The Collection

Pradeep’s CIO Insights letter of 1 September reports Anthropic’s Model Hardware Standard as a limited research preview. A common agent interface for microscopes, robotic arms, liquid handlers, lasers, and cameras. The standardised driver is read and write. The machine describes its own measures, settings, and enforced limits. Reachable as MCP, as CLI, or as code. This is partner-reported from a closed preview. The protocol has not been published for independent review. We print it as the letter printed it. Anthropic’s own Fable 5.1 blog, the same week, says they previewed the Standard last week. That corroborates existence. It does not add a number we were not given.

QuEra Computing put an agent on a laser used for neutral-atom quantum control, about $700,000. Overnight it read the instruments, changed the controls, observed, rewrote, tried again. Recovery from a lost lock: about 150 seconds down to about 6. Development success: 58 percent to 96 percent. A later test threw 700 randomised disturbances. The agent restored the lock 695 times. The hardest recoveries took 10 to 14 seconds. A specialist at the bench, the letter says, needed 5 to 10 minutes. Carnegie Mellon connected a robotic arm, a liquid handler, a plate reader, cameras, and three computers in about eight hours. Claude ran and revised a serial dilution. Vendor integration, the letter says, usually takes weeks. HHMI Janelia unified seven programs on a custom microscopy rig. AWS, Doosan Robotics, Tecan, Universal Robots, and QIAGEN are named as testing or exploring. Exploring is not production.

The laser locked. The protein bubbled.

Genentech is the failure the letter refuses to bury. Bubbles in a viscous protein. Claude treated the mess as software, retried with different parameters, and made more bubbles. Researchers explained the physics, a clean well, a gentler mix. Later it became a reusable skill. The production pattern in the letter is the standing order: explore inside a hard safety envelope, let independent sensors judge, convert stable behaviour into inspectable code. Write access without that envelope is just a faster way to foam the well.

At the World Humanoid Robot Games in Beijing a robot ran 100 metres in 8.86 seconds. Usain Bolt’s world record is 9.59. Last year the fastest robot on that distance was over 20 seconds. In February, at China’s Spring Festival Gala, Unitree robots danced in sync, spun, jumped, backflipped. Lee says they were not teleoperated. He also says it is still not useful work. Tesla’s Optimus serving drinks in October 2024 is the other exhibit: the letter says the hardware was fine and the software was not, and the run was apparently teleoperated. Karol Hausman of Physical Intelligence: robotics progress is hard to judge. Backflips impress. Picking up a Coke can is very hard. Théophile Gervet of Genesis AI puts roughly 90 percent or more of the value in manipulation. The floor does not fight back. Objects do.

Benjie Holson’s Humanoid Olympics, September 2025, set 15 simple tasks: doors, a peanut-butter sandwich, peel an orange. Physical Intelligence later demoed 10 of them. Turning a sock inside-out took 176 successful examples, about eight hours of collection. Four to ten times slower than a person. About 52 percent success. In January Holson added new challenges: make a bed, hammer a nail, catch an egg, plus long-horizon trash and a sunny-side egg in under five minutes. No announced solutions yet in the letter. Galbot, in a pharmaceutical warehouse in China, drops a box into a chute in about 40 seconds and trains a new SKU in five minutes. That is a job. Reyk Knuhtsen of SemiAnalysis on the Unitree G1: historically a few minutes of work, then a long cool-down. Improved, the letter says, to about 5 to 15 minutes of work with about 10 minutes of rest.

A backflip is a demo. A chute is a job.

Agility Robotics has deployed humanoids inside safety enclosures, with a plan to slow or stop near people. 1X is testing at home with a light, compliant design. The Wall Street Journal: families with young children are not in the testing. Rodney Brooks’s old warning still holds in Lee’s piece: new technology often takes under ten years from lab demo to a deployed robot, and it rarely succeeds faster. The self-driving parallel is sitting in the same letter. 2015 optimism. Waymo still in about 11 cities. No consumer fully self-driving car to buy. Write access on the track is not write access in the kitchen.

OpenAI showed Jalapeño, its custom silicon, at Hot Chips 2026. Richard Ho and colleagues. Initial RTL to tapeout on TSMC N3: nine months. A typical project, the letter says, takes at least 24. They cut the cycle roughly in half to two-thirds. A timeline slide marks Codex as enabling design and verification, the milestones in green. May 2026: “Codex runs on Jalapeño.” First silicon succeeded. That is the write-access story moved from the laser bench to the mask set. The model is in the loop that used to be a queue of specialists.

XLS, Accelerated HW Synthesis, is Google’s open-source high-level synthesis path, DSLX down to Verilog. OpenAI’s internal models search power, performance, and area against a human baseline. One slide: a 56 percent improvement on a BF16 multiplier. Late changes continued until RTL freeze. We do not claim AlphaChip was used on Jalapeño. The letter says that is unconfirmed. Anna Goldie and Azalia Mirhoseini left Google for Ricursive, with an ambition of months, then weeks, then days. We name that as adjacent. It is not a Jalapeño fact.

Hardware years. Model weeks. That is the constraint.

Ho’s SNUG 2025 dream, as the letter reprints it: AI and a human propose, measure, optimise. The remaining constraint is the one the laser already named. Models move in weeks. Hardware still moves in years, even when nine months looks like a miracle against 24. Write access on the RTL is still write access inside an envelope. Tapeout is the sensor that judges.

Claude Fable 5.1 is generally available. Mythos 5.1 is the same model with different safeguards, trusted access only, cyber and life sciences. Cache reads are $0.25 per million tokens, 75 percent cheaper. Input and output stay $10 and $50. Anthropic: typical workloads about 25 percent cheaper, highly agentic work up to about 45. On Terminal-Bench-Science 0.1, Fable 5.1 scores 52.6 percent against Fable 5 at 24.7, Anthropic’s setup, with noise of about 3.5 to 4.5 points either way. Terminal-Bench 4.0: Fable 5.1 at 55.8 percent, Mythos 5.1 at 60.9, Fable 5 at 42.0. Biology false positives down about 85 percent on elementary and medical benign prompts. Cyber: about 60 percent fewer interventions. Fable may discover vulnerabilities. It is not to develop exploits. Dual-use cyber still goes to Opus.

Encrypted file storage, when it ships, is customer-controlled cloud storage with the privacy of zero-data-retention plus misuse detection, rolling through the fall. Until then, eligible zero-data-retention stays on Fable 5.1. Watermarking for models after 2 August 2026 under the EU AI Act Code of Practice. C2PA on supported files. A detection API in private preview. Preserved thinking and anti-distillation for new organisations from today. Existing organisations, not yet. Anthropic claims scientific work: Mythos binders with partner validation, a Venus digital elevation model from Magellan, GPU kernel speedups up to 2.5 times on seven models. We print those as Anthropic claims, partner-validated where they say so. A rare-crash story from Millennium is an Anthropic-quoted partner anecdote. We print it as such. We do not stretch it.

Cheaper cache. The brief still overruns.

Katie Parrott and Dan Shipper, in Every’s free preview of the Fable 5.1 vibe check: after a summer on ChatGPT and Codex, Every is bringing work back to Claude. Faster. Clearer writing. Argues less when redirected. Still: overruns briefs, can invent quotations, sometimes keeps working when asked to stop. The full vibe check (Cora and Proof, Spiral tokens, the Opus 5 editing pipeline staying put) is subscribers-only. Unread paywall. We do not invent their scores. The desk and the laser are the same rule. If stop does not mean stop, do not give it the instrument.