Zan Digital › Companies
Anthropic
Every piece of Zan Digital analysis that covers Anthropic — 82 pieces, each figure traced to a named source.
- 001AI Capex 2026: The $725B Bet With No Payback MathFour hyperscalers plan roughly $725B of capex in 2026, up 77% on 2025. Here is where the money goes, what it earns today, and the three numbers to watch.$725B
- 002AI Circular Deals: Who Actually Pays Whom, and How MuchNvidia's up to $100B commitment to OpenAI was finalised at $30B. A map of the AI circular deals: what is signed, what is only intent, and why it matters.$100B → $30B
- 003AI Cost Management: Cheap Tokens, Bigger AI BillsToken prices fell as fast as 40x a year, yet 93% of enterprises blew past their AI budgets in 2026. Here is the arithmetic behind the bigger invoice.5x
- 004Anthropic IPO: The Enterprise Listing That Resets PricingAnthropic filed confidentially on 1 June 2026 for an October listing. A $47B run rate, 85% enterprise revenue, and why the mix matters more than the number.$47B
- 005Cursor at $2B ARR: Fastest SaaS Ever or Most Fragile?Cursor doubled revenue in three months to $2B ARR and sought funding near $50B. The growth is real. So is the question about whether the moat is.$2B
- 006GenAI ROI: Why 95% of Pilots Show No Return, and Who DoesMIT found 95% of GenAI pilots show no measurable P&L return. What the 5% did differently, where the money is really being made, and how to measure your own.95%
- 007OpenAI IPO: What a Delay Past 2026 Actually SignalsOpenAI filed confidentially in June 2026, then weighed a delay rather than list below $1T. What the numbers say about why, and what each outcome means.$852B
- 008Sovereign AI Funds: $30B and the Buyer of Last ResortSovereign AI spending pushed $30 billion through Nvidia alone in fiscal 2026. Here is what state programmes actually buy, and what it changes for vendors.$30B+
- 009The AI Revenue Leaderboard: Who Has Real RevenueCommitments, run rates and bookings get reported as if interchangeable. A ranking by revenue quality rather than headline number, with what each figure means.~$47B
- 010Vendor Financing in AI: The Lucent and Nortel LessonSuppliers lending customers money to buy their products inflated telecom revenue before 2001. Similar structures now appear across AI. The mechanics compared.2001
- 011Agent Marketplaces: What Distribution Is Actually WorthEvery major platform launched an agent marketplace. In one editor's data, 3 agents took 92% of turns. What marketplace distribution is really worth.92%
- 012AI Moats in 2026: Integration Depth Beats Model DepthThe top 4 AI labs sit within 25 Elo points of each other. Model quality no longer separates products. Integration depth does, and here is the evidence.25 Elo
- 013MCP: The Standards Fight That Decides the Agent StackMCP went from one company's experiment to a Linux Foundation standard with 97M monthly SDK downloads. What it settles, what it does not, and who benefits.~97M
- 014Software Procurement in 2026: The New CIO Scorecard45% of AI budget now comes out of existing software lines, per Redpoint's survey of 141 CIOs. The new evaluation criteria, and a scorecard you can copy.45%
- 015Agentforce vs Fin vs Breeze: 3 Prices, 3 MetersFin bills $0.99 per outcome, Breeze $0.50 per resolved chat, Agentforce $0.10 per action. The 10x gap is a unit mismatch. Normalised, the order inverts.10x
- 016AI COGS: How to Read a Vendor's Real Gross MarginC3.ai reported 31% GAAP gross margin in FY2026 and 46% non-GAAP. AI COGS hides inference, retrieval, evaluation and human review. How to read the real number.31% vs 46%
- 017AI Contract Negotiation: 6 Clauses That Beat DiscountFederal AI contract terms now specify 72-hour incident notice and 30-day model change notice. Six clauses CFOs are winning in 2026, and what each is worth.72 hours
- 018AI Provider Contracts Compared: Anthropic, OpenAI, GoogleAnthropic, OpenAI and Google differ on retention, training rights, indemnity and uptime. None of the 3 publishes a credit-backed uptime SLA on standard terms.99.48% v 99.94%
- 019AI Shrinkflation: Same Price, Fewer Tokens, Quieter LimitsAI shrinkflation is real but mis-stated. Vendors owe API buyers 60 days notice before a model dies. Subscribers got 31 before limits changed. How to detect it.Under 5%
- 020AI Unit Economics: When Power Users Lose You MoneyAI product gross margins ran 45% in 2025 against the 70-80% SaaS norm. Here is the contribution margin model that shows which customers cost you money.45%
- 021Anthropic Enterprise Pricing: $20 Seats, Uncapped UsageAnthropic cut the Claude Enterprise seat to $20 a month and unbundled every token. Here is the strategic logic, and who absorbs the new volatility.$20
- 022Rate Limits as a Pricing Lever: A 5-Test StandardGitHub's premium request multipliers span 173x, from 0.33 to 57. A 5-test transparency standard for AI rate limits, with four vendors graded against it.173x
- 023Usage-Based Pricing Problems: The Concessions Buyers Win79% of surveyed enterprises hit AI cost overruns in the past year. Usage-based pricing broke predictability. These are the contract terms buyers now win.79%
- 024AI Chief of Staff: 3 Tasks That Transfer, 4 That Do NotMeeting agents now handle scheduling, notes and follow-up. On 175 real office tasks the best agent finished 30%. Here is what transfers and what does not.30.3%
- 025AI Code Security Scanners Miss 61% of Real FlawsFour SAST tools combined found 38.8% of known flaws in production Java. Agents now write code 3 to 4 times faster. Why scanner placement beats detection.38.8%
- 026AI Coding Tools: Three Winners, Three ScoreboardsClaude Code leads developer satisfaction at 91% CSAT, Cursor sold for $60B, and Copilot sits inside 90% of the Fortune 500. Three scoreboards, three winners.91%
- 027AI Data Retention Policies, Ranked by What They PromiseEvery major AI vendor's data retention policy, scored on 5 criteria. Anthropic keeps covered-model prompts for 30 days even where zero retention was agreed.30 days
- 028AI for Excel Hits 34.9% on Multi-Sheet Model TestsOn real multi-sheet financial models, the best AI for Excel scores 34.9%. Here is what four benchmarks measure, where they disagree, and how to test.34.89%
- 029Autonomous Coding Agents Tested on Real RepositoriesTop systems resolve 61.5% of public benchmark tasks but 51.5% on private commercial codebases. What Devin, Factory and agent buyers should take from it.61.6%
- 030ChatGPT vs Claude vs Gemini Enterprise: Admin ControlsChatGPT, Claude and Gemini Enterprise are close on model quality. Governance decides: admin controls, 180-day audit logs, 10 data residency regions.180 days
- 031Chinese Open Models: The Sovereignty Cost of 90% SavingsChinese open models run 60% to 90% cheaper and hit 46% of US enterprise tokens. What self-hosting fixes, the one risk it does not, and how to decide.46%
- 032Claude Code vs Cursor vs Copilot: The 2026 VerdictGitHub Copilot's workplace adoption fell from 29% to 21% while Claude Code rose to 39%. A buyer-side comparison built on adoption data, not benchmarks.39%
- 033Clay vs Apollo vs ZoomInfo: Cost Per Verified ContactPublished list prices put one enriched contact between $0.02 and $1.12, a 56x spread. Here is the cost per verified contact math for Clay, Apollo and ZoomInfo.$650.5M
- 034Deep Research Tools Compared: Citation Accuracy AuditedCitation accuracy across deep research agents ranges from 77.96% to 93.68%, and the volume leader ships 13.3% dead URLs. What to measure before you buy.93.68%
- 035GitHub Copilot Market Share: The 51% Drop Nobody Can SourceThe claim that GitHub Copilot fell from 67% to 51% of developers is not in the survey it cites. The measured decline is 29% to 21%. Here is the evidence.21%
- 036Glean vs Copilot: The Permissions Gap Nobody DemosGlean vs Copilot is decided by permission plumbing. Microsoft syncs connector ACL changes only on a full crawl, daily by default. Glean falls back to 24 hours.30M
- 037Harvey vs Legora vs General AI: A Use-Case ScorecardLegal AI and general assistants both hit 80% research accuracy against a 71% lawyer baseline. How to score Harvey, Legora and ChatGPT by use case.80% vs 71%
- 038LangChain vs LlamaIndex vs Building It YourselfThe langchain package took 254 million downloads last month while practitioners argue for thin custom layers. What abstraction costs, and when to build it.254.5M
- 039Model Routing Savings: 31.7% Measured, Not 85%The largest independent routing benchmark measured 31.7% cost savings, not 85%. Here is the quality gap across five enterprise tasks, and where it hides.31.7%
- 040n8n vs Zapier vs Make: The Billing Unit DecidesZapier Pro allows 1,500 agent activities a month, capped at 40 per run, or about 37 full-depth runs. n8n bills the identical agent loop as one execution.1,500
- 041Open Weights in Production: Qwen, DeepSeek, Llama, GLMChinese models take 61% of OpenRouter tokens and about 1% of enterprise LLM spend. Comparing Qwen, DeepSeek, Llama and GLM on quality, cost and licence.61%
- 042Perplexity Enterprise vs Google AI Mode for ResearchPerplexity had the lowest error rate of 8 AI search tools tested and still missed 37% of citations. For research teams, source control decides this one.37%
- 043Agent Drift Is Silent. Only 14% Monitor Production AIOnly 14% of teams monitor LLM apps in production. Agent drift comes from model updates, data shifts and prompt sprawl. How to detect it before customers do.14%
- 044AI Agent Payback: Why Every Function Says 6 MonthsSix commissioned studies report AI agent payback in under 6 months across six functions. The one with 420 respondents says 15. What that gap means.6 of 7
- 045AI Change Management: Why 54% Bypass Your AI Tools54% of employees bypassed their company's AI tools in the past 30 days. Here is why passive resistance beats mandates, and the change management that works.54%
- 046AI Procurement Checklist: 5 Sections Most RFPs MissGartner expects over 40% of agentic AI projects to be cancelled by 2027. A 5-part procurement checklist for capability, data, cost, evals and exit terms.40%+
- 047AI Vendor Lock In: What Switching Models Really CostsAnthropic retired 9 models in 10 months, one on 61 days notice. The API call is the cheap part. Here is what model switching actually costs to redo.40%
- 048Build vs Buy AI Agent Framework: The 3-Test RuleAnthropic retired 9 Claude model IDs in 12 months and MCP rewrote its core in July 2026. Three tests decide whether a custom agent framework is worth it.9
- 049Context Engineering Replaced Prompt Engineering. Mostly.Anthropic shipped a 1M token window at standard pricing in March 2026, and its own best model still misses 21.7% of MRCR v2. Context engineering is why.78.3%
- 050Enterprise AI Cost: Why Inference Is the Small LineA full cost model for one production agent at 240,000 tasks a year. Inference is 5% of the bill. Review labour, evals and integration are the rest.5%
- 051Human in the Loop Is the Architecture, Not a Cop-OutClaude Code users approve 93% of permission prompts. That is what a bolted-on review gate produces. Here is how to design the loop before you buy the agent.73%
- 052LLM Evals: The Only Asset That Survives a Model SwapAnthropic retired 9 Claude models in 10 months and gives 60 days notice. Your eval suite is the only asset that survives. How to build one from scratch.60 days
- 053Multi-Agent Systems: 15x Tokens for a Narrow PayoffMulti-agent systems use about 15x the tokens of a chat, and lose to single agents under matched budgets. The 4 tests that decide the architecture.15x
- 054RAG Is Not Dead: 11 of 13 Models Fail at 32K TokensLong context did not kill retrieval. At 32K tokens, 11 of 13 models fell below half their short-context accuracy. How to diagnose and fix bad RAG.11 of 13
- 055AI Agent Liability: Authority Clauses for 2026 DealsAir Canada paid CA$812.02 for a promise its chatbot invented. Attribution is already settled law. The unsettled part is which clause moves the liability.CA$812.02
- 056AI Audit Log Design: 12 Fields Agents Must RecordOnly 21% of enterprises can see what their agents do at runtime. Here is a 12-field minimum audit record for AI agents, and what each field has to prove.21%
- 057AI Copyright Lawsuits: The $1.5B Repricing of Training DataAnthropic paid $1.5 billion, about $3,000 a book, to settle. Here is what the live AI copyright cases could do to model economics and licensing prices.$1.5B
- 058AI Cybercrime: 55% in Africa, 2% in FBI ComplaintsINTERPOL attributes 55% of reported African cybercrime to AI. The FBI's 2025 complaint data puts it near 2%. Why both are right, and which controls follow.55%
- 059AI Data Residency: What EU and India Buyers Demand51% of firms now rate data sovereignty as very important. What EU and Indian buyers actually demand, and where vendor residency claims stop short.51%
- 060AI Export Controls: The 19 Days Claude Went DarkUS export controls pulled Claude Fable 5 offline worldwide for 19 days in June 2026. The order, the reversal, and the precedent it set for your stack.19 days
- 061AI Red Teaming for Agents: Building a Real ProgrammeNIST red teamers raised agent hijacking success from 11% to 81%. Here is how to scope, staff, run and report an AI red teaming programme for agents.81%
- 062Data Processing Agreement Gaps: 7 Clauses to RewriteMost DPAs predate agent tooling and model updates. Here are the 7 clauses to rewrite, with the EDPB findings and vendor terms that made them stale.38%
- 063EU AI Act Transparency Rules: The 2 August ChecklistArticle 50 transparency duties applied from 2 August 2026, with fines to €15M or 3% of turnover. Only 9 of 27 member states have both enforcers named.€15M or 3%
- 064MCP Security: 200,000 Servers Exposed by One Design FlawOX Security found one MCP design flaw exposing up to 200,000 instances across 150 million downloads. Here is how to vet a connector before you trust it.200,000
- 065Personal AI Accounts: 67% of AI Use Is Off Your BooksVerizon's 2026 DBIR found 45% of employees now use AI on corporate devices, and 67% of them sign in with personal accounts. Source code leads the leak.67%
- 066Prompt Injection: 12 Defences Broken Above 90%Researchers broke 12 published prompt injection defences with attack success above 90%. Why the SQL injection fix has no equivalent, and what actually works.Above 90%
- 067Prompt Injection: 4 of 26 Models Paid the AttackerZscaler ran 26 models against live injected pages and 4 paid the attacker. Why prompt injection is the agent era's SQL injection, and what contains it.4 of 26
- 068SOC 2 AI Questions: 9 Artefacts Auditors Now WantThe AICPA has issued no AI-specific SOC 2 criteria. The 2017 trust services criteria already cover AI tooling. Here are the 9 artefacts auditors ask for.0
- 069AI Pricing Accuracy: Fix What ChatGPT Gets WrongAI assistants ship a significant issue in 45% of answers. Here is why they misquote SaaS pricing, and a six step fix for the pages they actually read.45%
- 070AI Referral Traffic: 4.4x Is a Model, 1.3x Is MeasuredThe 4.4x conversion claim for AI referral traffic is a model output. The datasets that publish a denominator land near 1.3x. What to measure instead.1.31x
- 071AI Visibility Tracking: Build the Dashboard in a DayThree AI models agree on the top recommended brand just 41.6% of the time. Here is the free AI visibility dashboard you can build in a single day.41.6%
- 072Block AI Crawlers? What robots.txt Actually ControlsCloudflare measured Anthropic crawling 70,900 pages per referral sent back. Blocking AI crawlers is defensible. robots.txt controls less than you think.70,900:1
- 073Gated Content Math: 99% of Your Leads Never CloseForrester puts inquiry-to-close at under 1%, so the gate fails 99% of the time. Here is what the form costs in AI visibility, and what replaces it.Under 1%
- 074Marketing Attribution Broke: 68% of Searches End Unclicked68% of US Google searches ended without a click in early 2026. Here is which attribution methods survive dark discovery, and which ones quietly invent numbers.68.01%
- 075Software Review Sites Sold for $110M: The ReckoningGartner wrote off $150M on Digital Markets, then sold Capterra, Software Advice and GetApp to G2 for $110M. What that says about software review sites.$110.0M
- 076Zero-Click Search Hit 68%: What to Report Instead68.01% of US Google searches ended without a click in early 2026, up from 60.45% two years earlier. The reporting metrics that replace organic sessions.68.01%
- 077AI Fluency Hiring: 9 Questions, Not Prompt TricksPrompt engineer missed Microsoft's top 10 AI hiring roles. One 2026 study found people took wrong AI answers 80% of the time. 9 questions that test judgement.822
- 078AI Tool Standardisation: What One Tool Really Costs70% of engineers run two to four AI tools at once, per a 906-person survey. Standardising improves governance and costs you goodwill. Here is the trade-off.70%
- 079AI Training Programmes That Actually Changed Behaviour79% of employees with more than five hours of AI training use it regularly, against 67% below that line. What separated the programmes that changed behaviour.79%
- 080Human Skills After AI: Which Ones Gained ValueEmployment for 22 to 25 year olds in AI-exposed jobs is 19% below trend. The human skills that gained value are the ones nobody wrote down. Here is the list.19%
- 081Product Manager Skills When AI Writes 75% of CodeAI now writes 75% of new code at Google. That moves the product manager job upstream, toward specification quality, evaluation design and problem selection.75%
- 082Technical Interviews When Everyone Ships With AI71% of engineering leaders say AI has made technical skills harder to assess. Portfolios and take-homes stopped signalling. Here is what replaces them.71%