Anthropic cuts Claude usage prices 40% for US government buyers

Overview

8 stories in this issue. The first 3 are today's priorities.

Hot Model Updates

1. Top · Anthropic cuts Claude usage prices 40% for US government buyers
2. Top · FlashMap speeds up Qwen3-8B decoding by as much as 6.14x with no loss in output

Global AI News

3. Top · Anthropic acquires Mithril Medicine to run drug programs with Claude
4. Inference cloud startup General Compute raises a $100M Series A
5. Meta quietly launches Waddle World, an AI pet journal app

Regional and Early Signals

6. Ant Group's Bailing open-sources Ling-3.1-flash-1019, scoring 44 on the Intelligence Index
7. Doubao Work can record a demonstration as a reusable skill and keep working while a Mac is locked
8. ByteDance denies training models on 150,000 film extras' likenesses

Jiufeng graphic based on the sources cited in this issue.

Hot Model Updates

Anthropic expanded its OneGov agreement with the US General Services Administration: government buyers get 40% off Claude usage, with no minimum purchase.

Anthropic announced the expanded GSA agreement on October 27th. The main terms:

• Discount: 40% off usage-based pricing for Claude models
• Commitment: minimum purchase commitments removed
• Eligible buyers: federal, state, local and tribal governments; Anthropic says territorial governments are also eligible
• Channels: buy directly from Anthropic, through resellers, or on major cloud platforms

Limitations: RuntimeWire notes that the deal comes after a year of disruption to Anthropic's federal purchasing status, and that the scale and durability of agency adoption have yet to be shown. The report gives neither the agreement's duration nor the undiscounted unit prices.

Image source: anthropic.

Source: Anthropic · RuntimeWire

Researchers from Polymathic AI, Princeton and other institutions released FlashMap, a speculative-decoding drafter that drafts a 16-token block in one forward pass. Drafter weights are available for Qwen3-8B and Qwen3-4B.

FlashMap comes from researchers at Polymathic AI, Princeton University, USC ISI, NYU and the Flatiron Institute. It replaces a discrete diffusion drafter with a continuous flow map, drafting a full 16-token block in a single forward pass. The tokens within each block stay consistent with one another.

• Drafter size: about 0.5B (500M for Qwen3-8B, 475M for Qwen3-4B)
• Deployment: runs on the same GPU as the target model; vLLM and SGLang plugins are provided
• Test setup: the paper's numbers use 1 NVIDIA H200 NVL at concurrency 1, in bf16
• Comparison: the fastest drafter tested on Qwen3-8B, about 7% above a retrained DFlash on average

Limitations: The 6.14x best case is GSM8K at temperature 0, with an average acceptance length of 7.69 tokens; the 2.28x worst case is Alpaca at temperature 1, so the two figures differ in both task and sampling temperature. Released drafter weights cover only Qwen3-8B and Qwen3-4B. The figures were measured on a single GPU at concurrency 1, and the source gives no numbers for high-concurrency serving.

Source: arXiv · Hugging Face · GitHub · MarkTechPost

Global AI News

Anthropic acquired the small biotech Mithril Medicine. Its founder will continue the company's muscular-dystrophy research, and the evidence so far is preclinical.

Anthropic said on October 27th that it acquired Mithril Medicine, founded by Cody Tranbarger. Tranbarger joins Anthropic to continue Mithril's muscular-dystrophy research and to start additional programs as part of Anthropic's drug-discovery work with Claude. RuntimeWire says the deal extends Anthropic from software partnerships into drug development that it operates directly.

Limitations: The report says the current evidence is preclinical, so the claim that one operator can run drug development with AI is only beginning to be tested. The report does not disclose the deal's price.

Source: RuntimeWire · Cody Tranbarger on X

General Compute raised $100 million in a Series A at an $800 million post-money valuation and plans to bring more than 75 MW of compute online next year.

Blackbird and Abstract Ventures led the round, with participation from Fuse Venture Partners, Village Global and other investors. General Compute builds cloud infrastructure for AI inference. According to SiliconANGLE:

• Capacity plan: more than 75 MW of compute coming online next year
• Customer contract: a signed deal worth more than $100 million in annual recurring revenue
• Earlier financing: $400 million in debt financing in July

Limitations: The first large-scale chip deployment won't come online until the first quarter of next year. SiliconANGLE also notes that Gartner and Elon Musk have both predicted a shakeout among neoclouds, and that CoreWeave, the largest neocloud, recently saw its stock fall to a 12-month low.

Source: SiliconANGLE

Meta released Waddle World, a pixel-art virtual pet app whose creatures chat with users and draw doodles to illustrate conversations. It is currently available only to adults in the Philippines.

According to its Apple and Play Store listings, Waddle World is a "16-bit pet care game", but the listing describes it as "a daily journal disguised as a pet." Users are encouraged to talk to their pets rather than type, and can play casual games such as stacking groceries and word games. The app is free. It was noticed by social media consultant Matt Navarra.

Limitations: Meta has not formally announced the app, and the report does not say which model powers it. The Verge says it isn't the only unusual app Meta has quietly launched recently.

Source: The Verge · Matt Navarra on X

Regional and Early Signals

Ant Group's Bailing team open-sourced Ling-3.1-flash-1019 and the base model Ling-3.1-flash-base, aimed at personal agent use cases.

On October 28th Bailing open-sourced Ling-3.1-flash-1019, which is further trained from Ling-3.1-flash. It scores 44 on the Artificial Analysis Intelligence Index v4.3.2, 3 points higher than the previous version. The base model released alongside it, Ling-3.1-flash-base, is available as BF16 weights under the MIT license. It uses a hybrid linear MoE architecture with about 560B total parameters and about 25B active per token. According to the company, it beat DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base on 32 of 47 benchmarks across 8 domains.

Limitations: Chinese-language source, single report. All scores are the company's own figures and have not been independently reproduced.

Source: QbitAI (Chinese-language source)

Doubao Work upgraded its computer-use features: users can save a demonstration as a reusable skill, and computer-use tasks keep running after the Mac locks.

Once a user grants permission, Doubao Work can view the screen and simulate mouse and keyboard input. A new "record actions to create a skill" option captures clicks, typed text and window contents, and users can add spoken instructions while recording. The recording is then turned into a personal skill. After users enable "allow Doubao to work while locked" in settings, tasks continue while the Mac is locked, and users can send tasks and check results from their phone.

Limitations: Chinese-language source, single report. The company says click accuracy and speed improved but gives no figures. Because recordings capture window contents, users are told not to show passwords or other sensitive information.

Source: Leiphone (Chinese-language source)

Liu Xing, head of PR for ByteDance's Doubao, called a rumor that Hengdian film studios was "feeding 150,000 extras' faces and voices to ByteDance's model" seriously inaccurate.

Liu said on Weibo that the project in question is the Hengdian Film and TV AI Service Platform. Its aim is to turn real filming locations and other assets into high-quality datasets that help productions scout and prepare, and to make AI easier for small film companies to use. Liu said any use of a person's likeness or voice in the project requires that person's consent. Public records show the project is at the contract-award stage.

Limitations: Chinese-language source, single report. Only ByteDance's side has been published, and the contract and data-consent terms are not public.

Source: Leiphone (Chinese-language source)