AI news straight from the labs. Twice a week, in five minutes.

Vol. I · No. 2

Tuesday, October 20, 2026

All issues

Anthropic's Claude designs 17 lab-validated protein binders against scorpion neurotoxins

Anthropic ran an internal research model inside Claude Science for 64 hours, with one initial prompt and four follow-ups developed with Citrini Research, to design minibinders against five scorpion neurotoxins. Adaptyv Bio and Twist Bioscience lab-tested 211 designs, and 17 met Anthropic's binding criteria. Anthropic is releasing all 211 designs and their data, and says it has not shown the binders neutralize venom.

By the numbers 17 of 211 experimentally tested designs met the binder criteria, each against a single toxin. The model generated 318,484 designs using 12 tools and screened 553,052 design-toxin pairs, using roughly 10,000 GPU hours. The two strongest binders, Cn2-B1 and Cn2-B2, showed single-digit nanomolar affinity at both labs. No multitoxin designs qualified as binders for more than one toxin; among 60 single-toxin designs, 11 were binders (18%).

Why it matters — It is a concrete, externally tested data point on how far autonomous AI agents can take an early-stage R&D workflow, while showing that lab validation and downstream drug development remain the bottleneck.

Read the Anthropic post →

United States

OpenAI to acquire Orchestra Bio to bring AI-powered operations to life sciences R&D

OpenAI announced it is acquiring Orchestra Bio. The company says the deal is meant to bring AI-powered operations to life sciences research and development. OpenAI frames the goal as helping biopharma teams speed up drug development and advance new treatments.

• OpenAI is acquiring Orchestra Bio to apply AI-powered operations to life sciences R&D.
• The stated aim is to help biopharma teams accelerate drug development and advance new treatments.

Why it matters — Biopharma and life sciences teams should expect OpenAI to push further into R&D operations tooling for their sector, which may affect vendor choices for AI in drug development workflows.

Read the OpenAI post →

Canada

Cohere's CALIBER trains reasoning models to report confidence before and after solving

Cohere researchers describe CALIBER, a training method in which a reasoning model states its expected chance of success before reasoning and then rates its specific answer afterward. Initial confidence is trained against how often the model solves a question across repeated attempts, while final confidence is trained against whether that particular answer is correct, alongside a reward for answering correctly. On a math benchmark the 7B model roughly halved calibration error versus the strongest single-confidence baseline, with gains also reported on three question-answering benchmarks.

• Cohere trained 7B and 30B models on 30,000 math problems and tested on 1,000 held-out problems plus three QA benchmarks.
• On BigMathDigits, the 7B model's expected calibration error dropped from 0.059 to 0.028, a 52.5% reduction versus the strongest single-confidence baseline.
• Without further training, the 7B model also reduced calibration error on non-math benchmarks.
• At 30B, CALIBER's calibration error was close to the CoCA-lite baseline, but it was much better at separating correct from incorrect answers; other baselines had slightly higher AUROC.

Why it matters — If this research approach reaches production models, confidence scores could become a more reliable way to decide which AI answers your team should double-check, though Cohere notes it has not yet tested uses like triggering verification or abstention.

Read the Cohere post →

Europe

Mistral releases Voxtral 3 speech models and a voice agent SDK

Mistral AI introduced Voxtral 3, a family of speech-to-text and text-to-speech models paired with a voice agent SDK built on LiveKit that streams listening, reasoning, tool calling, and speaking in one pipeline. The transcription models support more than 50 languages with diarization, context biasing, and end-of-turn detection, while the TTS model adapts emotion, pace, and emphasis across 35+ languages and 50+ dialects. The models are available via Mistral's API or for self-deployment.

• The Mini transcription model has 4B parameters and costs $0.0035 per minute.
• Voxtral Speech 3 text-to-speech is priced at $13 per 1M characters, which Mistral says is cheaper than several named competitors.
• The voice agent SDK runs its stages in parallel, so each conversational turn completes in under one second.
• Voxtral 3 is available now through the Mistral API on European, US, and global endpoints.

Why it matters — Teams building phone or web voice agents can access these models through a US endpoint or self-host them, and can swap the LLM, voice, or language without rewriting the agent.

Read the Mistral AI post →

Asia-Pacific

Sakana AI introduces QD-Rec, a method for generating diverse, reasoned recommendations

Sakana AI developed QD-Rec, a recommendation method that pairs each candidate (a company, customer, or paper) with a written hypothesis explaining why to choose it. An LLM repeatedly writes new hypotheses by mutating or combining strong past ones, and an archive keeps the ideas that are new or that beat their most similar existing entry. Across M&A, customer targeting, and paper recommendation tests, QD-Rec built larger archives with higher average scores than a baseline that writes each hypothesis from scratch in many settings.

• In many settings across three tasks, QD-Rec produced recommendation sets with higher average scores and larger size than a from-scratch baseline.
• The M&A experiment drew on data for about 6,300 US companies as of August 2024.
• In one example, a low-rated airline idea for JPMorgan was adapted to Liberty TripAdvisor, raising the LLM judge's score from 1 to 5 out of 10.
• The code and implementations for multiple domains are publicly available on GitHub.

Why it matters — Teams that shortlist acquisition targets, sales prospects, or research could use the open-source code to get a range of differently reasoned options instead of a single ranked list, though Sakana notes outputs need fuller, current data and human or expert review before real decisions.

Read the Sakana AI post →

Third Eye is published by Wave3 Labs. Every story links to its official source. Summaries are AI-assisted and reviewed by an editor before sending.