GPT-6 Rosalind, robot brains and AI that leaves the chat

The news, and what is confirmed

Most AI news is about a better chat window. This piece is about two places where the answer is not text on a screen: a lab bench, where a suggestion turns into weeks of experiments, and a factory floor, where a suggestion turns into a moving metal arm. Both raise the cost of being wrong, and both change what a good model looks like.

What was released?

GPT-6 Rosalind (30 October) is the newest model in OpenAI’s GPT-Rosalind series for life sciences, which began in April (OpenAI). Its system card calls it “a frontier reasoning model for biology, drug discovery, and translational medicine” and says something that changes how to read the launch: it “uses the same weights as GPT-6 Astra”, OpenAI’s flagship reasoning model. What differs is the deployment. It is “designed to be more useful for legitimate advanced biological research” with fewer refusals on dual-use biology questions, and it is only available through trusted access: OpenAI reviews each applicant’s use, security and governance first. According to OpenAI’s release notes, it is rolling out to approved Research customers in Work, Codex and the API, and not in the Chat tab.

Gemini Robotics on Atlas is not new this week. Boston Dynamics and Google DeepMind announced at CES on 5 January 2026 that they would bring Gemini Robotics foundation models to the new electric Atlas humanoid, aiming at industrial work starting in car factories (Boston Dynamics). Since then DeepMind has released Gemini Robotics 2, a vision-language-action model it says “controls entire humanoid bodies from feet to fingertips” (DeepMind), and on 2 October Boston Dynamics showed a new four-finger Atlas hand with behaviours “trained entirely in simulation” (Robotics & Automation News). We could not find a new joint announcement from late October or early November, so we treat it here as the state of play rather than as this week’s news.

Key takeaways

• GPT-6 Rosalind (30 October) is GPT-6 Astra’s weights deployed for life sciences: fewer biology refusals, trusted access only, Work, Codex and API.
• Gemini Robotics on Atlas was announced in January; there was no new joint announcement this week, and no Atlas success rates are published.
• Both move AI from producing text to producing actions in the world, which is what the rest of this lesson is about.

Biology as tokens

A genome is a string of four letters, and a protein is a string of twenty. That makes biology look like language, and it is, up to a point. How you cut the string into tokens decides what a model can even notice.

What is it?

A language model sees text as a sequence of tokens and learns which token is likely next or missing. Biology has its own sequences. Protein language models such as ESM-2 use one token per amino acid and learn by filling in masked ones. DNA models such as DNABERT cut DNA into overlapping k-mers: windows of k bases that slide one base at a time.

How does it work?

Each choice is a trade. One token per base keeps every detail but makes sequences very long, and attention cost grows with length. A k-mer packs k bases into one token, so the vocabulary grows to 4^k (4,096 for k = 6). Overlapping k-mers keep the token count close to the base count; chunked k-mers cut it by k but break when a sequence shifts by one base. Translating DNA into protein, three bases per amino acid, uses the genetic code itself as a tokeniser.

For a sequence of n bases and k-mers of length k.

The shift test is the point. With chunked k-mers or a translation read from the wrong starting base, one extra letter at the front changes every token after it, even though the biology barely changed. Overlapping k-mers and single bases are robust to that; they pay with longer sequences. The protein view of the insulin snippet, MALWMRLLPLLALLALWGPDPAAA, is the start of preproinsulin; reading it one base later gives a different string entirely.

Why does it matter for Rosalind?

GPT-6 Rosalind is a general text model, not a sequence model; it reads sequences as ordinary text tokens, which split DNA in ways that have nothing to do with codons. That is one reason science assistants lean on tools: OpenAI describes Rosalind as combining biological reasoning with “scientific tool use” (OpenAI), such as calling specialised models or analysis code rather than eyeballing a sequence itself. A sensible picture is a reasoning model as the planner, with sequence models and lab software as instruments.

Key takeaways

• DNA and proteins are sequences, so they can be tokenised: per base, per amino acid, or as k-mers with a 4^k vocabulary.
• Overlapping k-mers survive a one-base shift; chunked k-mers and wrong-frame translation do not.
• General models like Rosalind read sequences as plain text tokens, so they rely on specialised tools for sequence-level work.

From camera frames to motor commands

A chatbot can take five seconds to answer and nobody minds. A robot arm moving at half a metre per second travels 2.5 metres in five seconds. Turning a camera frame and a sentence into motor commands is a race against the clock.

What is it?

A vision-language-action (VLA) model takes images from the robot’s cameras, an instruction in words and, often, the robot’s own joint readings, and outputs actions as numbers. DeepMind’s On-Device 2 model card lists exactly that: text, images and “robot proprioception” in, “robot actions as numerical values” out. It is a multimodal model whose output vocabulary includes movement.

How does it work?

RT-2 showed the simplest version: write actions as tokens. Each continuous action dimension is “discretized into 256 bins uniformly”, so one action becomes 8 integers the language model can emit like words. The catch was speed: the 55B model ran at 1 to 3 Hz. Two ideas fix that. Action chunks (Zhao and colleagues) predict a short sequence of future actions at once, and a split architecture runs a large model in the cloud and a small decoder on the robot. Gemini Robotics reports a backbone latency “under 160ms”, about 250 ms from camera to action chunk, and “with multiple actions in the chunk, the effective control frequency is 50Hz”.

To keep the motors fed while the next chunk is being computed.

The demo shows the real trade-off. Short chunks keep actions fresh but stall when latency is high. Long chunks never stall, but the robot spends longer acting on an old picture of the world: with a 25-action chunk at 50 Hz and 250 ms of latency, the last action uses a frame 750 ms old, and a hand at 0.3 m/s moves 225 mm in that time. That is why fast on-robot models and low-level controllers that react to touch matter, and why a model that runs on the robot, like the on-device models, is attractive for a factory with patchy networks.

Key takeaways

• A VLA turns camera images, an instruction and joint readings into numeric actions; RT-2 did it by making actions tokens in 256 bins.
• Big models are slow (RT-2 55B: 1 to 3 Hz), so systems predict chunks of actions and split work between cloud and robot.
• Chunk size trades stalls against acting on stale views: actions per chunk must be at least latency × control rate.

Sim-to-real and step-by-step reliability

A chatbot that is right 99% of the time feels excellent. A robot whose every grasp works 99% of the time fails most long jobs. Errors in the physical world multiply, and they cannot be regenerated.

What is it?

A factory task is a chain of steps: find the part, grasp it, lift, carry, align, insert, release. The task succeeds only if every step does. If each step succeeds with probability p, independently, a chain of n steps succeeds with probability p^n. A chatbot gets a second chance whenever you press regenerate; a dropped engine part is on the floor.

Assuming independent steps. A retry raises per-step success to 1 − (1 − p)^(retries + 1).

How do robots get that reliable? Sim-to-real

Collecting millions of real attempts is slow and breaks hardware, so robots practise in simulation, often with reinforcement learning. The trouble is the reality gap: simulated friction, lighting and contact are never quite right, so a policy that is 99% reliable in simulation can be much worse on the real robot. Domain randomisation (Tobin and colleagues) attacks this by varying textures, lighting, masses and camera angles so widely in training that the real world looks like one more variation. Boston Dynamics says its new Atlas hand already shows “initial sim-to-real results in dynamic manipulation tasks, with behaviors trained entirely in simulation” (Robotics & Automation News).

Key takeaways

• Multi-step tasks succeed with p^n: 99% per step is only about 37% over 100 steps.
• Retries help only when a failed step can be retried; real failures cluster, so independent maths is optimistic.
• Robots practise in simulation and close the reality gap with domain randomisation; transfer must be measured on real hardware.

Reading a science model’s benchmarks

OpenAI’s system card reports Rosalind’s jumps on several life-science benchmarks. To read them well, ask three questions: what the tasks are, how many there are, and how they were run.

What was measured?

The system card gives two kinds of results. Capability benchmarks ask whether the model is useful: LifeSciBench Gold “uses research tasks contributed by practicing life scientists”, GeneBench Pro tests planning and running quantitative analyses of biological data, and MedChemBench covers chemical structures, potency and toxicity. Preparedness evaluations ask whether it could help cause harm, such as troubleshooting lab protocols, and compare results against thresholds OpenAI sets for higher risk. A model that gets better at the first kind usually gets better at the second too.

How should you read them?

Small benchmarks are noisy. With about 100 problems, luck in which problems were chosen moves a score by several points. The system card also says GeneBench Pro “used a benchmark-specific configuration that differs from production”, so the number describes a set-up customers may not get. And the same capability that solves a protocol puzzle for a vaccine lab helps someone with bad intentions, which is why the capability and safety numbers have to be read together.

Normal approximation for a proportion p over n problems.

Key takeaways

• Rosalind 6 beats Rosalind 5.5 by 17 to 21 points on three capability benchmarks of about 100 to 130 problems each.
• At that size, a rough 95% margin is about ±8 to 10 points per score and ±11 to 14 points per gap; one benchmark ran in a non-production configuration.
• Capability and misuse risk rise together, so science models are judged on both, and finally by experiments that work.

What to watch

Leaving the chat window changes what counts as a good model. Here is what would tell you these systems are working, and what to ask of the next announcement.

What would be convincing?

For science models: independent evaluations, not only the maker’s; error bars or problem counts with every score; and published cases where a model’s plan led to an experiment that worked and was replicated. For robots: success rates over whole tasks on real hardware, measured over long runs, with the number of human interventions per hour. A demo video shows that something can happen once; a factory needs it to happen thousands of times a day.

What is the common thread?

Both are agents: a model that plans, calls tools or motors, sees the result and tries again. In both, the expensive part is not the model’s output but what happens because of it, so access controls, human review, fast control loops and step-level reliability matter as much as benchmark scores.

The same week also brought GPT-6.1 Terra and OpenAI’s DevDay, which are covered in a separate piece.

Key takeaways

• Ask for independent tests, problem counts and real-world outcomes, not just headline scores.
• For robots, the number that matters is whole-task success over long runs on real hardware.
• Science and robot models are agents whose mistakes cost time, money or safety, so deployment rules matter as much as the model.

Check your understanding

Six scenarios, from choosing a biology model to sizing a robot’s action chunks and judging a benchmark gap.

A biotech asks why it cannot just use GPT-6 Astra in the Chat tab for the same results as GPT-6 Rosalind, since the weights are identical. What is the best answer?

References

Model details come from OpenAI’s GPT-Rosalind-6 system card and release notes, Boston Dynamics’ partnership announcement and Google DeepMind’s model pages, checked on 5 November 2026. Background on sequence tokenisation, vision-language-action models, action chunking and sim-to-real comes from papers.

Sources

1. [1]

   Introducing GPT-Rosalind for life sciences research(opens in a new tab)

   OpenAI, 2026

   April 2026 announcement of the GPT-Rosalind series, OpenAI’s line of models for life sciences research under a trusted-access programme.

2. [2]

   GPT-Rosalind-6 System Card Addendum(opens in a new tab)

   OpenAI, 2026

   30 October 2026. “A frontier reasoning model for biology, drug discovery, and translational medicine”, offered as GPT-Rosalind-6 Research to qualified scientists, institutes, companies and government partners. “Uses the same weights as GPT-6 Astra”; fewer refusals on dual-use biological queries, only via trusted access. Four-part safeguards: trusted access, model-level boundaries, monitoring and enforcement, contractual requirements. Capability results: LifeSciBench Gold 61.3% vs 44.0% (GPT-Rosalind-5.5), GeneBench Pro v13 37.7% vs 16.7%, MedChemBench v7 49.6% vs 28.9%; preparedness results such as ProtocolQA Open-Ended 55.6% and TroubleshootingBench 59.0%.

3. [3]

   GPT-Rosalind for life sciences research (help centre)(opens in a new tab)

   OpenAI, 2026

   Release notes, 30 October 2026: GPT-6 Rosalind Research is rolling out to a limited set of approved Research customers in Work, Codex and the API (model id gpt-6-rosalind). It needs Research approval for the organisation and the user and is not available in the Chat tab. GPT-5.5 Rosalind Research remains available.

4. [4]

   Boston Dynamics & Google DeepMind Form New AI Partnership to Bring Foundational Intelligence to Humanoid Robots(opens in a new tab)

   Boston Dynamics, 2026

   5 January 2026, CES. Joint research to bring Gemini Robotics foundation models to the new Atlas humanoid for industrial tasks, starting with automotive manufacturing; research “expected to kick off in the coming months”. Alberto Rodriguez: a partner to “establish new kinds of visual-language-action models for these complex robots”.

5. [5]

   Gemini Robotics 2(opens in a new tab)

   Google DeepMind, 2026

   Model page for the Gemini Robotics 2 vision-language-action model (July 2026): text and image in, actions out; “controls entire humanoid bodies from feet to fingertips”; private preview.

6. [6]

   Boston Dynamics unveils new four-finger hand for Atlas humanoid robot(opens in a new tab)

   Robotics & Automation News, 2026

   2 October 2026: a 13-degree-of-freedom, four-finger Atlas hand with tactile sensors in fingertips and palm; Boston Dynamics reports “initial sim-to-real results in dynamic manipulation tasks, with behaviors trained entirely in simulation”.

7. [7]

   Evolutionary-scale prediction of atomic-level protein structure with a language model(opens in a new tab)

   Lin, Z., Akin, H., Rao, R., et al., 2023

   ESM-2, a protein language model trained with masked prediction where each amino acid is one token, and ESMFold, which predicts structure from its representations. Science 379, 2023.

8. [8]

   DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome(opens in a new tab)

   Ji, Y., Zhou, Z., Liu, H., Davuluri, R. V., 2021

   A BERT model for DNA that tokenises sequences into overlapping k-mers (k = 3 to 6). Bioinformatics, 2021.

9. [9]

   Gemini Robotics On-Device 2: model card(opens in a new tab)

   Google DeepMind, 2026

   Published 30 July 2026. Inputs: text instructions, images from the robot’s view and robot proprioception; outputs: “robot actions as numerical values”. Runs on the robot; available to trusted testers. Limitations include generalising to out-of-distribution tasks.

10. [10]

    RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control(opens in a new tab)

    Brohan, A., Brown, N., Carbajal, J., et al., 2023

    Expresses robot actions as text tokens: continuous action dimensions “discretized into 256 bins uniformly”, so an action is 8 integers. The 55B model runs at 1-3 Hz and a 5B version at about 5 Hz.

11. [11]

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware(opens in a new tab)

    Zhao, T. Z., Kumar, V., Levine, S., Finn, C., 2023

    Introduces action chunking with transformers (ACT): the policy predicts a sequence of future actions at once instead of one per step, which reduces compounding errors.

12. [12]

    Gemini Robotics: Bringing AI into the Physical World(opens in a new tab)

    Gemini Robotics Team, Google DeepMind, 2025

    A VLA backbone hosted in the cloud plus a local action decoder on the robot. Backbone query-to-response latency “optimized from seconds to under 160ms”; end-to-end latency from observations to action chunks about 250 ms; “with multiple actions in the chunk, the effective control frequency is 50Hz”.

13. [13]

    Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World(opens in a new tab)

    Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P., 2017

    Trains on simulated images with randomised textures, lighting and camera positions so the real world looks like just another variation; the detector transfers to a real robot without real training images.