Anthropic rationed a rival coding tool's Claude capacity while shipping lower-refusal models to national security customers

WEEK OF MAY 31 – JUN 6, 2025

Anthropic rationed a rival coding tool's Claude capacity while shipping lower-refusal models to national security customers

Windsurf got under five days' notice that nearly all its first-party Claude capacity was gone. Two days later Anthropic announced Claude Gov, which refuses less on classified work — refusal behavior is now tuned per customer tier.

Windsurf CEO Varun Mohan said on June 3 that Anthropic cut off nearly all of Windsurf's first-party capacity for Claude 3.7 Sonnet and Claude 3.5 Sonnet with less than five days of notice, pushing the coding-tool startup onto third-party inference providers and a bring-your-own-API-key workaround that is more expensive and more complex for its users. Mohan said Windsurf had offered to pay for the full capacity; Anthropic's Steve Mnich replied that the lab is prioritizing capacity for sustainable partnerships that allow it to effectively serve the broader developer community. The cut followed Anthropic's May 22 Claude 4 launch, where Windsurf received no direct access while Cursor, Cognition's Devin and GitHub Copilot appeared to have it.

Two days later Anthropic announced Claude Gov: a custom model set for US national security customers that refuses less when engaging with classified information, handles intelligence and defense documents better, and is limited to classified environments. The contrast inside one week is the lesson rather than the irony — model access is a commercially negotiated allocation you can lose on five days' notice, and refusal behavior is a per-deployment configuration rather than a fixed property of the brand.

Elsewhere the week was about provenance and integration surface. Reddit sued Anthropic over more than 100,000 alleged scrapes, OpenAI made MCP a first-party connector surface inside ChatGPT business plans, and researchers argued DeepSeek's new R1 was probably distilled from Gemini outputs.

THU · Jun 5, 2025anthropicpolicyguardrailsgovernment

Anthropic ships Claude Gov: lower-refusal models for classified national-security work

Anthropic announced Claude Gov on Thursday, a custom model set for U.S. national security customers that refuses less when engaging with classified information, handles classified material better, has greater understanding of intelligence and defense documents, and shows enhanced proficiency in languages and dialects critical to national-security operations. Anthropic says the models are already deployed by agencies at the highest level of U.S. national security, are limited to classified environments, and underwent the same safety testing as commercial Claude.

It is the direct counterpart to OpenAI's ChatGPT Gov from January 2025, and follows Anthropic's November 2024 Palantir and AWS defense partnership. It also lands two days after the same lab cut off a third-party coding tool's Claude capacity, making the week's contrast explicit: refusal behavior is tuned per customer tier.

Anthropic ships Claude Gov: lower-refusal models for classified national-security work — Anthropic

WHY IT MATTERS

Guardrails have become a per-deployment configuration rather than a fixed property of a model, which matters when you pick a provider on the assumption that refusals are stable. If your product depends on particular safety behavior, verify it on the exact deployment path you buy — API versus partner cloud versus managed tier — and keep your own policy layer for the actions your software must never take, independent of the vendor's defaults.

Anthropic: Claude Gov models for U.S. national security customers The Verge: Anthropic launches new Claude service for military and intelligence use Ars Technica: Anthropic releases custom AI chatbot for classified spy work

WED · Jun 4, 2025legalredditanthropicscrapingdata licensing

Reddit sues Anthropic over 100,000+ scrapes after Anthropic said its bots were blocked

Reddit filed in San Francisco Superior Court on Wednesday alleging Anthropic accessed its platform more than 100,000 times since July 2024 — after Anthropic publicly said in mid-2024 that it had blocked its crawlers. The complaint describes Anthropic as the white knight of the AI industry that is anything but, and Reddit's chief legal officer Ben Lee said the alleged commercial exploitation of Reddit content, including deleted posts, could be worth billions.

Reddit licensed its corpus to Google in February 2024, in a deal Bloomberg reported at roughly $60 million a year, which is the pricing baseline the suit implicitly measures Anthropic against. The claims are state contract and tort counts rather than copyright, following the August 2024 authors' class action against Anthropic; Anthropic said it disagrees with the claims and will defend itself vigorously.

Reddit sues Anthropic over 100,000+ scrapes after Anthropic said its bots were blocked — platform.theverge.com

WHY IT MATTERS

Platform terms are hardening into a second licensing layer under the training stack, and the enforceable risk is moving into contract and ToS claims rather than copyright alone. If any of your pipelines ingest third-party user-generated content, record crawl provenance and identity accurately, honor crawler blocks, and make sure deletion semantics propagate into any snapshot you fine-tune or embed from — an unlogged scrape is an unpriced liability.

The Verge: Reddit sues Anthropic, alleging its bots accessed Reddit more than 100,000 times since last July Reddit's docket-stamped complaint (S.F. Superior Court)

WED · Jun 4, 2025openaimcpagent infrastructureenterprise

OpenAI makes MCP a first-party surface inside ChatGPT business plans

In its June 4 business-plans update, OpenAI shipped connectors that bring internal tools and content — Google Drive, SharePoint and other cloud apps — into ChatGPT, alongside custom connectors built on the Model Context Protocol, a new record mode, and more flexible pricing. The same release notes confirm connectors landed in beta across Business, Enterprise and Edu workspaces, with OpenAI pointing developers at MCP for wiring in their own internal sources.

It makes MCP the integration unit for enterprise ChatGPT roughly six months after Anthropic open-sourced the protocol, and it arrives in the same week Anthropic was rationing a rival's Claude capacity. OpenAI states that information accessed from connectors is not used to train models by default in those workspaces — a policy line that determines which internal data can be exposed at all.

WHY IT MATTERS

An MCP server is now the lowest-friction integration surface for reaching ChatGPT's business users, and the merge between connector and MCP server means your internal tool needs admin-gated installation, scoped auth and an audit trail before it ships. Design the server as a product: read-only by default, per-tool permissions, and explicit approval semantics for any action with an effect outside the chat.

OpenAI: updates to ChatGPT business plans OpenAI Help: ChatGPT Enterprise and Edu release notes (connectors in beta)

TUE · Jun 3, 2025anthropicmodel accessdev toolingsupply risk

Anthropic cuts Windsurf's first-party Claude capacity with under five days' notice

Windsurf CEO Varun Mohan said on Tuesday that Anthropic cut off nearly all of Windsurf's first-party capacity for Claude 3.7 Sonnet and Claude 3.5 Sonnet with less than five days of notice, pushing the coding-tool startup onto third-party inference providers and a bring-your-own-API-key workaround that is more expensive and more complex for its users. Mohan said Windsurf had offered to pay for the full capacity; Anthropic spokesperson Steve Mnich replied that the lab is prioritizing capacity for sustainable partnerships that allow us to effectively serve the broader developer community.

The cut follows Anthropic's May 22 Claude 4 launch, where Windsurf received no direct access while Cursor, Cognition's Devin and Microsoft's GitHub Copilot appeared to have it. Windsurf reported $100M ARR in April and is reportedly being acquired by OpenAI, at the same time Anthropic has been building its own competing coding product, Claude Code.

WHY IT MATTERS

Model access is a commercially negotiated allocation, not a utility: a single-provider integration can be degraded on five days' notice by a lab that is also your competitor in the coding-tool market. Architect for multi-provider routing and first-class bring-your-own-key fallback now, and treat per-token cost divergence between first-party and third-party inference as a real line item in your margin model rather than an edge case.

TechCrunch: Windsurf says Anthropic is limiting its direct access to Claude AI models

TUE · Jun 3, 2025deepseekopen weightsdistillationdata provenance

Researchers say DeepSeek's new R1 was likely distilled from Gemini outputs

Developer Sam Paech published evidence that DeepSeek-R1-0528, released the prior week, favors vocabulary and phrasing that tracks Google's Gemini 2.5 Pro, and AI2 researcher Nathan Lambert said he would do the same in DeepSeek's position: they're short on GPUs and flush with cash. TechCrunch reported the evidence is not conclusive, while noting Google has started summarizing model traces in AI Studio to make them harder to train on.

DeepSeek did not disclose the data behind R1-0528, and its own model card positions the update as approaching o3 and Gemini 2.5 Pro on maths and coding. Distillation accusations are not new — developers flagged in December 2024 that DeepSeek V3 kept identifying itself as ChatGPT — but OpenAI's terms of service explicitly prohibit using model outputs to build a competing AI.

WHY IT MATTERS

Provenance is now a compliance field, not trivia: open weights distilled from a competitor's API traces carry output-use restrictions you cannot audit from the weights alone. If an open model sits in a regulated or commercially sensitive product, you need model cards, licenses and upstream output-use terms documented before deployment — not after procurement signs.

TechCrunch: DeepSeek may have used Google's Gemini to train its latest model