Benchmarks & EvalsAnthropic
Evidence check
  • Preprint
  • No code released
  • Not independently replicated

New Tumor Context Benchmark Tests Whether AI Can Predict How Cancer Cells React to Their Neighbors

CellType and Inventia Life Science launched Tumor Context, a benchmark that grades AI predictions of tumor gene changes against new lab data from 3D bioprinted tumors. Claude Fable 5.2 leads with 0.76, against an experimental ceiling of 0.97.

Christopher Ort
By Christopher Ort · 3 min read
Share
A scientist wearing blue gloves pipettes pink culture medium into a multiwell plate next to a microscope in a lab.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. Background
  5. Why it matters
  6. What's still open
  7. About the analysis
  8. i10x Perspective

CellType and Inventia Life Science have launched Tumor Context, a new AI benchmark that tests whether models can predict how cancer cells change their gene activity when their cellular neighbors change. The benchmark is part of CellType’s BioSimBench and is graded against fresh lab measurements from engineered 3D tumors that no model has seen. The best model so far, Anthropic’s Claude Fable 5.2, scores 0.76, against an upper limit of 0.97 set by repeating the lab experiment.

i10X Quick Take

Quick Take

  1. The benchmark: Tumor Context asks models to predict how 50 genes change when fibroblasts or T cells are added to tumor cells.
  2. The data: Inventia printed the tumors on its RASTRUM 3D bioprinting platform and measured them with bulk RNA sequencing.
  3. Leaderboard: Claude Fable 5.2 0.76, Gemini 4 Carbon 0.73, Claude Opus 5.5 0.72, GPT-6.1 Sol and GPT-6 Astra 0.72. A simple rule using no lab data scores 0.62.
  4. Tools help: On a 40-question subset, all four tested models improved when they could use tools.
  5. Scope today: Three starting cultures from two pancreatic cancer cell lines, PANC-1 and AsPC-1.

In plain English

A tumor is not just cancer cells. It is a neighborhood of different cells that influence each other. CellType’s example: the immune signal CXCL9, which helps draw cancer-killing T cells in, stays almost off with fibroblasts or T cells alone, but jumps to 486 units when both are present. The benchmark asks whether AI models can predict effects like this, a step toward simulating biology instead of testing everything in the lab.

What happened

CellType published the benchmark with four question types: predicting gene changes from a description alone, predicting with each gene’s starting level shown, transferring a measured pattern from one or two other cultures to a new one, and naming which cell type makes each gene’s RNA. Questions are split by gene family so no gene appears in both training and test, and CellType says it runs scripted shortcut strategies against every task, shipping a task only if none of them scores above zero.

Results by model family rise with model size, which CellType reads as evidence that “this biology can be learned”. With a description alone, the best model reaches 0.68; shown each gene’s starting level, it reaches 0.80.

Background

AI for drug discovery is pulling in record money, from Isomorphic Labs’ $5 billion Series C to a wave of biology foundation models. Benchmarks grounded in new wet-lab data are rare, because most public datasets may already be in training sets. Our Health & Medical Research page tracks the field.

Why it matters

Whether a patient responds to immunotherapy depends heavily on which cells surround the tumor. A model that can reliably predict how changing the neighborhood changes the tumor could help researchers choose which experiments to run first. The gap between 0.76 and 0.97 shows today’s frontier models are useful but far from a lab replacement.

What’s still open

  • How results will change as more tumor types, drugs and patient-derived cells are added.
  • Whether specialized biology models outperform general frontier models on the benchmark.
  • How much of the score transfers to real patient tumors.

About the analysis

Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.

This independent i10x analysis is based on CellType’s Tumor Context benchmark page and its published results. Scores are as reported by CellType.

Sources (plain text, no links):

  • CellType: “Tumor Context, BioSimBench” benchmark page and technical report (November 2026)

i10x Perspective

Benchmarks are where the AI biology hype meets measurement. Using newly printed, controlled tumors is a smart answer to data contamination. Expect labs to start citing Tumor Context scores the same way they cite coding benchmarks, and more results in our Benchmarks & Evals section.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Timeline

What happened before

  1. Release

    Claude Haiku 5.5

    1M context · $0.1 / $0.5 per 1M tokens

    support.claude.com (source, opens in a new tab)
  2. Release

    Claude Sonnet 5.5

    1M context · $2 / $10 per 1M tokens

    support.claude.com (source, opens in a new tab)
  3. Release

    Claude Opus 5.5

    1M context · $4 / $20 per 1M tokens

    First model of the Claude 5.5 family; performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.

    anthropic.com (source, opens in a new tab)
  4. Partnership

    Accenture partnership on independent frontier-AI evaluation (≥$1B each over five years)

    With Accenture · Research

    anthropic.com (source, opens in a new tab)
  5. Release

    Claude Fable 5.1 & Mythos 5.1

    1M context · $10 / $50 per 1M tokens

    Anthropic's most advanced models for coding and knowledge work. Price/context shown for Claude Fable 5.1 (models overview).

    support.claude.com (source, opens in a new tab)
Full Anthropic timeline ›

Editor's picks

  1. Reflection AI Gates Laser, Its Autonomous Penetration Testing Agent, Behind a Trusted Access Program
  2. BrightAI Raises $97 Million to Replace Scheduled Infrastructure Inspections With Physical AI
  3. Remanence Raises $30 Million Seed to Turn Companies’ Agent Data Into AI Models They Own
  4. Astra for Law: OpenAI Opens Its Legal AI to All ChatGPT Enterprise Customers
  5. Proton Lumo Desktop Brings a Privacy-First AI Assistant to Windows, Mac and Linux