AI Agent Platform War: Conway, GPT-5.6, and the $2 Trillion IPO Race

The AI agent platform war just got a $2 trillion price tag. Two companies racing toward $1 trillion valuations revealed opposite visions for what AI agents should be. Anthropic filed its confidential IPO at a $965 billion valuation on June 1 (Fortune). Two weeks earlier, OpenAI filed its S-1 targeting $852 billion to $1 trillion (multiple sources). That is nearly $2 trillion in combined value riding on a single question: should an AI agent never sleep, or should it be so smart it does not need to?

Anthropic’s answer is Conway — an always-on persistent agent that runs in the background, watches your webhooks, and controls your browser while you sleep. OpenAI’s answer is GPT-5.6 — a frontier model so powerful it handles complex agentic workflows in a single pass.

The AI agent market hit $10.91 billion in 2026, up 43% year-over-year, and 40% of enterprise applications will have task-specific AI agents by end of 2026, up from less than 5% in 2025 (Grand View Research, Gartner). This is no longer a research demo. It is infrastructure.

The financial stakes make this concrete. Anthropic’s annual recurring revenue exploded from $87 million in January 2024 to $30 billion by April 2026 — an 80x jump that pushed it past OpenAI in valuation for the first time (VentureBeat, CNBC). OpenAI generates $25 billion in ARR but lost $1.22 for every $1 in revenue in Q1 2026 (Fortune).

This article breaks down what Conway and GPT-5.6 actually are, why these two companies bet on opposite architectures, and what the $2 trillion IPO race means for the professionals building on these platforms.

Conway Decoded — Anthropic’s AI Agent That Never Sleeps

Conway is not another chatbot upgrade. It is a persistent background agent that lives in a separate sidebar UI with three dedicated areas: Search, Chat, and System (TestingCatalog, April-May 2026). Think of it as hiring a junior analyst who works 24/7 — you give them tasks, they monitor triggers, and they report back when something changes.

The architecture is event-driven. Conway uses webhook triggers to wake up when specific conditions are met: a new commit hits your repository, a competitor changes pricing, a customer opens a support ticket. It executes tasks through Chrome browser control and Claude Code integration, then reports results. The system uses a proprietary .CNW extension format for storing agent configurations.

There is one critical constraint. Conway is limited to one agent instance per user. Anthropic is deliberately throttling concurrency, likely to maintain quality control over agent behavior before scaling.

Conway’s Ecosystem Play

Conway is not a standalone product. It is the tip of a broader agent ecosystem Anthropic has been building (TestingCatalog, May 31, 2026):

Orbit is a proactive assistant integrating Gmail, Slack, and GitHub — it does not wait for prompts but anticipates needs based on your workflow patterns.

Operon targets life sciences research. BugCrawl automates code bug detection. Both are vertical-specific agents that feed into Conway’s always-on framework.

Add the recently discovered file-based memory system and enhanced voice mode with mid-sentence language switching, and the picture becomes clear: Anthropic is building an agent operating system, not just an agent.

GPT-5.6 — What the Codex Canary Actually Tells Us

GPT-5.6 has not been officially announced. But the breadcrumbs are unusually specific. Internal codenames ember-alpha and beacon-alpha surfaced in OpenAI Codex logs on May 13, 2026, suggesting a possible dual variant — a flagship model and a lite/instant version (WaveSpeed Blog).

ChatGPT Pro users have reported behavior consistent with a 1.5 million token context window, 43% above GPT-5.5’s roughly 1 million tokens. If confirmed, this would be the largest context window in a commercial model.

The performance claims are aggressive. Multiple sources suggest GPT-5.6 could deliver Mythos-level reasoning at 2-3x lower cost than Mythos pricing (@pankajkumar_dev, X). Expected improvements span reasoning, frontend code generation, personality customization, and agentic workflows.

The Regulatory Holding Pattern

There is a reason GPT-5.6 has not shipped yet. Andrew Curran, a well-sourced AI analyst, stated: “Gemini Pro 3.5 and GPT-5.6 are both ready now, and both labs want to release them, but they are being held back for safety testing in a test flight of the new regulations in the forthcoming executive order” (@AndrewCurran_, X).

Polymarket prices the June 8-14 release window at 46.5% probability. The market thinks it is a coin flip, which itself tells you how close the launch is.

GPT-5.5 as a Baseline

To understand GPT-5.6’s leap, consider where GPT-5.5 stands today. It scores 88.7% on SWE-bench Verified, 81.2% on AIME, and 76% on MMMU-Pro. It was 25% faster than GPT-5.2 and — in a self-referential twist — was “instrumental in creating itself,” with GPT-5.3-Codex used during its own training (OpenAI).

The $2 Trillion IPO Showdown

FIG. 03 — $2T IPO SHOWDOWN

Anthropic vs OpenAI: The $2T IPO Race

Metric
Anthropic
OpenAI

Valuation
$965B
$852B-$1T

ARR
$30B
$25B

Q2 Revenue
$10.9B
~$6B

Profit
Approaching first profit
-$1.22/$1 (Q1 2026)

Agent Play
Conway (always-on)
GPT-5.6 + Codex

SOURCE: Fortune, CNBC, VentureBeat (June 2026)

The dual IPO filings create a natural experiment: which AI business model is Wall Street willing to underwrite?

Anthropic filed its confidential IPO on June 1, 2026, at a $965 billion valuation after raising $65 billion in total capital. Goldman Sachs, JPMorgan, and Morgan Stanley are leading the deal. The company is approaching its first profitable quarter (Fortune, June 1, 2026).

OpenAI filed its confidential S-1 on May 22, 2026, targeting an $852 billion to $1 trillion valuation. Goldman Sachs and Morgan Stanley are leading. But OpenAI lost $1.22 for every $1 of revenue in Q1 2026 and projects $14 billion in operating losses for the full year. Positive cash flow is not expected until 2030 (multiple sources).

Anthropic’s 80x Growth Story

The revenue trajectory is what made Wall Street take notice. Anthropic went from $87 million ARR in January 2024 to $1 billion by December 2024, then $9 billion by end of 2025, then $14 billion in February 2026, $19 billion in March 2026, and $30 billion by April 2026 (VentureBeat, CNBC). Expected Q2 2026 revenue is $10.9 billion in a single quarter.

That 80x growth in roughly two years is what eclipsed OpenAI in valuation for the first time. But the question for public markets is whether Conway-led enterprise stickiness can justify $965 billion — or whether this is 2021 all over again.

The Real Race — AI Agent Platforms, Not Models

The benchmark battle is razor-thin. Claude Opus 4.8 leads the Artificial Analysis Intelligence Index at 61.4 versus GPT-5.5’s 60.2. It wins SWE-Bench Pro (69.2% vs 58.6%), OSWorld-Verified (83.4% vs 78.7%), and Finance Agent v2 (53.9% vs 51.8%). GPT-5.5 wins Terminal-Bench 2.1 (BenchLM, multiple sources).

But the pricing gap tells a different story entirely. Opus 4.8 costs $5/$25 per million tokens (input/output). GPT-5.5 costs $1.25/$10. That is a 4x difference. If GPT-5.6 delivers Mythos-level performance at 2-3x lower than Mythos pricing, the cost advantage could widen further.

This creates two distinct go-to-market strategies. Anthropic’s play is the agent platform — Conway as the always-on backbone that locks enterprises into Claude’s ecosystem regardless of model cost. OpenAI’s play is model dominance — a model so capable and cheap that developers choose it as their default, making the agent layer a commodity.

The Platform Lock-In Question

The strategic question is whether agent platforms or model quality will create the deeper moat. Conway’s persistent architecture creates workflow dependency — once your business processes run on webhook triggers and .CNW configurations, switching costs are high. OpenAI’s model-first approach bets that raw intelligence at lower cost will attract the largest developer base, creating a different kind of lock-in through API integrations.

Google is watching closely. Gemini Spark, its desktop-native agent, and Gemini 3.5 Flash represent a third approach: embed agents directly into the operating system layer. Andrew Curran’s claim that Gemini Pro 3.5 is “ready now” but held for regulatory compliance suggests a three-way race.

The agent market data supports the urgency. With 40% of enterprise applications expected to have task-specific AI agents by the end of 2026, up from less than 5% in 2025, the platform that captures this buildout captures a decade of infrastructure revenue (Gartner).

What This Means for Korean Tech Workers

South Korean enterprises are not watching from the sidelines. A CIO survey found that 59.3% of Korean IT budgets are now directed to generative AI, making it the single largest technology investment category (ComWorld, DigitalToday).

The Korea Standards Association and Telecommunications Technology Association launched AI trustworthiness certification, signaling that agent deployment in Korean enterprise is moving from experimental to compliance-regulated.

The Local Agent Ecosystem

SKT, Naver, and Kakao are all building their own agent ecosystems. But the Conway-vs-GPT-5.6 showdown matters directly because Korean enterprise AI stacks increasingly sit on top of global foundation models. The choice between an always-on persistent agent framework and a raw-power model-first approach will ripple into Korean corporate IT architecture within 12 months.

For Korean IT professionals specifically, the shift from “prompting skills” to “agent platform skills” is accelerating. Understanding webhook-driven architectures, agent lifecycle management, and multi-tool orchestration is becoming more valuable than knowing how to write better prompts.

FAQ

Q. What is Conway and how does it differ from existing AI assistants?

A. Conway is Anthropic’s always-on persistent agent that operates in a background sidebar with webhook triggers, browser control, and a proprietary .CNW configuration format. Unlike traditional chatbots that respond only when prompted, Conway monitors events and executes tasks autonomously. It represents a shift from prompt-response to event-driven agent architecture.

Q. When will GPT-5.6 be released?

A. GPT-5.6 has not been officially announced by OpenAI, but canary-tested codenames (ember-alpha, beacon-alpha) appeared in Codex logs in May 2026. Polymarket prices the June 8-14 release window at 46.5% probability. According to analyst Andrew Curran, the model is “ready now” but held for safety testing under upcoming executive order regulations.

Q. How does the AI agent platform war affect career planning?

A. The $10.91 billion AI agent market growing at 43% year-over-year means that agent platform skills are replacing prompt engineering as the key competency. Understanding webhook-driven architectures, agent lifecycle management, and multi-tool orchestration is becoming more career-critical than model-specific prompt techniques, especially as 80% of enterprises now run at least one production AI agent.

Q. Why did Anthropic surpass OpenAI in valuation?

A. Anthropic’s $965 billion IPO valuation surpassed OpenAI primarily due to revenue trajectory. Anthropic grew from $87 million ARR in January 2024 to $30 billion by April 2026, demonstrating an 80x growth rate. Meanwhile, OpenAI, despite $25 billion ARR and 50 million consumer subscribers, loses $1.22 for every $1 in revenue, with positive cash flow not expected until 2030.

Disclaimer: This article is for informational purposes only and does not constitute investment advice. The valuations and financial projections discussed are based on publicly available reports and may change. Consult a qualified financial advisor before making investment decisions.