OpenAI and Anthropic:
The Pre-IPO Showdown

01Two Radically Different Prospectuses

IPO valuation can be decomposed into five variables:

Where N is user base, P is paid conversion rate, ARPU is average revenue per user, R is retention dependency, and M is the narrative multiplier. The narrative multiplier M can be further decomposed into M_growth (market confidence in growth) and M_quality (market confidence in revenue quality and margin sustainability). OpenAI’s advantages concentrate in N and P — a 900 million base + GUI-level low-barrier conversion; Anthropic’s advantages concentrate in ARPU and R — high-value enterprise contracts + deep workflow lock-in. This explains an apparent paradox: Anthropic’s revenue is higher ($47B vs. OpenAI undisclosed), yet OpenAI’s valuation is comparable ($852B vs. $965B) — because the market assigns a higher M_growth to OpenAI and a higher M_quality to Anthropic.

02CLI vs GUI: A Civilizational Fork in February 2026

The evolutionary paths of Claude Code and Codex reached a decisive fork in February 2026. That same month, Anthropic released Opus 4.6 (a model-layer upgrade) while OpenAI released the Codex desktop app (an interaction-layer upgrade). One dug deeper; the other spread wider. Over the following four months, this fork evolved from a “product choice” into an “ecosystem lock-in.”

Claude Code Evolution Timeline

Codex Evolution Timeline

03The Multimodal Product Loop Gap: Model Capability ≠ Interaction Loop

A strict distinction must be drawn between two levels: model-layer multimodal capabilities and product-layer multimodal interaction loops. Claude is not devoid of multimodal capabilities — Anthropic’s official documentation clearly states that it possesses vision capabilities for image understanding and analysis. However, the gap is significant when it comes to converting these capabilities into consumer-grade product loops.

Anthropic’s catch-up moves include the Vercept acquisition (February) and the Claude Design launch (April, supporting design prototypes, marketing assets, and Canva export). However, as of June, these remain in early-stage integration.

04Economic Fractures Within B-Side High Growth

Anthropic’s B-side revenue figures are striking — surging from $1B to $47B in 18 months, with Ramp data showing B-side AI billing share leaping from 10% to 60%. Yet during the same period, benchmark enterprise clients began exiting, and the tokenmaxxing phenomenon was declared “over” by Fortune. This is not a contradiction, but a time lag between leading and lagging indicators. The $47B in revenue is the continuation of growth inertia — contract cycles, migration costs, and organizational inertia ensure that revenue figures always trail shifts in usage behavior. The exits by Microsoft and Uber are early fracture signals; if such signals proliferate in Q3–Q4, the inflection point in revenue growth will follow.

May–June Fracture Timeline

Tokenmaxxing is not merely a pricing issue — it is fundamentally an organizational behavior failure. When token consumption is perverted into a KPI, Goodhart’s Law inevitably takes effect — the metric itself becomes the object of manipulation. Genuine AI productivity comes from workflow restructuring, not simple automation.

Goldman Sachs projects that Agent-driven token consumption will grow 24× by 2030. Even if per-token prices fall 90%, total enterprise bills will still rise under the current usage-based billing model. However, it should be noted that usage-based billing is not the final state — task-based pricing, outcome-based pricing, and hybrid subscription models may restructure the cost architecture. Compounded by the enterprise data security ceiling (the Samsung incident, EU AI Act transparency provisions taking effect in August 2026), B-side growth faces structural headwinds. Compliance pressure is not exclusive to Anthropic — the EU AI Act’s transparency rules and AI-generated content labeling requirements equally constrain C-side GUI Agent expansion in European markets.

The AI Layoff Boomerang

05The C-Side GUI Agent Demand Flywheel and the Data Double-Edged Sword

After the ChatGPT + Codex merger on June 2, 900 million users (industry tracker estimate) gained Agent capabilities overnight. “Default” is itself a moat.

The Data Flywheel: Strategic Asset or Noise Trap?

C-side growth generates a strategic asset that is exceedingly difficult to replicate: a dataset of human physical-operation behaviors — cross-application operational traces, hesitation nodes, error-correction records, and approval paths. This is far closer to real workflows than static corpora, and constitutes scarce fuel for next-generation Agent fine-tuning.

But the data flywheel is a double-edged sword. Signal-to-noise risk: Among 900 million mass-market users, a vast volume of meaningless clicks, illogical operations, and erroneous commands constitutes noise that may contaminate the model fine-tuning pool. By comparison, the clean, high-logic code data produced by the professional developers and enterprise users served by Anthropic may possess a higher signal-to-noise ratio on the path toward more advanced reasoning models (System 2 Thinking). Privacy-usability constraint: The strategic value of behavioral data depends on usability rather than sheer scale — Usable Data = Behavioral Data × User Consent × Retention Policies × Privacy Filtering × Enterprise Data Isolation Policies. In an increasingly strict data regulatory environment, the gap between “theoretically obtainable” and “practically usable” is significant.

06Behavioral Liberation Modules for Full-Stack AI Interaction

If AI can only be used while sitting at a desk, that is the greatest behavioral constraint imposed on humanity.

On-device inference (NPU/local models) is theoretically an architecture-level solution to C-side inference costs. But hardware reality is harsh: running a Computer Use-grade multimodal model fluently requires at minimum a high-bandwidth unified memory architecture (e.g., Apple 64GB+ unified memory), and the global penetration rate of PCs meeting this specification remains extremely low in 2026. On-device inference cannot save OpenAI’s balance sheet in the short term; it remains an option belonging to 2028 or beyond.

Human Supervisory Bandwidth: The Cognitive Ceiling of Agent Parallelism

Multi-Agent parallelism ≠ linear productivity gains. The cognitive bandwidth of the human prefrontal cortex is fixed — when a non-developer simultaneously monitors, approves, and intervenes across 3–5 parallel Agent workflows, they hit the physical limits of cognitive load. This means ARPU growth is not a linear function but faces a low ceiling determined by human attention. Unless AI achieves the leap from co-pilot to auto-pilot, human supervision costs will lock Agent product pricing ceilings. This constraint is more severe on the C-side (individuals supervising alone) than on the B-side (teams distributing supervisory duties).

The Liability Attribution Problem for GUI Agents

When a GUI Agent operates a computer, errors are no longer “wrong answers” but “wrong actions” — potentially misfiring emails, misediting quotes, deleting client records, or misallocating advertising budgets. C-side verification costs are more hidden, more distributed, and harder to quantify. When an Agent misoperation causes commercial losses, how should liability be distributed among the user, model provider, plugin provider, and platform? This is a legal and commercial risk that both C-side and B-side will face in H2 2026, and there is currently no industry consensus.

07Inertia, Constraints, and Multi-Polar Competition

Anthropic’s Most Likely Path

Short-term (Q3): Double down on coding. Opus 4.9/5.0, Dynamic Workflows GA. The reliability improvements in Opus 4.8 are actively rebuilding trust.

Mid-term (Q4 – Q1 2027): Forced acceleration of GUI and C-side. Vercept integration, Claude Design iteration. Constrained by the multimodal product loop gap, the most likely path is gap-filling through acquisitions or deep partnerships.

Anthropic’s Defensive Assets

Anthropic is not merely “CLI + enterprise.” For Fortune 100 companies, data privacy isolation, private deployment, and model transparency carry far more weight than interface usability. The AWS/Azure/GCP tri-cloud strategy forms a compliance moat. Opus 4.8’s emphasis on honesty and transparency raises procurement probability in finance, healthcare, and government. This is not a small TAM — it is a different kind of TAM: low headcount, high compliance, high contract value. Ramp data showing B-side share surging from 10% to 60% is the other side of the story, obscured by the cost narrative.

OpenAI’s Most Likely Path

Short-term (Q3): Embed Codex into ChatGPT, expand plugins from 6 to 20+, launch voice interaction.

Mid-term (Q4): IPO sprint. Requires dual growth in MAU and ARPU.

OpenAI’s C-Side Gross Margin Trap

When 900 million users invoke concurrent Agents at high frequency, per-user inference cost will very likely far exceed the $20 subscription fee. This is the mirror image of the B-side token crisis, but with an asymmetric structure — B-side costs are borne by clients (who can leave, directly impacting revenue), while C-side costs are borne by OpenAI itself (users continue using the platform, but the platform sustains ongoing subsidy losses). The former is a revenue risk; the latter is a margin risk, with different crisis transmission paths and detonation timelines. The $20/month bundle is a powerful pricing anchor, but its sustainability depends on on-device inference maturity and the exploration of new billing models.

Third-Pole Forces: Platform Power and Vertical SaaS Counterattack

The ultimate entry point for GUI Agents may be neither ChatGPT nor Claude. The competitive landscape may instead be a three-pole contest:

The OS layer: Microsoft revoking Claude Code licenses and pivoting to Copilot CLI is essentially a platform reclaiming the Agent entry point into its own ecosystem. Windows with built-in Copilot, macOS with Apple Intelligence integration — OS-level “default entry points” penetrate deeper than any third-party application’s distribution.

The browser layer: Once Chrome embeds Gemini, the Agent entry point for web-based workflows may bypass both ChatGPT and Claude. Google’s AI web traffic share surged from 5.7% to 21.5% within a year.

The vertical SaaS layer: Salesforce (Agentforce), Adobe, Figma, Canva, Notion, Atlassian, and ServiceNow will not readily cede workflow entry points. They may embed Agents natively into their products, eroding the entry-point advantage of general-purpose GUI Agents. When Codex connects to Salesforce via plugin, Salesforce’s own Agent is running natively within the same interface — whose default carries more weight?

Third-pole forces may transform the endgame from “who wins, OpenAI or Anthropic” into a multi-polar landscape of “application-layer Agents vs. platform-layer Agents vs. vertical Agents.” In this landscape, players who own OS or vertical SaaS entry points may pose a “platform tax” risk to both pure-play model companies.

08March Paper Validation, Corrections, and Falsifiable Predictions

Three Key Corrections

Correction 1: “Generation vs. Control” is not a binary phase transition but a convergence. Codex simultaneously performs generation + control + orchestration; reality is superposition, not replacement.

Correction 2: The deployment path is not “screen-recording training” but platformized plugin distribution.

Correction 3: From a “physical friction ladder” to a three-tier “behavioral freedom ladder” — mobile interaction liberation + on-device inference breakthrough + human supervisory bandwidth constraints.

Falsifiable Predictions: Q3–Q4 2026 Verification Window

09The Pre-IPO Showdown Is Fundamentally a Battle Over “Who to Serve”

Anthropic serves the small elite of developers and enterprise CTOs — with unmatched depth and a fortified compliance moat. But it faces the dual constraints of a token cost ceiling and data security limitations.

OpenAI serves the vast majority of ordinary knowledge workers — with insufficient coding depth (67% of blind testers rated Claude code superior), but with breadth advantages from GUI distribution, multimodal interaction, and mobile integration. Yet the economic constraints of inference costs are equally real.

Both are infiltrating the other’s territory. But the infiltration velocity is asymmetric: Teaching hundreds of millions to use a GUI Agent is far easier than getting tens of millions of programmers to switch terminal tools.

Third-pole forces — OS platforms, browser ecosystems, vertical SaaS — may simultaneously disrupt both companies’ distribution entry points, transforming the endgame from a duopoly showdown into a multi-polar contest.

In the second half of 2026, the five falsifiable criteria established in this paper will successively enter their verification windows. This research lab will continue to track developments and publish follow-up analyses.