What Changed in AI-for-Science: September 2026 Digest
Evo 2's 270B-parameter genomic model is the new reference point for sequence-level biology, ECMWF's ensemble weather AI extends forecast skill further, and three major tool pricing overhauls are worth knowing about.
A roundup of what’s actually new and usable (or worth watching) across the fields this site covers — sourced from lab announcements, agency releases, and peer-reviewed papers, not general AI hype newsletters.
Headline: Evo 2 sets a new baseline for genomic sequence modeling
Evo 2, published in Science in March 2025 by the Arc Institute and EvolutionaryScale, is a 270-billion-parameter foundation model trained on 9.3 trillion nucleotides of DNA across more than 100,000 species. It is the largest genomic sequence model released to date and the first capable of meaningful zero-shot prediction across the full diversity of life — bacteria, archaea, and eukaryotes.
What makes it practically relevant for researchers:
- Variant effect prediction — Evo 2 can predict whether a genomic variant is likely functional or neutral without any task-specific fine-tuning, which is relevant for interpreting variants of uncertain significance (VUS) in clinical genomics
- Regulatory sequence design — the model shows strong performance on designing synthetic enhancers and promoters, opening a path to in silico regulatory element design before committing to synthesis
- Cross-species transfer — because the training set spans 100,000 species, the model generalizes to organisms with sparse sequence databases where smaller species-specific models fail
The model weights and code are available on HuggingFace (arc-institute/evo-2) under an open-use license for academic research. Inference requires significant GPU memory (the full 270B model needs multi-GPU setup); a 7B distilled version covers most research use cases and runs on a single A100.
Status: Published and publicly accessible. Source: Nguyen et al., Science, 2025
Also worth knowing
ECMWF’s ensemble AI forecast (AIFS-CRPS) is now operational alongside AIFS-single. ECMWF has extended its operational AI forecasting beyond the deterministic AIFS model to include an ensemble version (AIFS-CRPS) trained using a probabilistic scoring rule. The ensemble output gives researchers probability distributions over forecast outcomes rather than a single best-estimate trajectory — more useful for risk-sensitive applications in agriculture, hydrology, and climate impact work. If you’ve been using GraphCast or AIFS results as a baseline, the ensemble version is now the more complete reference.
ESM3 from EvolutionaryScale is a multimodal protein model worth knowing about. Unlike earlier single-modality protein models, ESM3 jointly reasons over sequence, structure, and function in a single model. Released in 2024, it enables prompting across modalities — for example, conditioning structure generation on a functional description, or querying function from partial structure. The 1.4B open model is available via Hugging Face; the larger models (98B) are available via API. For researchers already using ESMFold or AlphaFold, ESM3 represents a meaningful capability expansion for generative tasks.
OpenAlex has matured into a credible free alternative to Scopus and Web of Science. OpenAlex (openalex.org) is an open catalog of academic works, authors, institutions, and citations derived from Microsoft Academic Graph and maintained by OurResearch. As of 2025 it indexes ~250 million works with full citation graph data, accessible via a free public API with no account required. For researchers at institutions with limited subscriptions, or for building custom literature pipelines without licensing a commercial database, it’s now worth a real evaluation. The coverage gap vs. Scopus in medical and social science literature has narrowed substantially.
Worth watching (research-stage / not yet independently usable)
Generalist biomedical agent systems are moving into clinical research evaluation. Several academic medical centers are running internal evaluations of “research agent” systems — LLM-based pipelines that can retrieve, synthesize, and reason across clinical literature and patient data — under IRB oversight. None are available as independent tools yet, and methodological standards for evaluating them are still being established. Worth watching: whichever research group publishes the first rigorous prospective evaluation (not just retrospective benchmark) will likely define what “acceptable performance” looks like for this category.
Correction / status update
Three tools on this site have had pricing changes since our last digest — all verified September 2026:
Julius AI overhauled its plan structure. The platform has expanded beyond data analysis to include presentation, report, website, and image generation. New tiers: Plus ($16/mo), Pro ($37/mo), Max ($166/mo), Business ($375/mo). The free tier remains available. Updated tool page →
Perplexity Pro dropped to $17/month when billed annually (from ~$20/month). A new Max tier at $167/month adds 35,000 monthly credits, frontier model access, and higher usage limits. The free tier is unchanged and still sufficient for most background research tasks. Updated tool page →
Elicit added a Scale tier at $169/month (or ~$103/month annually) for team collaboration, figure extraction, and larger paper sets. Pro remains at $49/month. The free tier now includes unlimited search and summaries. Updated tool page →
Have a development we should cover next month? Use the Submit page to suggest a story or share how you’re using one of these tools in your own research.