benefit corporation
Latest disclosed valuation$965bn[D] Series H post-money, May 2026 [3]
Disclosed revenue measure$47bn+[CR] Annualised run-rate, May 2026 [3]
Current frontier tiers5[D] Fable, Mythos, Opus, Sonnet, Haiku
Evidence markers. [D] Primary-source disclosed; [CR] company-reported metric, not independently audited; [CP] counterparty-disclosed; [C] calculated from cited facts; [I] inferred by this article; [ND] not publicly disclosed. Markers appear on decision-relevant claims rather than every sentence.
Scope. Prices, promotional claims and unnormalised benchmark rankings are deliberately excluded. A benchmark belongs here only when its task, version, settings, tool access, scoring and comparison conditions are sufficiently comparable; none is required to explain Anthropic’s system.
Essential questions
| Question | Concise answer |
|---|---|
| What exactly is Anthropic? | A private AI research and product company organised as a Delaware public benefit corporation. It develops Claude models, operates direct products and supplies a developer platform; Section 02 maps the complete system. |
| Who owns and controls it? | Investors own economic interests, but ownership, board governance and operational control are not identical. The conventional board governs the company while the Long-Term Benefit Trust holds defined board-election rights; Sections 03 and 10 separate these roles. |
| Do Amazon, Google or other investors control Anthropic? | Published investments establish financing and commercial relationships, not unilateral operating control. Each relationship must be read through its investment terms, cloud commitments and governance rights rather than inferred from headline capital alone. |
| How does Anthropic make money? | Through paid Claude applications, API consumption, enterprise agreements and distribution through cloud partners. Reported annualised run-rate is a measure of recent revenue pace, not profit, cash flow or audited full-year revenue; Section 09 explains the commercial model. |
| What is the difference between a Claude model and a Claude product? | A model provides underlying reasoning and generation capability. Claude.ai, Claude Code, Cowork, the API and partner services package models with interfaces, tools, context, permissions and operating infrastructure; Sections 05–07 separate these layers. |
| Why are Opus, Sonnet, Haiku, Fable and Mythos separate tiers? | They represent different capability, latency, access and deployment positions within the model portfolio. A tier name does not by itself establish identical context, tool, safety or availability characteristics across every product surface. |
| Where does Claude Code fit? | Claude Code is Anthropic's agentic software-development system, not a standalone foundation model. It combines Claude models with repository context, tools, execution, permissions and extensions; the companion Claude Code dossier provides the detailed architecture. |
| Is MCP an Anthropic-only technology? | No. Anthropic initiated the Model Context Protocol, but MCP is an open protocol used by multiple hosts, clients and servers. It connects systems; it is not a Claude model, a Claude Code feature set or an Anthropic-controlled application. |
| Who supplies Anthropic's computing infrastructure? | Anthropic uses a multi-provider infrastructure strategy involving Amazon Web Services, Google Cloud and specialised hardware relationships. Contracted capacity, installed capacity and usable production capacity are different measures; Section 08 explains the dependency chain. |
| Who are the principal customers? | Individuals, developers, enterprises, public-sector organisations and cloud customers access Claude through different channels. A cloud provider may simultaneously be an investor, infrastructure supplier, distributor and customer channel; Section 10 maps the roles. |
| How quickly is the product portfolio developing? | The model and product timeline shows unusually short intervals between major releases, including five Opus generations within 242 days. Release pace demonstrates execution cadence, but not uniform adoption, durability or economics; Section 04 provides the chronology. |
| What most constrains the roadmap? | Compute availability, energy and data-centre delivery, model reliability, safety controls, enterprise deployment requirements and distribution economics. Section 13 distinguishes released capability from published direction and the gates required to reach it. |
02
The Anthropic system and its seven objects
| Object | What it is | Do not confuse it with | Evidence basis |
|---|---|---|---|
| Company | Anthropic PBC: the legal, employing and contracting organisation. | Claude, a model or product brand. | Primary-source disclosed [1][2] |
| Research | Training, alignment, interpretability, evaluation and safety work. | A released model or customer product. | Primary-source disclosed [26][27] |
| Models | Claude model snapshots such as Fable, Opus, Sonnet and Haiku. | Chat, Code, Cowork or an access plan. | Primary-source disclosed [17–19][26] |
| Products and API | Interfaces, agent loops, tools and controls that package models for work. | The underlying model alone. | Primary-source disclosed where linked; maturity varies |
| Distribution and infrastructure | Cloud routes, accelerators, data centres, memory, power and service partners. | Anthropic ownership or operating control. | Primary-source agreements; delivery state separate [21–24] |
| Capital and governance | Economic ownership, financing, board authority and Trust rights. | Supplier, customer or distributor roles. | Primary-source disclosed; complete cap table not public [1–3] |
| Safety and policy | Principles, training methods, evaluations, thresholds, safeguards and escalation. | Proof that a system is safe. | Policy and company-reported evaluation evidence [25][27] |
- Compute → models.Training and inference depend on AWS Trainium, Google TPUs and NVIDIA GPUs, plus data centres, networking, power and memory.
- Models → products.The same Claude families can appear in Chat, Cowork, Claude Code, the API and partner clouds; a model release is not a separate customer product.
- Distribution → demand feedback.Direct subscriptions, API usage, cloud marketplaces and enterprise deployments reveal demand and finance further compute, research and product work.
Controlled ontology. Claude is the model and product brand; Anthropic is the company. AWS, Google Cloud and Microsoft can distribute Claude while also supplying infrastructure or capital. Overlapping roles do not make the objects equivalent.
03
From research laboratory to product system
The timeline tracks changes in the company’s operating system, not every feature release. It shows how models, applications, developer infrastructure and distribution have advanced together.
| Date | Model or company milestone | Product or distribution milestone | Why it matters |
|---|---|---|---|
| 2021 [1] | Anthropic established by former OpenAI researchers and operators. | Research-led company formed as a public benefit corporation. | Safety research and frontier-model development were joined from the beginning. |
| 14 Mar 2023 [4] | Claude introduced. | Chat interface and developer API opened beyond a closed partner group. | Research became a general-purpose commercial service. |
| 11 Jul 2023 [5] | Claude 2 released. | Claude.ai public beta launched in the US and UK. | Anthropic added a direct consumer surface alongside the API. |
| 4 Mar 2024 [6] | Claude 3 introduced Opus, Sonnet and Haiku tiers. | Claude API expanded internationally; cloud availability widened. | One model became a tiered portfolio balancing capability and speed. |
| 20 Jun–22 Oct 2024 [7][8] | Claude 3.5 Sonnet, an updated Sonnet and Haiku 3.5 arrived. | Artifacts and public-beta computer use extended Claude beyond text chat. | Product capability began expanding as quickly as base-model capability. |
| 25 Nov 2024 [9] | Model Context Protocol open-sourced. | Claude Desktop gained a standard path to tools and external data. | Anthropic began building an ecosystem layer not confined to its own applications. |
| 24 Feb–22 May 2025 [10][11] | Sonnet 3.7 introduced hybrid reasoning; Opus 4 and Sonnet 4 followed. | Claude Code moved from research preview to general availability; agent tools expanded on the API. | Reasoning, tool use and software agents became a coordinated product direction. |
| Sep–Nov 2025 [12][13] | Sonnet 4.5, Haiku 4.5 and Opus 4.5 shipped. | Agent SDK, native IDE work and Microsoft distribution expanded. | The model portfolio and agent infrastructure started shipping as one stack. |
| Feb–May 2026 [14–16] | Opus 4.6, Sonnet 4.6, Opus 4.7 and Opus 4.8 shipped. | Labs, Cowork, managed agents, workplace add-ins, Design and industry agents expanded the product surface. | Anthropic moved from chat and coding into broader professional workflows. |
| Jun–Jul 2026 [17–19] | Fable 5, Mythos 5, Sonnet 5 and Opus 5 shipped. | Claude Tag and Claude Science extended team and specialist work. | General and restricted model tiers now support multiple agent products. |
Two clocks are running. Model releases improve the reasoning engine. Product releases determine where that engine works, what context and tools it can reach, who controls it and how Anthropic earns revenue.
04
The pace of model development
[C] Five Opus generations arrived within 242 days from Opus 4.5 to Opus 5. The intervals shortened to as little as 42 days, calculated from Anthropic’s release dates, while Sonnet, Haiku and the Fable/Mythos tier evolved in parallel.
The important change is not a benchmark score. Each release progressively emphasised longer agentic work, configurable effort, context management, tool use, computer use, verification and professional artefact production. Release pace therefore affects both customers and Anthropic’s own product architecture.
| Opus release | Date | Days from previous Opus | Principal published change | Product implication |
|---|---|---|---|---|
| Opus 4.5 [13] | 24 Nov 2025 | — | More efficient agentic coding, effort control and longer-running work. | Broader Opus use across Claude products and the API. |
| Opus 4.6 [14] | 5 Feb 2026 | 73 | One-million-token context beta, adaptive thinking and stronger long-horizon execution. | Agent teams, compaction and richer professional work. |
| Opus 4.7 [15] | 16 Apr 2026 | 70 | More reliable difficult coding, stronger vision and stricter instruction following. | More work could be delegated with less close supervision. |
| Opus 4.8 [16] | 28 May 2026 | 42 | Improved computer use and collaboration; dynamic workflows and faster serving options. | Longer product workflows and browser-agent work became more practical. |
| Opus 5 [19] | 24 Jul 2026 | 57 | Greater verification, long-running agency and professional work. | Default Max-tier model and a general-access alternative below Fable. |
Cadence is not a promise. These are historical release intervals, not a forecast of the next model’s name or date. Availability, safeguards and migration behaviour can change independently of capability.
05
The current Claude model portfolio
Model state as of 24 August 2026. Model names identify capability and deployment tiers; product names identify the interfaces and controls that use them. Specifications come from Anthropic’s current model documentation.[26]
| Model | Release | Claude API ID | Access status | Deployment position |
|---|---|---|---|---|
| Claude Fable 5 | 9 Jun 2026 | claude-fable-5 | Generally available | Highest-capability widely released tier; safeguarded general use [17][26] |
| Claude Mythos 5 | 9 Jun 2026 | claude-mythos-5 | Invitation-only | Fable specifications with restricted Project Glasswing access [17][26] |
| Claude Opus 5 | 24 Jul 2026 | claude-opus-5 | Generally available | Complex agentic coding and enterprise work [19][26] |
| Claude Sonnet 5 | 30 Jun 2026 | claude-sonnet-5 | Generally available | Faster general agentic tier [18][26] |
| Claude Haiku 4.5 | 15 Oct 2025 | claude-haiku-4-5-20251001 | Generally available | Lowest-latency current tier [26] |
| Model | Context | Maximum output | Input → output | Thinking mode | Reliable knowledge cutoff |
|---|---|---|---|---|---|
| Fable 5 | 1m tokens | 128k tokens | Text/image → text | Adaptive; always on | Jan 2026 |
| Mythos 5 | 1m tokens | 128k tokens | Text/image → text | Shares Fable specifications | Jan 2026 |
| Opus 5 | 1m tokens | 128k tokens | Text/image → text | Adaptive | May 2026 |
| Sonnet 5 | 1m tokens | 128k tokens | Text/image → text | Adaptive | Jan 2026 |
| Haiku 4.5 | 200k tokens | 64k tokens | Text/image → text | Extended, not adaptive | Feb 2025 |
[ND] Published technical boundary. Anthropic publishes model IDs, modalities, token limits, thinking support and knowledge cutoffs. Parameter counts, expert or MoE structure, training-token volume, corpus composition, detailed compute, complete post-training recipe and internal architecture are not publicly disclosed.
Fable and Mythos are deployment variants, not unrelated foundations. Anthropic states that they share the same underlying model and published specifications but have different access and safeguard boundaries. Dateless Claude 4.6-and-later API IDs are pinned snapshots, not evergreen pointers.[26]
06
Products people and teams use directly
Anthropic sells more than model access. Its direct products package models with interfaces, context, tools, permissions, collaboration, connectors and administrative controls for particular forms of work.
Claude Code is strategically important because its agent loop and extensibility influence other products, but it does not contain the whole portfolio. Chat, Cowork, Tag, specialist workbenches and workplace integrations serve different users and operating environments.
| Object class and product | Primary work | Execution boundary | Primary customer | Status and evidence |
|---|---|---|---|---|
| Application · Claude Chat | Conversation, research, analysis and artefact creation. | Web, desktop and mobile Claude applications. | Individuals, teams and enterprises. | Operational · primary-source disclosed |
| Agent product · Claude Cowork | Multi-step knowledge work across local files, applications and connected systems. | Desktop agent with approvals and governed connectors. | Professional and enterprise users. | Operational; expanding · company-disclosed [20] |
| Agent product · Claude Code | Repository analysis, changes, terminal work, testing and delivery automation. | Terminal, IDE, web/cloud and automation environments. | Developers and engineering organisations. | Operational · primary-source disclosed [10][11] |
| Integration · Claude Tag | Delegated work inside team conversations. | Slack channels with selected tools, data and codebases. | Team and Enterprise organisations. | Beta · company-disclosed |
| Integration · Microsoft 365 add-ins | Document, spreadsheet, presentation and communication work. | Microsoft productivity applications. | Professionals and enterprises. | Staged rollout · company-disclosed |
| Specialist products | Scientific, design and industry workflows. | Science, Design, plugins and managed-agent cookbooks. | Specialist professional teams. | Mixed maturity · company-disclosed [20] |
Product does not equal model. Products add the workflow, context, tools and control plane around Claude. They can change without a new base model, and a new model can enter an existing product without changing that product’s purpose.
07
Claude Platform and developer infrastructure
The platform serves organisations that build their own applications and agents. Some components are hosted services; others are protocols, SDKs or capability packages used across Anthropic and third-party products.
| Layer | What it provides | Who operates it | Relationship to Claude Code |
|---|---|---|---|
| Claude API | Model inference, context, effort controls, streaming and usage metering. | Anthropic or supported cloud platform. | Claude Code consumes model inference but is a separate product. |
| Native tools | Tool use, code execution, web access, files, computer use, prompt caching and related primitives. | Anthropic plus the customer’s tool host where applicable. | Claude Code packages selected tools into its agent loop. |
| Model Context Protocol | Open client-server standard connecting AI applications to tools, resources and prompts. | Independent hosts and servers across the ecosystem. | Claude Code is one MCP host/client; MCP is not Claude Code-only. |
| Claude Agent SDK | The Claude Code agent loop exposed for custom Python and TypeScript agents. | Customer or Anthropic-managed environment. | Derived from the same agent infrastructure. |
| Skills, plugins and connectors | Reusable instructions, domain knowledge and governed access to external systems. | Anthropic, customers and ecosystem providers. | Extend Claude Code and other Claude products. |
| Claude Managed Agents | Hosted long-running agents with permissions, credential handling and auditability. | Anthropic-hosted control and runtime. | Applies agent infrastructure beyond interactive coding. |
Open protocol boundary. Anthropic created MCP and later donated it to vendor-neutral governance. MCP servers, clients and the Inspector are ecosystem infrastructure; their existence should not be treated as ownership by Anthropic.
08
Compute, cloud and distribution infrastructure
Anthropic does not manufacture accelerators or own the clouds through which Claude is distributed. It contracts across Trainium, TPU and NVIDIA GPU systems and sells Claude directly and through three major cloud platforms. This diversifies hardware and distribution while preserving substantial hyperscaler dependence.
| Organisation | Operating role | What it contributes | Transaction classification | Evidence boundary |
|---|---|---|---|---|
| Amazon / AWS | Primary cloud, training partner, investor and distributor | Trainium, Rainier, Bedrock; up to 5 GW. | Equity investment · compute agreement · cloud distribution [21] | Capacity is staged; investment does not establish control. |
| Google / Broadcom | Compute supplier and cloud distributor | TPUs, custom silicon, Vertex AI; 5 GW planned. | Compute agreement · cloud distribution [22] | New capacity expected from 2027; not operational today. |
| Microsoft / NVIDIA | Cloud, accelerator, investment and distribution partners | Azure/Foundry, GPUs, optimisation and contracted compute. | Investment commitments · compute purchase · cloud distribution [23] | The three transaction types are separate. |
| SpaceX | Compute supplier | Colossus capacity; over 300 MW and 220,000 GPUs disclosed. | Commercial compute agreement · capacity onboarding [24] | No ownership or distribution relationship disclosed. |
| Micron, Samsung and SK hynix | Infrastructure partners | Memory, storage and logic relationships. | Strategic supply relationships [3] | Volumes, terms and exclusivity are not disclosed. |
Controlled vocabulary. Strategic agreement, definitive compute agreement, operational capacity, capacity onboarding, future capacity, operational distribution and strategic investment describe different states. Contracted gigawatts are not the same as installed, energised and customer-serving capacity.
Dependency is not control
| Partner | Capital relationship | Governance evidence | Compute dependency | Distribution dependency |
|---|---|---|---|---|
| Amazon / AWS | [D] Equity invested and committed | [ND] No unilateral control disclosed | [D] Primary cloud and training partner | [D] Operational through Bedrock |
| Google / Broadcom | [ND] No capital amount classified in this agreement | [ND] No control rights disclosed | [D] Existing TPU use; future 5 GW | [D] Operational through Vertex AI |
| Microsoft / NVIDIA | [D] Investment commitments disclosed | [ND] No unilateral control disclosed | [D] Contracted Azure/NVIDIA capacity | [D] Operational through Foundry |
| SpaceX | [ND] No equity relationship disclosed | [ND] None disclosed | [D] Colossus capacity onboarding | [ND] None disclosed |
Analytical principle. Equity exposure, governance influence, compute dependence and distribution dependence are separate dimensions. A company can be operationally important without controlling Anthropic.
09
Who uses Claude, who pays and who distributes it
An end user, subscription customer, API customer, cloud customer, services partner and distributor can all touch the same Claude workload. The table identifies the contractual and operating role rather than describing every organisation as a generic partner.
| Customer or organisation | Relationship type | What it uses or provides | Who pays Anthropic | Evidence status |
|---|---|---|---|---|
| Individuals and professional users | End user and direct subscription customer | Claude Chat, Cowork, Code and specialist products. | Individual or employer under a direct plan or usage credits. | Operational at scale |
| Developers and software companies | Direct API or Claude Code customer | Models, tools and agent infrastructure embedded in products or engineering work. | Developer, startup or enterprise account. | Operational at scale |
| Netflix, Spotify, KPMG, L’Oréal and Salesforce | Named Claude Code customers | Software-development workflows. | Enterprise customer under undisclosed terms. | Customer use disclosed; contract economics undisclosed |
| Pfizer, Intuit and the European Parliament | Cloud-mediated customer | Claude applications deployed through Amazon Bedrock. | Customer pays through AWS commercial arrangements. | Named production examples |
| Accenture, PwC and Cognizant | Customer and services/distribution partner | Internal deployments, trained workforces and client implementation services. | Each firm may buy Claude while its clients fund separate implementations. | Strategic agreements · Operational deployments |
| AWS, Google Cloud and Microsoft | Cloud distributor and infrastructure provider | Claude access within Bedrock, Vertex AI and Foundry. | Cloud customer pays its cloud provider; Anthropic’s revenue share is undisclosed. | Operational distribution |
| Project Glasswing organisations | Vetted user or trusted-access customer | Restricted Mythos-class cybersecurity capability. | Commercial and programme terms vary and are not fully disclosed. | Restricted deployment |
Business-model takeaway. Direct products give Anthropic a customer relationship. The API gives developers building blocks. Cloud marketplaces reduce procurement friction. Services partners supply integration capacity. None of those channels by itself proves customer retention, profitability or contract value.
10
Capital, ownership and governance
Funding, board and Trust state as of 24 August 2026. Financing, economic ownership, board authority, Trust rights and operating dependence are separate dimensions.
| Category | [D] Current disclosed position | What it means | [ND] What is not disclosed |
|---|---|---|---|
| Corporate entity | Private Delaware public benefit corporation. [1][2] | Directors may balance shareholder interests with the stated public benefit. | Complete ownership register and current capitalisation table. |
| Executive management | Dario Amodei, co-founder and chief executive; Daniela Amodei, co-founder and president. [1] | Management runs the company under board oversight. | Detailed internal decision rights. |
| Board authority | Dario Amodei, Daniela Amodei, Yasmin Razavi, Reed Hastings, Chris Liddell and Vas Narasimhan. [1] | Directors provide corporate oversight under the governing documents. | Individual director voting power and reserved matters. |
| Trust authority | Neil Buddy Shah, Richard Fontaine and Ben Bernanke; Class T board-election rights. [1][2] | The financially disinterested Trust can elect and remove part of the board. | Non-public governing-document detail and future exercise of rights. |
| Equity capital | $65bn Series H at a $965bn post-money valuation; led by Altimeter, Dragoneer, Greenoaks and Sequoia. [3] | Company-reported financing for compute, research and product expansion. | Investor allocations, ownership percentages and control rights. |
| Strategic capital | Amazon, Microsoft and NVIDIA have disclosed investment or investment commitments alongside infrastructure relationships. | Capital and supply/distribution agreements can reinforce one another. | No evidence that any one strategic investor controls Anthropic. |
[D] Financing classification. The $65bn Series H is the announced round size; $965bn is its post-money valuation. Anthropic says the round includes $15bn of previously committed hyperscaler investment, including $5bn from Amazon. An investment commitment is not automatically funded equity, and none of these figures represents compute spending.[3]
Investment is not control. The Series H valuation and named investors establish financing facts, not the full ownership structure. The Trust holds governance rights but is designed to be financially disinterested.
11
Economics: can product growth outrun compute commitments?
Anthropic discloses selected run-rate, customer and infrastructure figures but not complete audited financial statements. The useful analysis therefore separates current revenue evidence, customer commitments, infrastructure commitments and future potential.
| Evidence layer | Published fact | What it establishes | What it does not establish |
|---|---|---|---|
| [CR] Company-reported metric | Annualised revenue run-rate exceeded $47bn in May 2026. [3] | Substantial recent revenue scale. | [ND] Audited revenue, margin, profit or cash flow. |
| [CR] Company-reported metric | Claude Code run-rate exceeded $2.5bn in February 2026. [28] | Coding had become a material direct product. | [ND] Product profit, retention or channel mix. |
| [CR] Company-reported metric | Over 1,000 customers exceeded $1m annualised spend in April 2026. [22] | Broader large-enterprise demand. | [ND] Contract duration, concentration, churn or realised revenue. |
| Infrastructure contracted | Large multi-year commitments span AWS, Azure, Google/Broadcom, SpaceX and US data-centre capacity. | Anthropic has reserved a diversified path to scale. | Utilisation, unit cost, depreciation, take-or-pay exposure or timing of all capacity. |
| Future commercial potential | Agent products are extending Claude from chat and code into broader professional workflows. | More addressable work and potential product revenue. | Adoption, defensibility, margins or returns on committed infrastructure. |
| Revenue path | Contractual customer | Commercial unit | Principal cost or dependency |
|---|---|---|---|
| Consumer and professional subscriptions | Individual or employer | Plan, seat and usage credits | Inference capacity and product operation |
| Enterprise applications | Organisation | Seats, usage and negotiated controls | Support, compliance, connectors and inference |
| Claude Code | Developer, team or enterprise | Subscription or usage | High-context agentic inference and execution infrastructure |
| Claude Platform | Developer or enterprise account | Model tokens, tools and hosted-agent usage | Training, inference, platform reliability and safety systems |
| Cloud distribution | Cloud customer through AWS, Google or Microsoft | Cloud marketplace consumption | Provider economics and commercial revenue-sharing terms |
Run-rate is not profit. Rapid revenue growth can coexist with negative cash flow when compute, research, infrastructure and product investment are expanding. Anthropic does not publish enough information to calculate current profitability or unit economics reliably.
12
Safety governance is part of the product architecture
[D] Safety-policy state as of 24 August 2026. The current Responsible Scaling Policy is version 3.4, effective 8 July 2026; Anthropic last updated the policy page on 14 August 2026.[25]
Anthropic’s safety system operates at company, model and request level. The public benefit charter and Long-Term Benefit Trust affect corporate oversight; the Responsible Scaling Policy links capability thresholds to safeguards; system cards document model evaluations; deployment classifiers and trusted-access programmes control particular uses.
This means “released” is not a single state. A model can be trained, evaluated, generally available with classifiers, restricted to approved organisations, temporarily suspended, redeployed or automatically routed to another model.
| Lifecycle stage | Policy or method | Trigger | Operational consequence | Evidence boundary |
|---|---|---|---|---|
| 1 · Principles | Claude’s constitution; Constitutional AI. | Model-training and evaluation design. | Principles shape critique, preference and response behaviour. | Method disclosed; current model implementation not disclosed [27] |
| 2 · Training | Supervised learning, AI feedback and post-training. | Creation of a deployable model. | Baseline instruction-following and refusal behaviour. | Research disclosed; model-specific recipe not disclosed [27] |
| 3 · Evaluation | Capability, alignment, safeguard and security assessments. | Candidate model approaches a risk threshold. | Classification, added testing or revised deployment configuration. | Company-reported; coverage and reproducibility vary |
| 4 · Thresholds | RSP v3.4; Frontier Safety Roadmaps. | Specified dangerous capability threshold is reached. | Required safeguards, risk report and deployment constraints. | Effective 8 Jul; page updated 14 Aug 2026 [25] |
| 5 · Safeguards | Classifiers, access controls, fallbacks and restricted programmes. | User, request or model enters a controlled risk category. | Block, reroute, restrict access or require verified participation. | Fable/Mythos states disclosed [17] |
| 6 · Monitoring | Adversarial testing, incident response and risk reports. | New behaviour, attack or control weakness appears. | Safeguard, access or deployment changes after release. | Processes disclosed; complete incident data not public [25] |
| 7 · Escalation | Management, board, LTBT and external authority. | Governance, policy or legal intervention is required. | External review, changed access, deployment limits or pause. | Governance disclosed; future decisions unknown [1][2] |
Safety policy is not proof of safety. It is a governance and control system. Its effectiveness depends on evaluations, implementation, threat coverage, external review, enforcement and the pace at which model capabilities change.
13
Roadmap and dependencies
Anthropic does not publish a fixed model-release calendar. The defensible roadmap is therefore built from announced capacity, operating product direction and explicit safety gates rather than invented model names or dates.
| Workstream | Current state | Published next stage | Required gate | What it enables |
|---|---|---|---|---|
| Claude models | Fable 5, Mythos 5, Opus 5, Sonnet 5 and Haiku 4.5 form the current portfolio. | No fixed public release schedule. | Training progress, evaluations, safeguards and serving readiness. | Higher-capability or more efficient products. |
| Agent products | Chat, Cowork, Code, Tag, workplace integrations and specialist products operate at different maturity levels. | More proactive, collaborative and domain-specific work surfaces. | Reliability, permissions, connectors, auditability and customer adoption. | Expansion from answers into completed workflows. |
| Claude Platform | API, tools, MCP, Agent SDK and managed-agent services. | Longer-running and more composable agents across clouds. | Context management, tool security, credentials, observability and uptime. | Customers can build production agents rather than isolated prompts. |
| AWS capacity | Operational Trainium fleet and Project Rainier; capacity expanding. | Nearly 1 GW of Trainium2/3 by end-2026; up to 5 GW under the wider agreement. | Chip delivery, data centres, power, networking and commissioning. | Higher direct and Bedrock serving capacity. |
| Google/Broadcom capacity | Existing TPU usage and Vertex AI distribution. | 5 GW of next-generation capacity expected from 2027. | Silicon delivery, sites, power and software optimisation. | Additional training and inference diversity. |
| Safety system | RSP v3.4, ASL controls, system cards, classifiers and trusted access. | Published Frontier Safety Roadmaps and iterative safeguards. | Capability coverage, external scrutiny and operational effectiveness. | Wider access without treating every capability as general access. |
Roadmap discipline. Contracted capacity must still be delivered, energised, networked and made reliable. Product announcements must still become adopted workflows. Model capability must still pass the applicable deployment and safety gates.
Analytical conclusions
- [I] Models are one layer, not the whole business.Observation: Claude appears across several products and clouds. Mechanism: interfaces, agents, distribution and compute convert model capability into work. Implication: product execution matters alongside model quality.
- [I] Release cadence is an advantage and a burden.Observation: five Opus generations arrived within 242 days. Mechanism: replacement propagates through products, evaluations and customer systems. Implication: faster improvement also increases migration and governance work.
- [I] Infrastructure diversity does not remove dependence.Observation: Anthropic contracts across several accelerator and cloud providers. Mechanism: capacity must be delivered, energised, networked and utilised. Implication: diversification reduces single-provider exposure but adds execution gates.
- [I] Investment can reinforce supply without proving control.Observation: some partners supply capital, compute and distribution. Mechanism: these are separate contracts and rights. Implication: economic exposure and operational dependence can be material even when unilateral control is not disclosed.
- [ND] The decisive unknowns remain financial and architectural.Observation: profitability, unit economics, cap table, utilisation and detailed model architecture are not public. Mechanism: private-company disclosures cover selected metrics. Implication: precise returns, control and training economics cannot be calculated reliably.
14
Key risks and unresolved questions
- Private-company disclosure. Run-rate revenue, valuations and selected customer metrics do not provide a complete income statement, balance sheet, cash-flow statement or capitalisation table.
- Compute commitments. Multi-gigawatt and multi-year infrastructure agreements can secure supply but create utilisation, financing, power, delivery and concentration risks.
- Supplier overlap. Amazon, Google, Microsoft, NVIDIA and other firms can be investors, suppliers and distributors in different combinations, complicating incentives and bargaining power.
- Release cadence. Fast model replacement can improve products while increasing migration, evaluation, safety and customer-operating burdens.
- Safety execution. Capability evaluations and classifiers can miss risks, block legitimate work or lag behind newly discovered behaviours.
- Regulatory and geopolitical exposure. Export controls, national-security policy, data rules and AI regulation can change access by model, user, sector or jurisdiction.
- Product sprawl. Chat, Code, Cowork, Tag, specialist products and platform services must share infrastructure without confusing customers or duplicating development.
- Competitive pressure. Model capability can diffuse quickly across competitors, placing more weight on product execution, distribution, reliability and economics.
- Security and misuse. More capable models, tools and computer access enlarge both beneficial automation and cyber, biological, prompt-injection and data-exfiltration risk.
- Customer evidence. Named deployments and annualised spending establish use, not long-term retention, contract value, workload quality or profitability.
15
Evidence ledger and primary sources
- Anthropic company and governanceCurrent purpose, operating functions, board and Long-Term Benefit Trust trustees; accessed 24 August 2026.
- The Long-Term Benefit Trust19 September 2023; PBC, Class T governance rights and independence design.
- Anthropic Series H financing28 May 2026; valuation, investors, company-reported run-rate and infrastructure relationships.
- Introducing Claude14 March 2023; product and API launch.
- Claude 211 July 2023; model and Claude.ai public beta.
- Claude 3 model family4 March 2024; Opus, Sonnet and Haiku tier structure.
- Claude 3.5 Sonnet20 June 2024; model, vision and Artifacts milestone.
- Updated Claude 3.5 models and computer use22 October 2024; model and product expansion.
- Introducing the Model Context Protocol25 November 2024; open protocol purpose, architecture and initial ecosystem.
- Claude 3.7 Sonnet and Claude Code24 February 2025; hybrid reasoning and Claude Code research preview.
- Introducing Claude 422 May 2025; Opus 4, Sonnet 4, Claude Code general availability and agent API tools.
- Claude Sonnet 4.5 and Agent SDK29 September 2025; model and agent infrastructure.
- Claude Opus 4.524 November 2025; beginning of the measured current Opus sequence.
- Claude Opus 4.65 February 2026; long context, adaptive thinking, agent teams and compaction.
- Claude Opus 4.716 April 2026; coding, vision and long-running work.
- Claude Opus 4.828 May 2026; model and product update.
- Claude Fable 5 and Claude Mythos 59 June 2026; shared underlying model and deployment boundaries.
- Claude Sonnet 530 June 2026; current Sonnet tier and product availability.
- Claude Opus 524 July 2026; current Opus tier and deployment safeguards.
- Introducing Anthropic LabsProduct-incubation path covering Cowork, Skills and MCP-based work.
- Anthropic and Amazon compute agreementPrimary-cloud role, Trainium capacity, investment and long-term commitment.
- Google and Broadcom compute agreement6 April 2026; next-generation TPU commitment and company-reported customer evidence.
- Microsoft, NVIDIA and Anthropic partnershipsAzure capacity, NVIDIA engineering and investment commitments.
- SpaceX compute agreementColossus capacity and the wider multi-provider compute map.
- Anthropic Responsible Scaling PolicyVersion 3.4 effective 8 July 2026; page updated 14 August 2026.
- Claude models overviewCurrent API IDs, modalities, token limits, thinking modes, cutoffs and snapshot behaviour; accessed 24 August 2026.
- Constitutional AI15 December 2022; supervised and reinforcement-learning stages using AI feedback.
- Anthropic Series G financing12 February 2026; company-reported Claude Code run-rate and business-customer evidence.
Research cut. Facts are current to 24 August 2026, 16:10 ICT. All cited sources are primary sources: Anthropic announcements, documentation, research and policy pages. No secondary reporting is used. Company-reported metrics remain company disclosures rather than independent verification.