Google DeepMind shipped two models in two days. Gemini Robotics 2.5, published November 10, is its newest vision-language-action (VLA) model, the layer of a robot stack that turns camera frames and instructions into hand and arm motion. AlphaGenome 2, published November 11, predicts how a stretch of DNA and its variants behave in one specified cell type at a time. Both come with large numbers that Google measured itself; only one comes with something an outside engineer can call this week.
The robot model: steep curves, partners only
DeepMind presents 2.5 as the most capable VLA model it has built. Across over 200 distinct tasks, it reports success rates 30% to 100% higher than previous baselines, depending on task type, with tasks finished 20% to 30% faster. The per-task figures, all against Gemini Robotics 2, show where the jump lands: packing pastries with tongs went from 5% to 100%, closing a toolbox lid from 0% to 93%, unhanging a mug from 29% to 75%. The bottom of the same list is the more instructive half: sweeping trash into a dustpan reached 52%, folding pants and packing a rice cooker 40%.
The test hardware was Apptronik's Apollo 2 humanoid with a five-finger, 22-degree-of-freedom SharpaWave hand, plus a Franka Duo rig with two-finger grippers. The headline field number comes from a Mercedes-Benz kitting pilot on production lines in Kecskemét, Hungary, and Sindelfingen, Germany, where Apollo 2 robots reached a 96% success rate, with the Gemini Robotics ER reasoning model tracking the workflow while 2.5 moved the hands.
Access is narrow. The post says the model is rolling out to robotics partners and mentions no API endpoint, SDK, pricing or waitlist. It credits "Gemini 4 Argon", a model of unstated availability, with speeding development: generating simulated practice runs and a digital twin from a real camera scene, then reading failure video to propose fixes. On safety, Google reports 33% better safe task completion around obstacles, 43% around nearby people, and 99.5% resistance on its ASIMOV-Jailbreak benchmark for Argon against 82% for Gemini Robotics ER 2.
The genome model: one cell at a time
AlphaGenome 2 changes the question rather than just the score. The first version predicted many tissues in parallel; the new one takes a single condition (a tissue, cell type or cell state) plus a million letters of DNA and predicts effects for that condition alone. A cell encoder builds a condition from a researcher's own single-cell data, from as few as a few dozen cells, without retraining. Readouts cover gene expression, histone marks, chromatin accessibility and now DNA methylation, and the model can impute an unmeasured readout from a measured one. Training used bulk and single-cell human and mouse data: over 107 million cell profiles across 68 tissues.
On TenK10K, a single-cell map of 28 primary immune cell types, AlphaGenome 2 identified 30% of causal variants against 16% for a baseline Google describes as the original AlphaGenome fine-tuned on single-cell data plus unnamed specialised tools. For the direction of a variant's effect on a gene, it reports 43% against 23%.
Non-commercial research access is live through the AlphaGenome API with Colab tutorials; commercial research is in private preview, and AlphaGenome 2 Skills for Google Antigravity are listed as coming soon. Use-case submissions close November 30, with extra quota and credits possible for selected projects. The first AlphaGenome's weights sit on Kaggle and Hugging Face behind non-commercial model terms, code under Apache 2.0; the new post says nothing about weights.
Where the numbers stop
A 96% pilot rate is not a production rate. Tech Times noted that Google gave no trial count, success definition or human-intervention rate, and DeepMind itself frames the Mercedes work as a proof of concept. The eight charted tasks are the ones Google chose to itemise, and tasks at 40% fail more often than they succeed: a fine lab result, not yet a shift's work.
On genomics, "causal" is a benchmark label. The post does not explain how TenK10K's causal variants were established, and a model ranking a variant highly produces a hypothesis for the bench, not a mechanism. Its case studies, including a distant variant raising ETS2 in gut macrophages in inflammatory bowel disease, reproduce published findings in hindsight. AlphaGenome 2 predicts molecular effects; it does not diagnose or treat anyone, and Google says both versions "have not been validated or approved for clinical or diagnostic purposes."
Neither post benchmarks against rival VLA or genomics models, only against Google's own predecessors, so cross-vendor standings cannot be read off these releases.
What engineers can actually do with this
For robotics, nothing callable, but two patterns worth stealing: a reasoning model that tracks the workflow while a separate action model executes, the planner and executor split of many software agents; and an evaluation loop that turns one camera scene into thousands of simulated retries. For bioinformatics teams in academia, the API is usable today; treat its output as a ranked candidate list for wet-lab validation, and read the terms before building anything you plan to sell.
What we could not verify
- Trial counts, success definitions and intervention rates for the 200-task suite and the Mercedes pilot. Small samples would make 96% far less firm.
- Whether "30% to 100% higher" is relative or in percentage points; with zero baselines the difference is large.
- What Gemini 4 Argon is and whether it is available.
- How TenK10K defines causal variants and which tools formed the 16% baseline. A weak baseline would shrink the doubling.
- The AlphaGenome terms page did not render for us; non-commercial status comes from Google's post. AlphaGenome 2 weight availability is unstated.
- No figure here has been independently replicated.
The 30,000-foot read
Both releases make the same move: Google grades its work against its own earlier work and publishes the grades it likes. That is normal for a launch and insufficient for a decision. The robot model is a lab result wearing a hard hat; the genome model is a research instrument with a usage policy. The second is the more useful today, because researchers can test it rather than admire it.
Watch for three things: a partner-published reliability figure with a denominator, independent replications of the TenK10K result, and commercial terms for AlphaGenome 2. Until then, the curves are steep and the error bars invisible.
Sources: Google DeepMind, Gemini Robotics 2.5 (November 10 date, VLA description, 200-task results, per-task figures, hardware, Mercedes pilot and 96%, partner rollout, Gemini 4 Argon, safety and ASIMOV figures); Google DeepMind, AlphaGenome 2 (November 11 date, conditional design, cell encoder, readouts, training data, TenK10K figures and baseline, access tiers, November 30 form, case studies, research-use-only statement); Tech Times (missing trial count, success definition and intervention rate); google-deepmind/alphagenome_research on GitHub (AlphaGenome 1 weights and licences). All benchmark figures are Google-reported and unreplicated. Correction to early reports: Gemini Robotics 2.5 was published November 10, not November 11; the 96% Mercedes figure is Google-reported (we did not check Apptronik's own channels); the 16% TenK10K baseline is not AlphaGenome 1 alone but includes fine-tuning and other tools. The AlphaGenome 2 post rounds the profile count to more than 100 million in its introduction and gives over 107 million in its body; we use the latter. SiliconANGLE's combined report could not be opened, so every figure was checked against DeepMind's posts instead.