跪拜 Guibai
← Back to the summary

GPT-5.6 Gets a Government Gatekeeper, Claude Moves Into Slack, and the Chip War Heats Up

July 2026 AI Earthquake: GPT-5.6 Restricted by Government, Claude Embedded in Slack, Anthropic Develops Own Chips

One-sentence summary: In July 2026, the AI industry is shifting entirely from a "technology race" to an "ecosystem game" — three main threads of regulation, hardware, and agentification are erupting simultaneously.

1. What Happened in the AI World This July?

If you thought AI development in the first half of 2026 was already fast, July will completely reset your understanding.

The cycle for major AI updates has compressed from once per quarter in 2023 to once every 1.5 days in 2026. You read that right — not every month, but a major update every 36 hours.

And the first few days of July 2026 were unprecedentedly dense:

This is not a series of isolated events, but a complete restructuring of the AI landscape. This article will sort through the logic behind each event and how they collectively define the turning point of the AI industry in 2026.

2. GPT-5.6: The Strongest Model Ever, "Blocked" by the Government

Three-Tier Architecture, Clear Positioning

OpenAI did not take the "one model rules all" route this time, but divided GPT-5.6 into three tiers:

  1. Soul — Flagship edition, the strongest model, and the one restricted by the government
  2. Terra — Balanced edition, performance on par with the previous flagship, at half the price
  3. Luna — Fast and cheap edition, suitable for large-scale high-frequency calls

On the programming benchmark Terminal-Bench 2.1, Soul scored 88.8% in standard mode, surpassing Claude Mythos 5's 88.0%; after enabling Ultra mode (using sub-agents to accelerate complex tasks), it reached 91.9%.

Why Was It Restricted?

There is one core reason: CyberSecurity.

GPT-5.6 Soul's ability to discover software vulnerabilities is so strong that the US government needs to audit it before release — a first in the history of AI development. Just a few weeks earlier, Claude Fable 5 experienced a similar fate.

This raises a question worth pondering: these models are trained on data (code, articles, ideas) left by all of humanity on the internet, but ultimately, the public who created this data cannot use the strongest AI they helped "feed."

A tiered "AI access rights" era has arrived. The most powerful models are only open to vetted partners; ordinary users can only use secondary versions.

3. Claude Sonnet 5 + Claude Tag: AI Is Really Starting to "Work"

Sonnet 5: The Strongest Agent Model in the Sonnet Series

On July 1, Anthropic released Claude Sonnet 5. Key features:

Claude Tag: Your AI Colleague Now Lives in Slack

If Sonnet 5 is a capability upgrade, Claude Tag is a paradigm disruption.

Claude Tag is an AI colleague permanently embedded in Slack. You @ it in any channel, give it a task —

"Help me set up auto-export"

It will pull files from Google Drive on its own, break down task steps, update documents, and then reply in the channel with a completion checklist.

Anthropic internal data: 65% of product team code is already written this way.

The most critical design lies in the permission model: Claude can see contracts in the legal channel and edit code in the engineering channel, but cross-channel access is architecturally impossible — not a toggle setting, but a hard boundary. Every credential use is logged. Claude has its own company account.

A former MIT professor's assessment is "Trojan horse" — not that Anthropic is doing anything bad, but that incentives are fundamentally distorted: you can switch AI models at any time, but the workflows, customer commitments, and "we tried that solution and it didn't work" lessons your company has accumulated are all deeply embedded in this one platform.

The lesson is clear: models can be swapped, but your company memory cannot.

4. AI Chips: From "Software Race" to "Hardware Land Grab"

Anthropic's Own Chips

Following OpenAI, Microsoft, Google, and Amazon, Anthropic has also joined the ranks of developing its own chips. The latest news shows that Anthropic has started early development of AI chips and is in talks with Samsung Electronics for 2nm process manufacturing cooperation.

What does this mean? The foundation of AI computing power is undergoing a major reshuffle:

Company Chip Strategy Foundry
OpenAI Jalapeño ASIC (in-house, design completed in 9 months) Partnering with Broadcom
Anthropic In-house AI chip (early stage) Planning to use Samsung 2nm
Google TPU v7 In-house
Microsoft Maia 2 In-house
NVIDIA Blackwell Ultra / TwoTower TSMC

OpenAI's Jalapeño Chip

Interestingly, OpenAI even used AI to help design its own chip. Jalapeño is a custom inference chip with industry-leading performance-per-watt. Microsoft has already secured about 40% of the capacity.

IBM Breaks Below 1nm

IBM demonstrated another path: successfully integrating nearly 100 billion transistors onto a fingernail-sized chip, breaking through the technical barrier of the 1nm process.

South Korea's $576 Billion Investment

The South Korean government announced a $576 billion AI investment plan.

5. Industry M&A: SpaceX xAI Acquires Cursor for $60 Billion

SpaceX's xAI division will acquire Cursor's parent company Anysphere for $60 billion in stock.

6. Agentification: The Qualitative Change from "Can Chat" to "Can Do"

Multi-Model Orchestration Becomes Mainstream

Japan's Sakana AI's Fugu series adopts a completely different approach — letting models work in teams.

AI Embedded in Productivity Tools

7. Global AI Governance: The Era of Self-Regulation Is Over

Anthropic CEO Dario Amodei published "Exponential AI Policy," publicly calling on global governments to implement mandatory regulation of frontier AI.

8. Summary

  1. From scale competition to efficiency competition
  2. From single model to model ecosystem
  3. From cloud to edge
  4. From conversation to autonomous action
  5. From self-regulation to external regulation

This article is based on the latest public information as of July 2026