Anthropic made the weekend's biggest corporate-governance move, committing 5% of its shares — with an escalating formula tied to company value — to a new Anthropic Public Benefit Fund. Combined with prior commitments, co-founder Jack Clark says roughly 20% of all Anthropic equity is now dedicated to the public good. Separately, Anthropic is pitching Congress a narrow draft federal AI safety bill that would let the government block a frontier model only on a court order, and only above a $1B-damages / 50-casualty threshold — drawing debate over state-law preemption.
On the model front, Chinese labs kept up the pressure. DeepSeek's V4.1-Flash-1017 checkpoint became the #1 open model on Code Arena: WebDev at 1670 points, passing Kimi K3 (Max) at roughly 10x lower cost, and shipped alongside new Search & Fetch APIs. StepFun open-weighted its 600B-parameter (27B active) Step 5 Preview with a 1M-token context built for long-horizon agentic work. Google Cloud AI Research open-sourced Harness², a recursive framework that lets agents rewrite their own prompts, tools, and memory — lifting GLM-5 by 11.5 points on WorkBuddy-Bench without touching model weights.
Reflection AI reports serving 1 trillion tokens of its Beam model in a single 24-hour window — a signal of how quickly inference demand is scaling beyond the big-three labs. Ethan Mollick flags that internal usage stats at OpenAI and Anthropic already show developers becoming a small fraction of total token consumption, shifting the economics toward long-running, high-value tasks.
Anthropic announced a new wholly owned entity holding 5% of its shares for public benefit, growing by 2% for every 50% rise in company value over a baseline, subject to shareholder approval. The fund targets AI's societal impacts, including labor-market disruption, and lobbying is off limits. Combined with prior commitments, roughly 20% of Anthropic equity is now pledged to the public good.
Per POLITICO, the draft proposal would limit when the federal government can block the release of an advanced model — only for a foreseeable, imminent risk of more than $1B in damages or 50+ deaths or serious injuries, and only via federal court order — while preempting some state AI safety laws and making Commerce the lead agency. Critics and supporters are sparring over the thresholds and the preemption clause.
The new V4.1-Flash checkpoint scores 77.0 on DeepSWE, opens Search & Fetch APIs, and is live on the website, app, and API at the same price as the September release. It now ranks #1 open model on Code Arena: WebDev with 1670 points — ahead of Kimi K3 (Max) at roughly 10x lower cost, and ahead of Claude Opus 5 (High) overall.
BF16 and FP8 weights are live on Hugging Face and ModelScope alongside a technical blog. The sparse-MoE model (600B total, 27B active) takes native text and vision input, supports a 1M-token context window, and is trained for long-horizon agentic work — with reports putting it around 1/6 the cost per task of Claude Opus 5.5 and #2 on OpenRouter's weekly chart.
Google Cloud AI Research open-sourced Harness², which improves agents by having them rewrite prompts, tools, memory, skills, and orchestration — without retraining or touching model weights. It lifted GLM-5 by 11.5 points on WorkBuddy-Bench with frozen weights. The repo is Apache 2.0, with a README noting it is not an officially supported Google product.
Selected posts from the edition, with links to the originals.
We're committing a portion of Anthropic's shares to the Anthropic Public Benefit Fund, a new wholly owned entity dedicated entirely to the public benefit. It will start off with 5% of our shares. For every 50% increase in the company's value over a set baseline, we can [add more].
View post on X ↗Exciting news: DeepSeek-V4.1-Flash-20261017 (Max) by @deepseek_ai is now the #1 open model in Code Arena: WebDev, surpassing Kimi K3 (Max) at 10x lower cost! Highlights: - #1 open model, ahead of Kimi K3 (Max) - #12 overall with 1670 pts, ahead of Claude Opus 5 (High) and within…
View post on X ↗Our team worked with OpenCode to have their users who are also OpenAI API customers back in business and any credits refunded which would have been drawn in the attack. Also, you can request Daybreak Blue access to use our cyber permissive setup for defense. Do not wait until…
View post on X ↗We just served 1 trillion tokens of Beam in 24 hours. To celebrate, we've reset everyone's limits. Go build something big
View post on X ↗Developers are going to be a small part of total token usage over time, as the internal stats on usage at OpenAI and Anthropic both show. And for long-running high value tasks in areas where verification is harder, the benefits for higher 'intelligence' models are quite large.
View post on X ↗who benchmarks the benchmarkers? a good eval is: - ephemeral. it snapshots a moment in time and capabilities. - grounded. it directly measures performance on tasks and data you care about. - specific. it operates on the desired results and not some proxy of them.
View post on X ↗Some strong embeddings models have been released recently. Ours is still better. We benchmarked Embed 5 Fast against Google's EmbeddingGemma 2 across 163 datasets using RCP-nDCG@10. Embed 5 Fast leads on over 95% of them.
View post on X ↗In harness engineering, open source means free distribution. There is almost no way to defend software in this domain unless the software contributes to model quality. In that case, software is eaten up by the model so quickly that you delete it all eventually.
View post on X ↗AI agents that hunt security bugs spend 60.4% of their tokens understanding code and reasoning about how to trigger the bug, so that is where cost cuts should aim. AI agents hunting security bugs can burn millions of tokens without producing an input that crashes the program.
View post on X ↗Going to make hermes agent into a DIY game dev studio
View post on X ↗