Anthropic Blocks Third‑Party Access—How Xiaomi’s MiMo Launches a Silent Counterstrike

1. Anthropic Closes the Door on Third‑Party Tools

On April 4, Anthropic’s Claude Code lead Boris Cherny announced on the official blog that, starting April 5, Claude subscribers would no longer be able to consume their quota through third‑party integrations such as OpenClaw. Users must purchase an additional pay‑as‑you‑go package to keep using those integrations.

OpenClaw, an open‑source AI‑Agent framework with more than 310 k GitHub stars, had previously allowed developers to “piggyback” on Claude’s subscription for free; the ban removed that core usage path.

Boris Cherny explained that the subscription model was not designed for the usage patterns of third‑party tools and that capacity must be managed for customers who use Anthropic’s own products and API.

Anthropic offered transitional compensation: a one‑time credit equal to a month’s fee (usable before April 17), up to a 30 % discount on pre‑purchased quota packs, and a full‑refund request via email.

2. The “Buffet Dilemma” Behind Third‑Party Tools

The article cites The Decoder’s analogy that third‑party agents behave like “sumo wrestlers in a buffet”: inefficient context handling causes API request volumes that are several times higher than native Claude usage, making a fixed‑price subscription unsustainable under continuous Agent‑driven workloads.

Core contradiction : Chat‑era pricing estimates cost per dialogue turn, which is predictable. In the Agent era, a single task can trigger dozens of calls, repeatedly filling and refreshing the context window, causing token consumption to grow exponentially—comparable to buying an all‑you‑can‑eat meal for a fixed monthly fee.

The shift from Chat to Agent paradigms is the underlying reason Anthropic acknowledges the need to restrict third‑party access.

3. Luo Fuli’s “Silent Strike” Commentary

On April 5, the same day the ban took effect, Xiaomi’s MiMo large‑model lead Luo Fuli posted a lengthy comment on X (Twitter), describing the event as a “silent strike” caused by the rapid Chat‑to‑Agent transition.

She organized her analysis into five points:

Subscription design : Claude’s internal compute allocation is sound, but third‑party tools’ inefficient token usage makes it unprofitable.

Cost reality : Tools like OpenClaw generate API request volumes that are “multiple times” higher than native usage, inflating costs.

Market pressure : Short‑term cost increases will force third‑party tools to improve token efficiency and caching.

Industry direction : She opposes unsustainable price wars, emphasizing that “long‑term stable delivery of high‑quality models” outweighs short‑term low pricing.

Future vision : Success in the Agent era depends on “co‑evolution” of efficient tools and powerful models rather than simply consuming more compute.

4. OpenClaw Founder’s Response

Peter Steinberger, OpenClaw’s founder (now at OpenAI), acknowledged Anthropic’s move as “sad for the ecosystem” because Anthropic first incorporated popular open‑source features into its closed system before blocking the alternatives. He noted that Boris Cherny’s transitional measures were not a blunt “kill‑switch.”

OpenClaw remains a high‑profile open‑source project with over 310 k GitHub stars, 58 k forks, and 1 200+ contributors; losing Claude quota support is a significant blow.

5. MiMo Token Plan – A Precise Market Entry

One day before Anthropic’s announcement (April 3), Xiaomi’s MiMo platform launched a Token Plan that explicitly supports all major third‑party tools, including OpenClaw, OpenCode, Kilocode, Cline, and Roocode.

The plan offers four overseas pricing tiers (MiMo‑V2‑Pro, ‑Omni, ‑Flash, ‑TTS) with per‑million‑token rates ranging from $0.10 to $1.00 for input tokens and $0.30 to $3.00 for output tokens, plus domestic monthly subscriptions starting at ¥39 for four tiers.

On launch day the platform processed over 1 trillion tokens, demonstrating rapid adoption.

6. MiMo‑V2‑Pro’s Competitive Edge

The MiMo‑V2‑Pro model, released under the “Hunter Alpha” alias on OpenRouter, was initially mistaken for DeepSeek V4. It later proved to be Xiaomi’s own model.

Key metrics:

Parameter count: >1 trillion total, 42 billion active MoE parameters.

Context window: 1 million tokens.

PinchBench: 81.0 (second only to Claude Opus 4.6 = 81.5).

ClawEval: 61.5 (close to Claude Opus 4.6 = 66.3).

OpenRouter market share: 22.3 % (rank 1).

7. Luo Fuli’s Personal Journey

Luo Fuli, born 1995, holds a B.Sc. in Computer Science from Beijing Normal University and an M.Sc. in Computational Linguistics from Peking University. She authored eight solo papers at ACL 2019, later joined Alibaba DAMO Academy to lead the multilingual pre‑training model VECO, and contributed to DeepSeek V2 before being recruited by Xiaomi in November 2025 as MiMo’s head.

Within five months, MiMo‑V2‑Pro was released, achieving top rankings on OpenRouter and embodying her “silent strike.”

8. Deeper Issue: AI Industry Business‑Model Conflict

The episode reveals a fundamental contradiction for the AI industry in the Agent era: Chat‑era products are “question‑answer machines” with limited token consumption per interaction, whereas Agent‑era products act as “automated workers” that plan tasks, read files, execute code, and repeatedly call APIs, potentially consuming millions of tokens per task.

A $20 monthly subscription that broke even in the Chat era becomes a loss‑making proposition in the Agent era.

Two competing routes :

Anthropic’s route : Closed ecosystem, prioritize native Claude experience, charge third‑party usage per API call – controls cost but reduces openness.

MiMo’s route : Open ecosystem, embrace all third‑party tools with competitive pay‑as‑you‑go pricing – aims to capture market share but requires strong cost‑control capabilities.

Luo concludes that success will belong to the side that balances model capability, token efficiency, and commercial sustainability.