跪拜 Guibai
← All articles
Artificial Intelligence · ChatGPT · Claude

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

By 米小虾 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A tiered access regime for frontier models is now operational reality, not speculation: the strongest AI is gated behind government audit, and ordinary users train models they cannot use. Meanwhile, Claude Tag's Slack integration makes platform lock-in about accumulated workflows and institutional memory, not model quality — switching models is easy, but leaving behind years of AI-mediated decisions and task histories is not.

Summary

The pace of major AI releases has accelerated to one every 36 hours, and the first week of July 2026 delivered a cluster of events that redraw the industry's boundaries. OpenAI shipped GPT-5.6 in three tiers — Soul, Terra, and Luna — but the flagship Soul model was immediately restricted by the US government pending a cybersecurity audit, a first for a frontier model. Days later, Anthropic launched Claude Sonnet 5 with a 1M-token context window and agent capabilities, alongside Claude Tag, a persistent AI colleague embedded directly in Slack channels with hard architectural permission boundaries.

Behind the software headlines, the hardware landscape is fragmenting fast. Anthropic entered early-stage development of its own AI chips on Samsung's 2nm process, joining OpenAI (whose Jalapeño ASIC was partially designed by AI), Google, and Microsoft in the race to reduce dependence on NVIDIA. IBM demonstrated sub-1nm transistor integration, and South Korea committed $576 billion to AI infrastructure.

The week also saw SpaceX's xAI division acquire Cursor parent Anysphere for $60 billion in stock, while Anthropic CEO Dario Amodei publicly called for mandatory global regulation of frontier AI. The through-line across all these moves: AI is shifting from a pure technology sprint into a contest of ecosystem control — who owns the chips, who sets the access rules, and whose platform holds the institutional memory.

Takeaways
GPT-5.6 Soul scored 91.9% on Terminal-Bench 2.1 in Ultra mode, but the US government blocked public access pending a cybersecurity audit — the first pre-release restriction on a frontier model.
Claude Sonnet 5 ships with a 1M-token context window, agent tool-use capabilities, and performance approaching Opus 4.8, at a lower price through August 31.
Claude Tag is a persistent AI agent embedded in Slack with hard architectural permission boundaries: it can see contracts in legal channels and edit code in engineering channels, but cross-channel access is structurally impossible.
Anthropic reports that 65% of its own product team's code is already written through Claude Tag workflows.
Anthropic has started early-stage development of its own AI chips, targeting Samsung's 2nm process, joining OpenAI, Google, Microsoft, and Amazon in the custom-silicon race.
OpenAI's Jalapeño inference ASIC was designed in nine months with AI assistance and delivers industry-leading performance-per-watt; Microsoft has secured roughly 40% of capacity.
IBM integrated nearly 100 billion transistors onto a fingernail-sized chip, breaking the 1nm process barrier.
SpaceX's xAI division is acquiring Cursor's parent company Anysphere for $60 billion in stock.
South Korea announced a $576 billion AI investment plan.
Anthropic CEO Dario Amodei publicly called for mandatory global government regulation of frontier AI models.
Conclusions

Government pre-release auditing of frontier models over cybersecurity capability is now a repeatable pattern — GPT-5.6 Soul and Claude Fable 5 both hit the same gate within weeks of each other, suggesting a de facto regulatory regime has already formed.

The Claude Tag permission model solves cross-channel data leakage at the architecture level rather than through configurable policies, which makes the security guarantees auditable and hard to misconfigure — a design choice that enterprise compliance teams will notice.

Anthropic's 65% internal adoption figure for Claude Tag code generation is a signal about where software development workflows are heading: AI isn't assisting developers in a chat window; it's a persistent team member with its own credentials and task history.

The platform lock-in argument has shifted. The moat is no longer model quality but institutional memory — years of AI-mediated decisions, failed approaches, and workflow patterns embedded in a specific agent platform that cannot be exported.

Every major AI lab is now designing its own inference silicon, which fragments the hardware supply chain away from NVIDIA's dominance and toward a multi-foundry landscape split across TSMC, Samsung, and in-house fabrication.

OpenAI using AI to design its own chip in nine months closes a loop that has strategic implications: AI accelerating the hardware that accelerates AI.

Concepts & terms
Terminal-Bench 2.1
A programming benchmark that evaluates AI models on real-world software engineering tasks executed in a terminal environment, measuring their ability to write, debug, and deploy code autonomously.
1M Token Context Window
The amount of text an AI model can hold in active memory during a single interaction — roughly 750,000 words or the length of the entire Lord of the Rings trilogy — enabling it to process very long documents without summarization or chunking.
2nm Process
A next-generation semiconductor manufacturing node that packs transistors at roughly 2-nanometer scale, offering significant power efficiency and density improvements over current 3nm and 4nm processes used in leading-edge chips.
ASIC (Application-Specific Integrated Circuit)
A chip custom-designed for a single purpose — in this context, AI inference — as opposed to general-purpose GPUs. ASICs trade flexibility for dramatically better performance-per-watt on their target workload.
Agent Model
An AI model capable of autonomous multi-step task execution: it can plan, use tools (browsers, terminals, APIs), verify its own output, and iterate toward a goal without human prompting at each step.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗