ANG⁺ Briefing #2 — Anthropic's Public Benefit Fund & Hiring Data, DHS Bars DeepSeek, MiniMax M3.1 Flash, Mercury 3, and Reflection Beam

Welcome to the second ANG⁺ Briefing. This edition draws on six first-party sources published between October 16 and 20, 2026 — Anthropic's newsroom and economics research, a U.S. Department of Homeland Security notice filed with the Federal Register, and launch posts from MiniMax, Inception, and Reflection AI. Two items follow up on Briefing #1: what agent-scale coding is doing to Anthropic's own workforce, and a new U.S. federal order on DeepSeek.

1. Anthropic commits 5% of its shares to a new Public Benefit Fund — plus 2% for every 50% rise in company value

On October 18, Anthropic announced the Anthropic Public Benefit Fund, a new entity whose stake is designed to grow as the company grows. Subject to stockholder approval, 5% of Anthropic's shares are committed now, with an additional 2% added at the board's discretion for every 50% increase in the company's value above a set baseline.

• Structure and limits: The Fund is a wholly-owned LLC rather than a 501(c)(3), to allow flexibility between different forms of giving. Anthropic says none of the equity held by the LLC will be used for lobbying or political giving, in the US or abroad.
• What it could fund: Possible uses include large donations to traditional nonprofits or direct distributions to specific populations. Anthropic names labor market impacts as one example, noting it has already committed $200 million to studying and addressing the issue.
• Roughly 20% for the public good: At founding, Anthropic's co-founders pledged to donate 80% of their equity holdings, and approximately half of employees take part in an equity donation match program. Combined with the Fund, Anthropic says roughly 20% of its shares will be dedicated to the public good, plus any valuation-based additions. Governance details and the valuation milestones will be published "in the coming months."

Our take: Anthropic frames this as redirecting the kind of escalating equity awards tech companies usually give executives. The baseline valuation and the Fund's governance are still undisclosed — who decides how the money is spent will matter more than the headline percentage.

Source: Investing in the public benefit (Anthropic)

2. Follow-up: Anthropic's own hiring data — junior technical hires fell from 27% to 9% as agent hours overtook human hours

Briefing #1 covered how agent-driven coding pushed Anthropic off Git. On October 16, Anthropic economists published "How employment is changing at Anthropic," drawing on internal employment records, AI usage telemetry, output metrics, surveys, hiring scorecards, and 17 employee interviews.

• Fewer juniors in technical teams: The share of junior-level employees in technical departments slipped from 21% to 14% since early 2025, and the share of junior technical hires fell from 27% to 9%. Technical departments' share of total employment fell from 61% to 56% in the last year, even as technical hiring kept growing in absolute terms.
• Agents now out-work their handlers: In March 2026, total agent work hours passed human work hours for the first time; by August, employee-directed agents worked 6 times as many hours as the people running them. About 80% of Claude session time is spent idle, waiting for human input.
• Output and new kinds of work: Code commits per engineer reached 5x their pace of six months earlier. Coding-focused agent sessions among non-technical staff rose from 2 per month (August 2025) to 90 (May 2026), and written standing instructions for agents grew from a few hundred to 8,500 in a year. The authors caution that adoption is not random (they compare 192 adopting engineers with 74 non-adopters).

Our take: Here a frontier lab is measuring its own staff, not its customers. The pressure falls first on entry-level technical roles, and the new bottleneck is human attention — four-fifths of agent session time is spent waiting on a person.

Source: How employment is changing at Anthropic (Anthropic)

3. Follow-up: U.S. DHS orders DeepSeek products excluded and removed from federal civilian IT — Federal Register notice dated October 20

Briefing #1 covered DeepSeek-V4.1-Flash-1017. On October 19, the U.S. Department of Homeland Security filed a notice (FR Doc. 2026-21447, publication date October 20, 2026) announcing an order by the Secretary of Homeland Security under the Federal Acquisition Supply Chain Security Act of 2018 (41 U.S.C. 1321–1328), issued after review of a Federal Acquisition Security Council recommendation.

• Procurement exclusion: Information technology products and services developed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd, including affiliated entities as defined in 48 CFR 2.101, are excluded from federal civilian agency procurements.
• Removal, including third-party products: The order requires removal of all DeepSeek and DeepSeek-affiliate products, and third-party products and services developed using DeepSeek implementations, from civilian agency information technology.
• Where to find the order: The signed order is attached as an appendix and filed under docket DHS-2026-0628 on regulations.gov. Implementation questions go to CISA's FASCSA Order Implementation Team. The notice is signed by James H. Percival II, DHS General Counsel.

Our take: The third-party clause is the part vendors should read twice. Open MIT-licensed weights do not exempt a product — anyone selling to U.S. civilian agencies now needs to know whether a DeepSeek model sits anywhere in their stack.

Source: Order: Exclusion and Removal of DeepSeek Products from Federal Civilian Executive Branch IT Systems and Procurements (Federal Register, public inspection)

4. MiniMax rolls out M3.1 Flash at $0.15/$0.60 per 1M tokens — anonymous "Space Bunny" was an early version

On October 19, MiniMax officially rolled out MiniMax M3.1 Flash, revealing that the anonymous model "Space Bunny," which processed more than 200 trillion tokens, was an early version. MiniMax says online inference is nearly three times faster and agentic-task usage cost is 70% lower than its predecessor.

• Flat pricing: $0.15 input · $0.60 output · $0.015 cached input per 1M tokens.
• Company-reported benchmarks: 73.7% on DeepSWE v1.1, 76.2 on BankerToolBench, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam (first among Chinese models on both of the last two), and 82% on CyberGym.
• Speed and availability: Output speeds reach 200+ tokens per second in everyday use (P50 170+). Available now via the MiniMax API, M Plan, and MiniMax Code; MiniMax says model weights will be released later, and a Pro model will follow.

Our take: Next to Claude Haiku 5.5 ($0.10/$0.50 for prompts up to 100k tokens, Briefing #1), the small agent-model tier is now clustered around $0.10–$0.20 per million input tokens. Competition is shifting from price to speed and agent benchmarks.

Source: MiniMax M3.1 Flash: Redefining Agent Performance and Cost (MiniMax)

5. Inception's Mercury 3 diffusion LLM reaches 1,068 tokens/sec with a 1M context — launch price $0.10/$0.35

On October 19, Inception introduced Mercury 3, which it calls "the most capable diffusion-based LLM released to date." Unlike autoregressive models that write one token at a time, diffusion LLMs refine many tokens in parallel. Mercury 3 is more than 40x larger than Mercury 1, and Inception says standard-benchmark intelligence rose 3x from Mercury 1 (mini).

• Speed and context: 1,068 tokens per second on NVIDIA GPUs; Inception says it is 4–8x faster than GPT-6 Luna and Claude Haiku 5.5 with on-par-or-better quality. Context grows to 1M tokens, up from 260K in Mercury 2.5.
• Pricing: $0.20 input / $0.70 output per 1M tokens, discounted 50% at launch to $0.10 / $0.35; cached tokens are 90% off the input price.
• Early results and access: In a partner's web-agent harness, Mercury 3 completed 50% of 50 live Mind2Web tasks, "significantly more than GPT-6.1-Sol"; a search partner measured a median generation time of 0.9 seconds. Available on the Inception API Platform, OpenRouter, Baseten, and Vercel, OpenAI API-compatible, with 100 million free tokens to try.

Our take: The post does not publish full benchmark tables, so the quality claims are the company's own. But if diffusion keeps scaling, latency-bound jobs such as voice agents, tutoring, and browser automation get a genuinely different option.

Source: Diffusion at Scale: Introducing Mercury 3 (Inception)

6. Reflection AI releases Beam 1.0 weights under Apache 2.0 — 501B MoE, 23B active, 1M context

On October 16, Reflection AI released the weights of Beam, its first open-weight model (unveiled October 5), on Hugging Face under the Apache 2.0 license, together with reference agent integrations, deployment containers, and the Mirror CLI agent harness.

• Architecture: A sparse mixture-of-experts model with 501.3B total / 23.4B active parameters (256 experts) and a 1M-token context, developed predominantly for English. Reflection says it pretrained on about 23.8 trillion tokens and ran reinforcement learning on 10.5K NVIDIA GB300 GPUs for four weeks.
• Model-card scores (BF16): 81.6% SWE-bench Verified, 80.1% Terminal Bench v2.1, 90.6% GPQA Diamond, 97.7% AIME 2026. Reflection claims scores comparable to GLM-5.2 on advanced reasoning with 3–4x less inference compute.
• Deployment options: BF16, FP8, and NVFP4 checkpoints plus a speculative-decoding drafter; day-zero support in Transformers, vLLM, SGLang, Unsloth, and Axolotl; a Reflection API public beta that is "free to start"; and on-premises deployment, including air-gapped, through Dell and HPE.

Our take: A U.S.-built, Apache 2.0, 500B-class open model lands in the same week as the DHS DeepSeek order — giving public-sector and regulated buyers a self-hostable option that is easier to clear on supply-chain grounds.

Source: Beam is now available (Reflection AI) · Beam-1.0-501B-A23B model card (Hugging Face) · Introducing Beam (Reflection AI)