Anthropic reports Claude drug-development results with six partners, DeepMind unveils the Genie 4 world model, Liquid AI opens the LFM3 small-model family
The thread today: models moving from demos into real workflows. Anthropic brings concrete numbers from drug labs and regulatory documents, and GitHub lets agents run as the organization with a model screening permission requests. Meanwhile, Ai2 shows that when a model is the judge, rewording the question can flip its verdict, so make evaluation stable before putting models into the workflow.
The partners are BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group; the work is still mostly preclinical.
Divergence Labs generated about 35,000 antibody designs with Claude in about a month; BigHat made and tested about 1,010 variants in about 10 weeks, with the best binding more than 10x more tightly.
On 1,061 pages of Enveda manufacturing and filing documents, Claude caught all 11 issues human reviewers found and flagged 7 they missed, cutting human review from over 25 hours to under 3 (89%).
Anthropic also rolled out the Claude Science research workbench, Claude credits for life scientists, and a Life Sciences Verification Program that lets trusted organizations use its most powerful models with more permissive safeguards.
Builder's takeFor those of us outside biotech, the Enveda numbers are the ones to copy: Claude did not replace review, it did a first pass and listed issues for humans to confirm, cutting time by nearly 90% while finding 7 more problems. Scoring reports in AI Interview and document Q&A in AI Cloud Drive can use the same split: the model does exhaustive first-pass screening with citations, a person makes the call.
Claude Enterprise, Claude for Government and the Claude API all get a 40% discount, covering federal agencies plus state, local, tribal and territorial governments, with no minimum spend and pay-as-you-go.
Direct purchases qualify today; purchases through cloud providers follow "in the coming months". The discounts are estimated to save the government $1 billion.
About 500 US federal, state and local organizations use Claude; weekly Claude for Government active users are up about 8x since February, and federal civilian usage grew about 20x in a year.
Claude for Government became generally available on September 30 with FedRAMP High authorization; Claude Enterprise is in process for FedRAMP Moderate through FedRAMP 20x.
Builder's takeRead with the October 26 S-1, government business is one of the cards Anthropic brings to public markets. For teams selling to government or regulated industries, note the compliance path: FedRAMP High and IL4/IL5 land on the cloud providers first, then flow downstream. Confirm the compliance status of your model on the target cloud early, not when the tender arrives.
Genie 4 is natively multimodal and generates audio alongside the visuals.
It builds on the Gemini family: Gemini Omni anchors real-world appearance and behavior, while other Gemini models provide reasoning for coherence and continuity.
It is a research preview for trusted testers, coming to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio, with a waitlist form.
The announcement gives no resolution, frame rate, duration or pricing.
Builder's takeWith no resolution, duration or pricing, there is no way to judge product fit yet. I would join the API early-access waitlist and watch whether it becomes a per-second API the way Imagen and Veo did. Teams building games, teaching demos or simulation environments for agents can prepare a fixed set of evaluation scenes now and compare on day one.
The lineup is LFM3-1.7B (text), LFM3-VL-2B (images and video), LFM3-Audio-2B (speech in and out), LFM3-Omni-2B (all modalities), plus d1-2B and d1-omni-2B for calibrated answers, all open-weight on Hugging Face.
LFM3-1.7B scores 38.42 on WebArena-Verified versus 3.02 for LFM2.5-2.6B, with 35% fewer parameters.
LFM3-VL-2B raises the average across 62 vision benchmarks from 57.6 (LFM2.5-VL-3B) to 71.4, at 1.9B versus 3.1B parameters.
Every model fine-tunes on a single GPU, with cookbook Colab notebooks for SFT, DPO, GRPO and continued pretraining; the post does not state a license.
Builder's takeA 2B-class model going from 3 to 38 on web-agent tasks says small on-device models are starting to do real work. For high-volume, low-difficulty jobs in AI Cloud Drive such as file classification and screenshot reading, I would fine-tune LFM3-VL-2B on our own data. Until the license is clear, keep it to internal evaluation, not production.
OpenAI and Oracle are expanding their partnership to bring Oracle business applications and AI agents into ChatGPT.
The partnership will also power new Oracle offerings.
Builder's takeI could not open the full post this time, so only these two points from the official summary are confirmed; which apps and when will need the details. The direction is clear: employees working ERP and HR systems from inside ChatGPT. B2B SaaS teams should decide whether their product gets called as an app inside ChatGPT or keeps asking users to open its own interface.
In public preview: grant a GitHub App the read-only "Copilot Requests" organization permission, install it on the organization, and pass a short-lived installation access token to the Copilot SDK.
Usage is attributed and billed to the organization that owns the installation, and its model policies, budgets and spending controls apply.
Organization owners can review which apps have access, find app-attributed usage in AI usage CSV reports via the app_client_id field, and revoke access by suspending or uninstalling the app.
Builder's takeBackground coding agents have mostly run on some employee's PAT, breaking when that person leaves and billing usage to an individual. With enterprise GitHub App installations GA on October 23, usage metrics split by client on October 25, and now this, the chain is complete: agents run as the organization, on short-lived tokens, with separate billing. Teams with homegrown CI agents can schedule replacing personal tokens.
In assisted mode, an LLM evaluates Copilot's requests to run a command, access a file system path or open a URL; requests classified as clearly safe proceed automatically and the rest still need approval.
It is generally available in Copilot CLI, the Copilot app and Visual Studio Code.
Enterprise administrators can block Allow All while keeping Assisted, disable Assisted separately, and set the default mode for new sessions; if the default is unavailable, sessions start with manual approval.
Builder's takeApproval fatigue is a risk in itself, and a model doing the first pass beats people clicking "yes" by reflex. But it is a probabilistic call. My setup: make Assisted the team default, block Allow All, and pair it with logs like the Claude Code safety-layer events from October 21, then spot-check what was auto-approved so the classifier does not become a black box no one reads.
2.1.311 (October 26) fixes standard-context sessions stopping with "Usage credits required for 1M context" when a request ran over 200K tokens; Claude Code now compacts the conversation and continues.
2.1.311 fixes SubagentStop hooks running about twice a minute for each running subagent instead of once when it finishes.
2.1.312 (October 27) caps prompt and agent hooks at 24 running at a time per session, subagents included, plus at most 12 that have waited past their event timeout; a timeout under 1 second counts as 1 second.
2.1.312 adds oidc.user_id_claim to the Claude apps gateway, to identify people by an id_token claim other than sub, such as Microsoft Entra ID's oid.
Builder's takeTeams using hooks for audit or blocking should note two things: SubagentStop used to fire repeatedly, so any counts based on it need correcting; and under the new concurrency cap slow hooks will queue, so give each hook a sensible timeout and check queuing under load. Companies on Entra ID single sign-on can switch user identity to oid.
SOCI indexing fetches only the files needed to start the container and streams the rest in the background; job definitions, entry points and data channels are unchanged, with a fallback to a full download if there is no index or lazy loading fails.
Measured image download times: 1.8 GB from 48 s to 3–11 s, 13.2 GB from 3 min 42 s to about 6 s, 35.4 GB from 7 min 17 s to 3–11 s; GPU idle time during image pulls dropped by up to 90%.
The first epoch takes 3–16 s longer, after which time per epoch matches the standard image; each tested job finished 1–6 minutes sooner overall.
It is available in all AWS Regions for Amazon ECR private repositories only; images under about 4 GB see small gains, and the runtime is containerd, so code relying on the Docker CLI or socket may need changes.
Builder's takeSaving a few minutes per job looks minor, but for frequent short jobs such as evaluations, small fine-tunes and data processing, the GPU-hour bill drops noticeably. Before switching, check whether your training scripts call docker; if they do, that is the only thing to change. Images under 4 GB are not worth the effort.
Data comes from ICLR 2026 reviews: 154 high-novelty and 145 lower-novelty human ideas, plus the 154 high-novelty ideas paired with 154 from a simple LLM generator.
On human-versus-generated pairs, rewording the prompt alone moved GPT-5.4 from 93.5% to 40.9% accuracy, below chance; on the human-only set, Claude Opus 4.6 shifted by 21 points.
A simple baseline (Claude Opus 4.6, low reasoning, majority vote over three calls) scored 0.854 macro-F1 at $2.59, versus 0.642 at $30.66 for Idea Novelty Checker and 0.772 at $82.05 for the Sakana AI Scientist evaluator.
The benchmark, data (allenai/novelty-judge-bench) and code (allenai/novelty-eval) are public; the post does not state a license.
Builder's takeThis is a warning for any product that has a model assign scores: interview scoring or automated content-quality review can swing when the prompt is merely reworded. I treat scoring prompts as code and rerun a fixed human-labeled set on every change. Cheap majority voting over a few calls is often steadier and cheaper than an elaborate agentic judge.
Researched and drafted with AI assistance; editorial standards and views set by Darius.
Every Monday, the week in AI in one email
Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.
Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →