PromptArmor：供应商接入 Fireworks 等推理商后往往跟进 Kimi、DeepSeek、GLM 等国际模型

Your vendors

Glean

Harvey

Zendesk

Which of your vendors are adopting new inference hosts and international models?

Vendors are turning to international AI models

Vendors with AI features are increasingly turning to international and open-weight models due to the cost difference compared to models from leading U.S. labs, and because these international models are becoming increasingly competitive on benchmarks.

Recently, Kimi K3 and DeepSeek V4 Flash 0731 have made the news, offering competitive performance at a steep discount compared to models like Anthropic's Fable and OpenAI's GPT Sol.

International vs U.S. model token costs

International model

U.S. model

Input, then output

DeepSeek

V4 Flash

in $0.14

DeepSeek

V4 Pro

in $0.435

Z.ai

GLM 5.2

in $1.40

Mistral

Medium 3.5

in $1.50

OpenAI

GPT-5.6 Terra

in $2.00

Moonshot

Kimi K3

in $3.00

Anthropic

Sonnet 5

in $3.00

Anthropic

Opus 5

in $5.00

OpenAI

GPT-5.6 Sol

in $5.00

Anthropic

Fable 5

in $10.00

International

U.S.

DeepSeek V4 Flash

In

$0.14

Out

$0.28

DeepSeek V4 Pro

In

$0.435

Out

$0.87

Z.ai GLM 5.2

In

$1.40

Out

$4.40

Mistral Medium 3.5

In

$1.50

Out

$7.50

OpenAI GPT-5.6 Terra

In

$2.00

Out

$12.00

Moonshot Kimi K3

In

$3.00

Out

$15.00

Anthropic Sonnet 5

In

$3.00

Out

$15.00

Anthropic Opus 5

In

$5.00

Out

$25.00

OpenAI GPT-5.6 Sol

In

$5.00

Out

$30.00

Anthropic Fable 5

In

$10.00

Out

$50.00

USD per million tokens. Provider list prices, read August 9, 2026. Sources: DeepSeek, Z.ai, Mistral, Moonshot, OpenAI, Anthropic.

Furthermore, open-weight models do not impose API-level restrictions on security and biology work, unlike leading U.S. model providers.

More on How Hugging Face Turned to Zhipu due to Claude's Guardrails

Hugging Face's incident responders were blocked from analyzing their own intrusion because the attack commands in the logs tripped Anthropic's safety guardrails. They moved to Zhipu AI's GLM 5.2 to complete their investigation.

While organizations are not all aligned on the potential for risks such as model backdoors, most organizations are in agreement that data processing for AI inference must remain within the U.S.

Because of this, vendors that want to add international or open-weight models turn to inference providers whose job it is to run those models within U.S. data residency.

Inference provider changes predict model changes

Across the vendors we monitor for our customers, we see AI subprocessor changes frequently. From new features that come with new models to updating features with the latest models, subprocessor and model changes are common occurrences. Throughout these changes, there have been some notable trends.

One sequence we have seen play out again and again is that shortly after a vendor adds a US-based inference provider like Fireworks, Baseten, Together, Groq, or Cerebras, they begin rolling out open-weight or international models.

Below are samples of vendors we have seen take this path, and vendors who have recently added an inference provider but are yet to announce their more-than-likely function: supporting the adoption of international or open models.

Recent examples from vendors

Vendor

Subprocessor added

Date

Model added

Date

International models following new inference subprocessor

Glean

Fireworks

Jun 16, 2026

Z.ai GLM 5.2

Jul 16, 2026

Harvey

Baseten

Fireworks

Jul 3, 2026

Mistral Medium 3.5

Aug 5, 2026

Vercel

Cerebras

Baseten

Jan 29, 2026

Z.ai GLM 5.1

Apr 8, 2026

GitHub Copilot

Fireworks

Aug 29, 2025

Moonshot Kimi K3

Aug 6, 2026

New inference subprocessor, new models coming soon

Sierra

Baseten

Fireworks

Together

Modal

Jun 3, 2026

Coming soon

Zendesk

Groq

Fireworks

Baseten

Jul 17, 2026

Groq, Dec 2025; Fireworks, Mar 2026; Baseten added and Fireworks broadened to hosting and inference, Jul 17, 2026

Coming soon

Ada

Groq

Baseten

Jun 2026

Groq, Jan 2026; Baseten, Jun 2026

Coming soon

Deepgram

Baseten

May 2026

Coming soon

International models following new inference subprocessor

Glean

Subprocessor added

Jun 16, 2026

Fireworks

Model added

Jul 16, 2026

Z.ai GLM 5.2

Harvey

Subprocessor added

Jul 3, 2026

Baseten

Fireworks

Model added

Aug 5, 2026

Mistral Medium 3.5

Vercel

Subprocessor added

Jan 29, 2026

Cerebras

Baseten

Model added

Apr 8, 2026

Z.ai GLM 5.1

GitHub Copilot

Subprocessor added

Aug 29, 2025

Fireworks

Model added

Aug 6, 2026

Moonshot Kimi K3

New inference subprocessor, new models coming soon

Sierra

Subprocessor added

Jun 3, 2026

Baseten

Fireworks

Together

Modal

Model added

Coming soon

Zendesk

Subprocessor added

Jul 17, 2026

Groq, Dec 2025; Fireworks, Mar 2026; Baseten added and Fireworks broadened to hosting and inference, Jul 17, 2026

Groq

Fireworks

Baseten

Model added

Coming soon

Ada

Subprocessor added

Jun 2026

Groq, Jan 2026; Baseten, Jun 2026

Groq

Baseten

Model added

Coming soon

Deepgram

Subprocessor added

May 2026

Baseten

Model added

Coming soon

More predictive alerts for AI in vendors

Beyond the identification of international and open-weight model use, we have observed a number of trends that can be extrapolated to inform decisions about AI in vendors. Here are a few:

• Cyber-capable model adoption. Anthropic added a “Covered Models” section to its Service Specific Terms on June 8, 2026, one day before launching Claude Fable 5 as a covered model on June 9, with supplemental terms that permit human safety review and explicitly supersede zero data retention commitments for their models that have advanced cyber capabilities. The upstream data processing requirements became a signal for the adoption of these models. For example, Glean introduced a Limited Retention Addendum on June 10, two days after Anthropic’s terms change, and added Claude Fable 5 within a month.

• AI feature expansion. Across industries, adoption of new tools and AI capabilities usually occurs in groups. For example, Skills first spread across coding agents, and then shortly after they spread across knowledge work. Once one major player in an industry adopts a trending AI capability, competitors are quick to adopt - so by tracking the first entrant in an industry, we can predict when other industry players will adopt. This is playing out right now with Legal AI; Legora introduced Skills on June 3, 2026; since then, Harvey and other competitors have begun rolling out Skills. This pattern has played out multiple times now, including with Memory, Web Search, MCP, and recently, 'Coworkers' and always-on-agents.

• Agentic platform access via nth party apps. As applications are increasingly accessed via AI connectors, vendors have begun to shift their posture to reflect the utilization of their services via third-party AI services, like Claude and ChatGPT. Vendors have begun to release terms changes that indicate when and how users are liable for access to their services via third-party AI, such as MCP servers and connectors. In addition to legal changes, a common tendency is to have a staggered rollout of ChatGPT and Claude connectors: if we see a vendor become accessible via ChatGPT or Claude, we can infer that it will soon be accessible through the other.

Vendors are Deflecting Liability for Anything 'AI'

• Agentic scope expansion. Vendors love to describe their features as 'agents' and 'agentic', but it is not uncommon to find 'agents' with limited practical capabilities beyond a normal chatbot. In the process of reviewing these agents, we've come across a pattern: from the first time an 'agentic' system surfaces in marketing, it takes about six months to go from a 'chatbot with a specialized prompt' to true agentic, autonomous, or always-on capabilities (MCP, web browsing, recurring workflows, code execution, etc.). Monitoring 'agents' for this transition has become a priority for many organizations that have approved these 'agents' on the basis of the limited toolset they originally presented.

• Connector and integration releases. Recently, we released research that found that Claude and ChatGPT connectors change on average every 9 minutes. But the big vendors are not the only ones with tumultuous connector ecosystems. Across vendors of varying sizes with AI features, we’ve seen that the overhead to create one connector vastly lowers the prerequisites for future connectors. Shortly after we detect a vendor adding its first connector, they often enter a period of rapid growth in which dozens or hundreds of connector options are added, before the increase tapers back down and lands at a steady rate.

Claude and ChatGPT Connectors Change Every 9 Minutes

We monitor tens of thousands of vendors across industries, which has enabled us to convert insights across tools into predictive alerts. This helps our customers stay ahead of the curve as their vendors expand and change their AI posture.