The Agentic AI Valuation Framework
How to value an agentic AI company when the model is no longer the moat -- DCF, comps, and the new rules driving Anthropic toward a $950B raise.
1 · Why valuation suddenly matters
Anthropic is in talks to raise $30-50B at a pre-money valuation of $900-950B -- pushing it past OpenAI's $850B and within sight of a $1T mark OpenAI itself just targeted in its May 22 S-1 filing. Anthropic's ARR has compressed from $9B (late 2025) to $44B (May 2026), an 80x lift in five quarters. Cowork, Claude Code, and the 10-agent Finance pack have collapsed the per-seat SaaS model into per-task economics -- and that breaks every valuation framework built around DAU and ARPU. So what: Every CFO, investor, and founder is being asked the same question this quarter -- what is an agentic AI company actually worth, and how do you defend the multiple when the underlying
| Primitive | Definition | 2026 benchmark |
|---|---|---|
| Active Agents (AA) | Distinct agent identities (SVID/DID) calling your platform per month. | Anthropic: ~1.5M Claude Code dev agents, Cowork undisclosed |
| Tasks per Agent (TPA) | Completed multi-step tasks per agent per month. 5-50x richer than chat turns. | Salesforce Agentforce: 18,500 customers, ~190 tasks/agent/mo |
| Revenue per Task (RPT) | Blended price across model routing tiers (nano/mid/frontier). | $0.05-$0.50 typical; legal redline $2.00; payroll recon $0.30 |
| Task Completion Rate (TCR) | % of attempted tasks completed without human escalation. The 99.9% threshold. | Best-in-class: 98-99.5%. The 9.9% gap is where 90% of valuation lives. |
2 · DCF rebuilt for per-task economics
The classic SaaS DCF leans on seats × ARPU × net retention. Agentic businesses don't sell seats -- they sell completed tasks. Rebuild the model around four primitives: Forward revenue then becomes AA × TPA × RPT × TCR. Discount it at 12-15% (frontier model risk + EU AI Act enforcement risk + agent-economy regulatory tail). The terminal multiple is no longer 8-10x SaaS FCF; it's anchored to the marketplace position you've earned in the skills ecosystem (Day 60-62) -- typically 15-25x for
| Primitive | Definition | 2026 benchmark |
|---|---|---|
| Active Agents (AA) | Distinct agent identities (SVID/DID) calling your platform per month. | Anthropic: ~1.5M Claude Code dev agents, Cowork undisclosed |
| Tasks per Agent (TPA) | Completed multi-step tasks per agent per month. 5-50x richer than chat turns. | Salesforce Agentforce: 18,500 customers, ~190 tasks/agent/mo |
| Revenue per Task (RPT) | Blended price across model routing tiers (nano/mid/frontier). | $0.05-$0.50 typical; legal redline $2.00; payroll recon $0.30 |
| Task Completion Rate (TCR) | % of attempted tasks completed without human escalation. The 99.9% threshold. | Best-in-class: 98-99.5%. The 9.9% gap is where 90% of valuation lives. |
3 · The 2026 comp sheet
ARR multiples have widened -- investors have stopped paying 50x just for the AI label. The new ceiling is reserved for companies that own a workflow and the data feedback loop that makes the agent uniquely reliable. Two patterns are visible. First, frontier-lab valuations have decoupled from current ARR and re-anchored to distribution surface (OpenAI's 900M WAU, Anthropic's enterprise share). Second, EU-sovereign players (Mistral) trade at extreme premiums on geography alone -- a direct consequence of the August 2 EU AI Act
| Primitive | Definition | 2026 benchmark |
|---|---|---|
| Active Agents (AA) | Distinct agent identities (SVID/DID) calling your platform per month. | Anthropic: ~1.5M Claude Code dev agents, Cowork undisclosed |
| Tasks per Agent (TPA) | Completed multi-step tasks per agent per month. 5-50x richer than chat turns. | Salesforce Agentforce: 18,500 customers, ~190 tasks/agent/mo |
| Revenue per Task (RPT) | Blended price across model routing tiers (nano/mid/frontier). | $0.05-$0.50 typical; legal redline $2.00; payroll recon $0.30 |
| Task Completion Rate (TCR) | % of attempted tasks completed without human escalation. The 99.9% threshold. | Best-in-class: 98-99.5%. The 9.9% gap is where 90% of valuation lives. |
4 · The new moat: skills + governance, not model
Model performance has compressed to under 3% across Claude Opus 4.7, GPT-5.5, Gemini 3.5 Flash, and Qwen3.7-Max. That means the model is no longer a moat -- it's a commodity component inside model-routing stacks (Day 43). The valuation multiple now lives in three places:
| Company | ARR (run-rate) | Implied valuation | ARR multiple |
|---|---|---|---|
| OpenAI | $25B | $850B-$1T (S-1) | ~34-40x |
| xAI | ~$5B | $250B (SpaceX deal) | ~50x |
| Cursor | $2B | $50B (rumoured) | ~25x |
| Databricks | $4.8B | $134B | ~27.9x |
| DeepSeek | Undisclosed | China-only ~$80B | n/a |
| Cohere | $200M | $5B | ~25x |
| Mistral | $120M | $11.5B | ~96x (sovereignty premium) |
5 · Treating agents as customers with LTV
The biggest break with SaaS valuation is that agents are now customers. Coinbase reports 165M x402 transactions across 69K active agent wallets in April 2026; $600M+ now sits in 500K+ active agent wallets. These agents pay, renew, churn, and refer other agents -- and they need their own LTV/CAC model:
| Company | ARR (run-rate) | Implied valuation | ARR multiple |
|---|---|---|---|
| OpenAI | $25B | $850B-$1T (S-1) | ~34-40x |
| xAI | ~$5B | $250B (SpaceX deal) | ~50x |
| Cursor | $2B | $50B (rumoured) | ~25x |
| Databricks | $4.8B | $134B | ~27.9x |
| DeepSeek | Undisclosed | China-only ~$80B | n/a |
| Cohere | $200M | $5B | ~25x |
| Mistral | $120M | $11.5B | ~96x (sovereignty premium) |
6 · Pure-skill, pure-marketplace, pure-orchestration
Three new startup archetypes need bespoke valuation frames:
| Metric | Human SaaS analogue | Agent-economy translation |
|---|---|---|
| Net retention | Expansion / churn | Tool-call expansion as orchestrator agents discover more of your skills |
| Churn signal | Login drop-off, NPS | OTEL gen_ai span drop, cosine<0.7 drift, model rotation by orchestrator |
7 · Public market read-through
Three public tape signals validate the framework: (a) Microsoft's M365 E7 tier -- the first new enterprise SKU since E5 (2015) -- added a measurable ARPU lift credited explicitly to Agent 365 in MSFT's Q3 FY26 call. (b) Salesforce traded to a fresh agentic premium after Agentforce 3.0 crossed $500M ARR with 18,500+ customers. (c) ServiceNow's Autonomous Workforce drove gross margin expansion of ~180 bps QoQ as 90%+ of IT tickets moved off human agents. Public markets are now paying explicitly for the same three moats -- skills ecosystem, governance evidence, and distribution lock -- that private agentic comps trade on.
| Metric | Human SaaS analogue | Agent-economy translation |
|---|---|---|
| Net retention | Expansion / churn | Tool-call expansion as orchestrator agents discover more of your skills |
| Churn signal | Login drop-off, NPS | OTEL gen_ai span drop, cosine<0.7 drift, model rotation by orchestrator |
8 · Viral AI app of the day
OpenHuman by tinyhumansai -- climbed to #1 on GitHub Trending May 13-16 by inverting the agent UX playbook. Instead of waiting for a first prompt, it pre-loads a 1B-token Memory Tree from 118+ services via one-click OAuth (Gmail, GitHub, Slack, Calendar, Drive, Stripe, Linear, Jira). A 'subconscious loop' runs every 20 minutes without user input, reading to-dos and recent memory to decide what to act on. The desktop mascot can join Google Meets as a separate participant, transcribing into the Memory Tree and speaking back. All local-first (Rust + Tauri, MIT, SQLite + Obsidian wiki) -- no data egress unless explicitly escalated. Valuation read: OpenHuman is a pure-skill / pure-orchestration hybrid with zero ARR today -- but the 1B-token Memory Tree is exactly the kind of data feedback loop investors are paying 15-20x ARR for. Watch for a paid SaaS tier or A2A registry play within 90 days.
| Metric | Human SaaS analogue | Agent-economy translation |
|---|---|---|
| Net retention | Expansion / churn | Tool-call expansion as orchestrator agents discover more of your skills |
| Churn signal | Login drop-off, NPS | OTEL gen_ai span drop, cosine<0.7 drift, model rotation by orchestrator |
9 · Practical takeaway for tomorrow
If you're a CFO, build a one-page agentic valuation deck with five blocks: (1) AA × TPA × RPT × TCR forward revenue, (2) ARR multiple anchored to 22x (Anthropic comp) discounted by your skills ecosystem position, (3) Skills + governance + distribution moat scorecard, (4) Agent-LTV/CAC for any third-party agent integrations, (5) Annex III governance evidence as a discount-rate reducer. If you're a founder, publish a SKILL.md into all four marketplaces this week -- the asymmetry window closes in 2027 when revenue shares formalise. Tomorrow (Day 64): Agentic AI in Financial Services -- Anthropic's 10 finance agents in production, JP Morgan's COiN at scale, and the new FinCEN AML rules landing Q3 2026.
| Metric | Human SaaS analogue | Agent-economy translation |
|---|---|---|
| Net retention | Expansion / churn | Tool-call expansion as orchestrator agents discover more of your skills |
| Churn signal | Login drop-off, NPS | OTEL gen_ai span drop, cosine<0.7 drift, model rotation by orchestrator |
multi-agent customer by 5-20x. Re-slice by tasks billed and agent identities served -- the same artefact that produces your EU AI Act Annex III audit pack.