Anthropic's moat has shifted from "model scores" to compound lock-in across the protocol, interface, and application layers—three protocol-level moves in MCP, Computer Use, and Mythos are harder to overtake than a 0.5 percentage point SWE-bench improvement every week.
Hybrid Reasoning was the right bet—Anthropic rejected OpenAI's "standalone reasoning model" path; a year later, OpenAI merged reasoning and standard modes in the GPT-5 series, essentially adopting Anthropic's judgment
MCP donated to Linux Foundation (2025-12-09)—trading "single-company governance" for "de facto standard status"; monthly downloads still growing after OpenAI, Google, and Microsoft fully integrated
Mythos autonomously discovered thousands of zero-days in weeks—a 27-year OpenBSD zero-day, a 23-year Linux zero-day—the White House went from fear to procurement in about 3 weeks
The SWE-bench gap between Opus 4.7 and GPT-5.5 has compressed to 1.1 percentage points + 7 days—the window for model capability as a differentiator is closing
"Highest model score" is failing as a sales pitch—in early 2024, the gap between Claude 3 and GPT-4 was about 6 months; by mid-2026, the gap between Anthropic and OpenAI has compressed to about 7 days. What truly determines moat depth is the compound lock-in across three layers—protocols (MCP), interfaces (Computer Use / Agent SDK), and applications (Claude Code / Cowork)—not who leads the next benchmark.
2025 Q1 was the first watershed (3.7 Sonnet + Claude Code), 2025 Q3–Q4 was the second (Opus 4.5 broke 80%), and 2026 Q1–Q2 was the third (4.6/4.7 combined with Cowork platformization). Each watershed was not just a "score change"—it was accompanied by new product forms emerging from the model layer.
When Claude 3.7 Sonnet launched in February 2025, Anthropic made a choice with far-reaching implications for the industry—rejecting OpenAI's "o1/o3 standalone reasoning model" path in favor of "hybrid reasoning": a single model that can both answer normally and reason deeply, with developers setting a thinking budget via the API (up to 128K tokens).
"Reasoning should be an integrated capability of frontier models, not a standalone model."
The Extended Thinking framework went through four generations of evolution: 3.7 Sonnet (binary enable) → Claude 4 (Interleaved thinking, can call tools during reasoning) → Opus 4.5 (Effort parameter low/medium/high) → Opus 4.6/Sonnet 4.6 (Adaptive thinking, model self-determines whether to think deeply) → Opus 4.7 (xhigh effort + Task Budgets).
OpenAI's o-series was forced to merge "reasoning + standard" into the GPT-5 series by late 2025—essentially adopting Anthropic's judgment. This was a battle Anthropic won at the level of product philosophy.
On November 25, 2024, Anthropic released the Model Context Protocol (MCP), based on JSON-RPC 2.0, likened by the industry to "USB-C for AI." It solved the traditional N×M integration dilemma—every LLM application had to write a separate connector for each data source; unification at the protocol layer reduced this matrix to N+M.
Key adoption milestones clustered in mid-2025: 2025-03-26 OpenAI fully integrated, Sam Altman publicly stated "People love MCP"; 2025-04 Google DeepMind integrated; 2025-05 Microsoft fully integrated into Copilot Studio / Semantic Kernel / Azure OpenAI. The spec gradually absorbed enterprise requirements: OAuth 2.1 + PKCE, RFC 8707 Resource Indicators, CIMD replacing dynamic client registration.
The most critical step came on December 9, 2025: Anthropic donated MCP to the newly established Agentic AI Foundation (AAIF) under the Linux Foundation—co-founded by Anthropic, Block, and OpenAI, with support from AWS, Google, Microsoft, Cloudflare, and Bloomberg. This was the first time OpenAI and Anthropic co-founded a governance entity at the protocol layer.
Growth metrics are undeniable: monthly SDK downloads of approximately 97 million (within 16 months), 10,000+ officially registered servers (community estimates 16,000–18,000+). In May 2026, Anthropic密集 launched 8 official connectors in creative domains (Blender, Autodesk Fusion, Ableton Live, Adobe Creative Cloud, SketchUp, etc.), marking MCP's ecosystem leap from "DevOps toolset" to "cross-professional domain operating system."
The strategic implication of the donation: Anthropic traded "relinquishing single-company governance" for "de facto standard status." When MCP was no longer Anthropic's protocol, everyone became more willing to adopt it—yet the core code, maintainers, and documentation ecosystem are still led by Anthropic. Protocol donation ≠ influence relinquishment.
When everyone is reading your code and calling your protocol,
your design patterns become the de facto standard. — Anthropic Panorama Series · Model Lineage & Protocols
When Claude 3.5 Sonnet v2 launched in October 2024, Anthropic did something unexpected—the industry's first commercial computer use. Claude could view screenshots, click the mouse, type on the keyboard, and operate standard desktop applications without dedicated API integration. Over 18 months, the OSWorld score went from 22.0% (3.5 Sonnet v2) to 61.4% (Sonnet 4.5) to 72.5% (Sonnet 4.6)—a 3.3× increase.
Product derivatives along this curve: Claude in Chrome (2025-08), Claude Cowork (2026-01, desktop), Claude for Excel (2026-Q1 beta, vertical office). Computer Use is the key technology for Anthropic to cross the "API → application layer" threshold—without it, Anthropic could only build a better chatbot; with it, Anthropic can directly "take over" applications within the user's work environment.
In April 2026, Anthropic released Mythos Preview—a research preview open only to 11 cybersecurity/critical infrastructure partners (AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks). Capability metrics shook the industry: autonomously discovered thousands of zero-day vulnerabilities in weeks—a 27-year OpenBSD zero-day, a 23-year Linux zero-day, a 17-year FreeBSD zero-day—single exploit cost $50–$2,000, zero human intervention.
Anthropic's official framing was "too dangerous to release"—this narrative is itself marketing: isolating Mythos from the "regular LLM" customer base, pricing at $20K/scan, directly benchmarked against Palantir / Booz Allen / Mandiant B2G tools. The subsequent Project Glasswing was the real strategic move: 12 core partners + 40 expanded access + $100M usage credits.
The White House went from fear to procurement in about 3 weeks: Mythos release + Cybench 100% saturation → Treasury Secretary and Federal Reserve urgently convened Wall Street CEOs → 2026-04-16 White House opened Mythos to federal agencies. The pivot logic was simple: restricting government use of Mythos wouldn't make vulnerabilities disappear; it would only leave government defense capabilities behind attackers who could already use similar tools.
Every API upgrade was accompanied by a price reduction: Opus 4.5 cut pricing from $15/$75 to $5/$25 (a 67% reduction). This is classic AWS economics—when unit costs decline faster than volume grows, gross margins actually expand. Anthropic pushes model capability past SWE-bench 80% on one hand, while pricing lets mid-market enterprises dare to use Opus on the other, simultaneously compressing inference costs through in-house Trainium / TPU chips—the scissors spread between the two is the profit margin.
In June 2025, the Anthropic engineering blog published a widely cited study: multi-agent systems outperformed single-agent Claude Opus 4 by 90.2% on research tasks. The key pattern was Orchestrator-Worker: Opus 4 as the "leader" decomposes tasks, while multiple Sonnet 4 instances act as subagents executing in parallel. The earlier December 2024 "Building effective agents" defined five core patterns widely referenced across the industry: Prompt chaining, Routing, Parallelization, Orchestrator-Workers, and Evaluator-Optimizer.
Paired with the Harness Engineering series (Initializer Agent + Coding Agent + Generator-Evaluator separation), Anthropic fully externalized its engineering team's internal experience as industry best practices. The key insight is counterintuitive:
"Having an Agent evaluate its own output" is one of the primary sources of Agent quality degradation.
Hence the need for Generator + Evaluator separation, letting an independent assessor rather than the generator self-evaluate quality—this view became near-consensus in multi-agent design by the end of Q1 2026.
Opus 4.7 SWE-bench Verified 87.6%, but OpenAI GPT-5.5 is already at 88.7% (released 7 days later). In early 2024, the gap between Claude 3 Opus and GPT-4 was about 6 months; by mid-2026, the gap between Anthropic and OpenAI has compressed to about 7 days—the benchmark score differential among top-tier models has been compressed to single-digit percentage points.
But Anthropic doesn't compete on "highest model scores"—it moves value up three layers: Protocol layer (MCP, which became a de facto standard after being donated to the Linux Foundation), Interface layer (Computer Use / Agent SDK, providing the "OS interface" for models), Application layer (Claude Code / Cowork, directly facing end-user workflows)—these three layers have higher barriers and greater switching costs than the "model layer." Once an enterprise integrates MCP into Claude and installs Claude Code into its development workflow, the cost of switching at the model layer is amplified layer by layer.
Anthropic has moved past the era of "winning on model scores"—it is betting that the compound lock-in of stacked protocol, interface, and application layers is harder to overtake than simple model leadership.
If this judgment is right, the winner of the next decade will be the company that establishes the operating system position, not the one posting new model scores every week—this is also the main thread that the next article (Safety Moat & OS-ification) will unfold.
Whoever controls the road network decides how the next generation of AI applications runs.
First published 2026-07-15