OrbitalHub

On October 26, 2026, OpenAI posted a preprint with a blunt title: The Exponential Time Hypothesis is False. The paper carried a date of October 17, a revision of October 25, a byline reading “OpenAI,” and a Lean formalization in the company’s public mathematics repository. Within hours, Anthropic published its own answer to the same problem, A 2^(o(n)) Time Algorithm for 3-SAT, with a Lean 4 development and build instructions. Both labs say a model found the result on October 16. The conjecture under attack was proposed in 1999, and much of complexity theory has leaned on it since.

The problem underneath is 3-SAT. An instance is a list of clauses, each one a disjunction of at most three literals over Boolean variables, and the question is whether some assignment of truth values satisfies every clause at once. Checking an assignment is trivial. Finding one by brute force means trying 2^n possibilities, and since 3-SAT is NP-complete, a polynomial-time algorithm for it would settle P versus NP the other way. The Exponential Time Hypothesis is a sharper bet. Put s₃ for the infimum of all exponents δ such that some algorithm decides 3-SAT on n variables in time O(2^(δn)). ETH says s₃ > 0: some fixed positive δ sits out of reach, so 3-SAT always costs a constant fraction of the full exponential. Russell Impagliazzo and Ramamohan Paturi proposed the hypothesis in 1999 and refined it with Francis Zane in 2001.

What makes it more than a guess about one problem is how much rests on it. ETH implies P ≠ NP, and its negation does not imply P = NP. Under ETH, k-Clique has no algorithm running in f(k) n^(o(k)) time, dynamic programming on bounded treewidth is optimal up to a log factor, and a long list of parameterized algorithms is optimal in the exponent. ETH also underwrites conditional circuit lower bounds: a theorem of Ryan Williams says that if ETH is false, the class E^NP has no linear-size Boolean circuits. Impagliazzo called the hypothesis a “principled educated guess” and put the odds at “a fair coin flip” in a 2012 lecture; Williams gave ETH about 70 percent odds of being true in 2019.

Both labs announced within hours of each other, and both shipped Lean proofs. OpenAI’s route was to enlist the people who know the terrain. Impagliazzo and Paturi at UC San Diego, who proposed the hypothesis, and Williams at MIT, who had already refuted strong versions of it, now lead an open examination of the claim, with a conference funded by OpenAI to work through the fallout. Williams told TechCrunch that in the OpenAI proof some parts are routine, “but then yes, there are two or three alien-looking parts,” and those are the pieces that convinced him the field needs to look at them together. His first question to OpenAI’s science lead Dan Roberts was whether any cryptographic system is now insecure; they had checked, he reports, and were reasonably confident nothing breaks. The release ran wider than 3-SAT: OpenAI also claimed the local smoothing conjecture in every dimension, which implies Kakeya, and existence for Yang-Mills theory in four dimensions, settling half of a Clay Millennium Prize problem. That manuscript stays with the group David Gross leads, a choice OpenAI’s advisory panel has publicly opposed.

Anthropic’s artifact is unusually well documented. Its paper, posted on the Hexagon preprint server under a CC BY licence, states the result in the abstract: a deterministic algorithm for 3-SAT running in 2^(o(n)) time, which also counts satisfying assignments, the harder #3-SAT problem. The provenance statement says the core algorithm was found on October 16 by an internal Claude research model using 14 million output tokens and no human input, after which Anthropic researchers worked with Claude on the write-up. The GitHub repository pins Lean 4.33.1 and the matching Mathlib, explains how to build the development, and notes that a second, independently written kernel called nanoda re-checks the proofs alongside Lean’s own. The README reports 1,050 modules, written by AI models and reviewed by other instances of the same models. Its strangest entry is a condensed description of the proof, enough, the README says, for a current Claude model to rederive the refutation on demand.

The mathematics behind Anthropic’s paper starts with the sparsification lemma, which lets you assume the formula is sparse, so its clause count is proportional to its variable count. The counting problem then becomes a sum of products over a graph whose nodes are clauses and whose edges carry the shared variables, and it can be split recursively as long as the bits crossing each cut stay few. The lever that makes it work is an identity two centuries old: Degen found it in 1818, Graves rediscovered it in 1843 through the octonions, and it says a match between two pairs of 3-bit vectors can be written as a product of sign functions, 2[x=v][y=u] = [x⊕y⊕u⊕v=0](σ(x,y)σ(v,u) + σ(x,u)σ(v,y)), where σ is a table of plus and minus signs. It spans 12 bits and can be checked at all 4,096 settings. Laying the two terms out as a grid turns each crossing into a choice between merging rows and merging columns, and iterating that, with pieces built at rate ρ = 1/√2, yields a runtime of 2^{O(min{km/√(log m), k²n log log n/√(log n)})}.

OpenAI’s paper reaches the same conclusion by a different road, giving a deterministic algorithm that counts satisfying assignments exactly, built from polynomial interpolation and sparse parity tests on low-degree forms, running in (n+m)^{O_k(1)} 2^{n/(log n)^{1/2+o(1)}} time with polynomial space. Both results are asymptotic in a way that matters: Anthropic’s bound, for a formula with the usual five clauses per variable, does not beat exhaustive search until the variable count passes 10^(10,000), and its bound in n alone does not bite until n passes 10^(100 million). Brute force is already unusable at 100 variables. Neither paper makes anything faster in practice.

What the Lean files do and do not establish is worth sitting with. Anthropic’s MainResults.lean holds no proofs at all, only definitions: def ETH : Prop := 0 < s 3 and def ETHIsFalse : Prop := ¬ ETH, followed by the main counting theorem. The proofs live in a separate file, and the README is blunt about the limit of machine checking: “Lean checks the proofs, but no machine checks that this file says what the paper says.” The formalization proves the stated proposition, not that the stated proposition is what the field means by ETH, and that gap is now a human responsibility. The paper’s appendix adds that the Lean development was written by AI models, reviewed by other models, and that no person is known to have checked it. OpenAI’s side has the same shape and a bigger pile: 721 manuscripts in 373 families, some formalized and some not, produced by posing roughly 4,000 problems to an unreleased model, each result averaging about three hours of thinking compute.

The consequences, if the work holds, are wide and uneven. The counting corollary gives s_k = 0 for every k ≥ 3, so s_∞ = 0 and both ETH and SETH are false. A literature scan carried out by Claude turned up more than 800 papers that prove a result of their own only under ETH, or under a hypothesis implying it. The Orthogonal Vectors hypothesis is untouched, though, and many fine-grained lower bounds rest on that rather than on ETH, so a large share of the lower-bound literature survives. P versus NP survives with it, with Williams describing the effect as a shift in priors rather than a solution.

What changed most is procedural. Two frontier labs have put the same kind of artifact in public view on the same day: a theorem the field spent decades formulating, a proof written by a model, and a machine-checked formalization that settles the steps but not the meaning. The checking is being organized as a community practice, with the proposers of the hypothesis running the examination. Credit is not, and the release policy is still being argued over in public, which is why the advisory panel OpenAI convened has been pressing for the physics manuscript to be released.

At 7:04 a.m. Eastern on 20 October 2026, a Falcon 9 lifted off from Space Launch Complex 40 at Cape Canaveral Space Force Station carrying more than thirty small satellites and one squat, uncrewed cylinder. The cylinder was a Starfall reentry capsule, and it belonged to the Air Force. Inside sat orange cargo pods holding the sort of gear you would want if the supply chain stopped working: military ration boxes, water, batteries, a hand-crank radio, a flashlight, a GPS tracker and a satellite beacon. The pods rode as a passenger on a commercial rideshare mission called Bandwagon-5, and when the capsule came back down in the Pacific a few weeks later, engineers would crack them open and see what the ride had done to the contents (AFRL, 20 October 2026; Air & Space Forces Magazine, 26 October 2026).

For five years the Air Force Research Laboratory has been trying to make a strange proposition look obvious: that the fastest way to move freight to a place with no runway is to put it in orbit first. The program, Rocket Cargo, began in 2021 with a fact sheet that set out the doctrine plainly. The lab would not build rockets. It would “leverage commercial rocket capabilities as a leased service,” and its stated goal was “up to 100 tons of cargo to be delivered anywhere on the planet within tactical timelines” (AFRL Rocket Cargo fact sheet AFRL-2021-1732, June 2021). SpaceX took the first contract in 2022, five years and $102 million (Air & Space Forces Magazine, 26 October 2026). AFRL now works with Blue Origin, Rocket Lab, Sierra Space, Anduril, Inversion Space and Outpost Space alongside SpaceX (AFRL, 20 October 2026).

What came back out of the Pacific was an experiment rather than a delivery. The pods splashed down inside the capsule and were recovered, each carrying sensors that logged temperature, shock and vibration through launch, a spell in orbit and a controlled reentry (AFRL on X, 26 October 2026). The lab also fitted a UNI-PAC III life raft inside the capsule to see how it fits, in preparation for carrying one on later flights. “Any day you launch a rocket is a good day, but today was more than just another big rocket launch,” said Brig. Gen. Douglas P. Wickert, who commands AFRL and serves as the Air Force’s technology executive officer (AFRL, 20 October 2026).

Starfall is new hardware, but the machinery around it is not. The Federal Aviation Administration’s May 2026 environmental assessment for the vehicle describes a squat cylinder about 0.75 metres tall and 3.1 metres across, with an empty mass near 2,100 kilograms, room for roughly 1,000 kilograms of payload, and a maximum reentry mass around 3,100 kilograms (FAA Final Environmental Assessment, May 2026, as summarized by Gunter’s Space Page). It has no propulsion of its own, survives the trip down under parachutes, and settles in the water roughly 700 nautical miles west of the United States coast, in international waters outside the economic zones of both the United States and Mexico. The FAA’s licence modification covers reentry of two such vehicles: a licence for two flights rather than for a fleet.

This is where the engineering gets interesting, because point-to-point cargo has to solve two problems at once: leaving orbit without cooking itself, and coming back with enough of its payload still working. The deorbit is unglamorous arithmetic. To drop from a circular orbit of radius r_a down to a perigee of r_p, the minimum retrograde burn is Δv = v_c * (1 - sqrt(1 - e)), where v_c = sqrt(mu / r_a) and the eccentricity of the new orbit is e = (r_a - r_p) / (r_a + r_p). From a 400-kilometre orbit, something like 90 to 120 metres per second buys a perigee low enough for the atmosphere to finish the job. That is a trivial amount of propellant, which is why the rocket is the expensive part of this system and the capsule is the cheap part.

Heat is the harder problem, and the numbers get worse fast. Convective heating on a blunt-body reentry goes roughly as q = k * sqrt(rho / R_n) * v^3, with rho the atmospheric density at the peak-heating point, R_n the nose radius and v the entry speed. The velocity term is cubed, so a vehicle that sheds a little speed early pays for it many times over. Low Earth orbit entry runs near 7.7 km/s. A lunar trajectory arrives at about 11 km/s, and the heating rate climbs with the cube. The work-around is a blunt forebody that stands off a detached bow shock, pushing the heat away from the vehicle, combined with an ablative shield that chars and erodes away, carrying the energy with it. SpaceX has already flown that trick at scale: Dragon’s forebody shield uses PICA-X, a phenolic-impregnated carbon ablator developed at NASA Ames from the material first flown on the Stardust sample-return capsule in 2006 (NASA Ames). Starfall’s own shield has not been published, so the physics of entry is well understood while the vehicle’s performance envelope is not.

What the lab was actually measuring decides whether any of this survives contact with commerce. Cargo is not a satellite. It is a wrench, a water purifier, a blood bag, a radio that has to work when nothing else does. The lab’s list of monitored items is a list of things with electronics and moving parts that do not forgive heat, shock or vibration: a reverse osmosis desalinator, batteries, a satellite-communications tracker, emergency medical equipment. A capsule that arrives intact but shakes a desalinator into scrap has not solved point-to-point delivery. The pods are a payload wrapped inside a payload.

The ride itself is what quietly changed. Bandwagon-5 launched more than thirty separate spacecraft for paying customers alongside the Air Force’s capsule (Gunter’s Space Page, 20 October 2026). The capsule did not buy a rocket; it bought a seat on someone else’s rocket. AFRL’s fact sheet was explicit that the program would not fund rocket development, only lease it, and a rideshare manifest is exactly that kind of lease. It divides the most expensive minute in the program, the launch, across three dozen customers, and leaves the reentry capsule as the only genuinely reusable piece of the chain. Drag the capsule’s cost per kilogram down far enough and the arithmetic stops looking like airlift.

The bottleneck is no longer hardware. It is airspace, and the paperwork. On 17 September 2026 the FAA released a draft programmatic environmental assessment that would streamline licensing for up to 70 capsule reentries a year in the Pacific, 22 in the Gulf of America and 10 in the Atlantic (FAA, 17 September 2026). Reentry has become a traffic problem, and traffic needs rules, airspace closures and a coast guard on the line. That paperwork moves slower than a rocket, and it now stands between a demonstration and a service.

Expect more of this rather than less. Starfall’s first demonstration flight in June 2026 was a vehicle test with no Air Force payload aboard; the October flight carried the cargo pods and the life raft, and the next round will carry more of both (Gunter’s Space Page; AFRL, 20 October 2026). At some point in that sequence the capsule stops being a capsule and becomes a truck. The version of this story that matters is the dull one: a pod comes down in the Pacific every week, gets unloaded by a machine onto a barge, and goes inland by road. The interesting part was always the question of who pays for the rocket.

On 21 November 2026, a spacecraft that has been decelerating since 2018 will finally stop. Partway through the day, more than 200 million kilometres from the people flying it, the stacked BepiColombo orbiters will turn their chemical thrusters against their own motion and fire. If the burn behaves, the joint ESA-JAXA mission becomes the first to hold two spacecraft in orbit around Mercury simultaneously. If it does not, eight years of trajectory design has to be rebuilt from scratch around a rock that most spacecraft are barely engineered to survive.

That frame is among the last things the mission photographed with the hardware it no longer needs. M-CAM 1 is bolted to the Mercury Transfer Module, the ESA-built propulsion stage that carried both science orbiters across the inner Solar System and was discarded on 3 September 2026 at a point 63 million kilometres from the Sun (ESA, September 2026). What remains is roughly 1,500 kg of spacecraft still moving fast enough that Mercury cannot simply grab it.

The module carried three of those 1,024 by 1,024 pixel black-and-white cameras, which exist less for science than for triage: they let engineers check hardware health from a distance when the main science instruments are pointed the wrong way or blocked. They also became the mission’s accidental souvenir shop, recording Earth, both Venus encounters and all six Mercury passes. The final swing-by on 8 January 2025 came within about 295 km of the surface and was the last gravitational tug before the long approach.

Mercury’s reputation as a hostile destination is mostly a matter of velocity. It orbits at 0.387 astronomical units, where the Sun’s pull is roughly seven times stronger than at Earth, and a spacecraft launched from Earth arrives there carrying far too much speed. Run the numbers with the vis-viva equation, v^2 = GM(2/r - 1/a), and a direct transfer from Earth’s orbit to Mercury’s comes out near 57 km/s at the arrival point, against Mercury’s own orbital speed of about 48 km/s. That surplus has to go somewhere. Entering a low orbit around a small planet means surrendering most of it to the planet’s gravity, and the only cheap way to surrender velocity is to let a planet take it away for you.

That is what the nine flybys were for. BepiColombo launched from Kourou on an Ariane 5 on 20 October 2018 at 4,100 kg, comprising a 1,230 kg orbital module carrying 85 kg of instruments and a 255 kg Japanese magnetospheric orbiter carrying 45 kg (ESA fact sheet). It swung past Earth once, Venus twice and Mercury six times, shedding heliocentric energy in small instalments, and used solar-electric thrust between encounters to trim the orbit inward. It is named for the Italian engineer Giuseppe Colombo, whose 1970s work on gravity-assist trajectories let Mariner 10 visit Mercury three times. ESA notes that orbit insertion at Mercury demands more energy than sending a probe past Pluto, which is why a nearly $2 billion mission needed a nine-stop itinerary.

The itinerary almost did not survive 2024. On 26 April, while the module was mid-manoeuvre, its power system began drawing unexpected currents between the solar array and the unit that extracts and distributes power, starving the ion engines. Engineers at ESOC in Darmstadt restored thrust to about 90 percent within eleven days but could not recover full power, and the flight dynamics team concluded in September that the reduced thrust could no longer reach the December 2025 insertion date (ESA, September 2024). The hardware was unfixable, so the trajectory moved. The fourth Mercury flyby on 4 September 2024 was flown roughly 35 km lower than planned, a small shave of altitude that conserved propellant and let a redesigned path carry the spacecraft to Mercury a year late while keeping the science programme intact (ISAS).

September 2026 was the handover. When the transfer module let go, the remaining pair had to inherit power generation, propulsion, pointing and thermal control in one shot, at a distance where a radio round trip takes hours. Confirmation came through Doppler shift as the two halves changed velocity, read out from Cebreros in Spain and Malargüe in Argentina (ESA, 3 September 2026). By mid-September the instruments that had spent years staring at the back of a dead engine deck were working. The BepiColombo Laser Altimeter fired more than 700,000 pulses over 21.5 hours, MERTIS returned about a gigabit during a 22-hour look at empty space through its planet and calibration ports, and SIMBIO-SYS returned its first image in space (ESA, 6 October 2026). A thruster rehearsal on 24 September produced 11.4 m/s over 8 minutes 42 seconds, the first trajectory correction of the arrival sequence.

The propulsion that carried it here deserves the credit. ESA loaded 581 kg of xenon into the module’s three tanks at Kourou in June 2018, and that fuel went through four QinetiQ T6 gridded ion thrusters, 22 cm across and rated around 4.5 kW each, producing 75 to 145 mN of thrust at a specific impulse up to 4,300 seconds (Hofer et al., AIAA 2010-7114). Do that through Tsiolkovsky’s rocket equation, Δv = v_e × ln(m_0 / m_f), with an effective exhaust velocity near 39 km/s, and burning a bit over a fifth of the stack’s mass buys on the order of 10 km/s of velocity change. The same job on a chemical engine with a 3.3 km/s exhaust would demand a mass ratio near 20 to one, meaning a spacecraft more propellant than spacecraft. That arithmetic is the whole reason solar-electric propulsion exists.

The other constraint was sunlight. At 63 million kilometres the solar flux is roughly 5.6 times what Earth receives, and heat limits how squarely the array can face the Sun, so ESA canted the wings and accepted a deployed span near 30 metres to collect the power four thrusters needed (ESA fact sheet). The September 2024 power fault sat uncomfortably close to that limit, which is why the recovery plan was flown conservatively rather than pushed.

What happens on 21 November is closer to a chemical burn, because capture cannot be done slowly. Arrival commands go up days ahead and execute autonomously, because at Mercury’s distance there is no opportunity to watch a burn and correct it. After capture, the pair separates on 9-10 December, MPO settles into a 480 by 1,500 km polar orbit with a 2.3-hour period, and Mio stretches out to 590 by 11,640 km to study the magnetosphere from outside. Sixteen instruments begin formal operations in April 2027 (ESA fact sheet).

The paired-orbit arrangement matters more than it sounds. MPO works close in, repeatedly crossing the planet’s surface to build topography, map composition and weigh the interior from gravity and radio tracking. Mio works far out, crossing the magnetopause and sampling the solar wind that Mercury sits inside rather than outside. No single spacecraft can do both jobs at once, and building two cost an extra decade and an extra module to carry them.

What MESSENGER left open in 2015 is what these two orbiters are built to close, and the arrival itself is an argument for slowness in a field that rewards speed. BepiColombo spent five years producing thrust no desk lamp could outrun, because that was the cheapest way to buy the velocity change it needed, and the bill for that patience came due in a single burn that nobody will be able to touch.

On October 20, 2026, NASA named Sierra Lobo Inc. as the supplier of the closed Brayton-cycle converter for Space Reactor-1 Freedom, or SR-1. Electricity is the converter’s product: heat from the fission core will drive its machinery to generate current, while the reactor itself is being designed at the Department of Energy’s Idaho National Laboratory. The launch is set for late 2028. Success would mark the first departure of a fission-powered craft from Earth orbit. Startup is deliberately delayed: only after SR-1 has cleared Earth orbit will the reactor be allowed to run, keeping an operating reactor away from the launch pad (NASA, Oct. 20, 2026).

The assignment matters because it links two machines that have to become one spacecraft. The receiving hardware is listed on NASA’s SR-1 page: a power-and-propulsion bus and a 12-kilowatt Hall thruster (NASA SR-1 mission page). The converter’s job is to close that gap with reactor heat, turning an announced Mars goal into work that can be scheduled, reviewed and tested against flight hardware.

The machine belongs to an old idea that has had a long wait for flight hardware. The United States launched SNAP-10A in 1965, the first and only U.S. fission reactor flown in space, then spent decades developing reactors without sending one to another planet (NASA Space Nuclear Propulsion). NASA separates the field into two broad designs. Nuclear thermal propulsion heats a propellant directly and sends it through a nozzle, producing high thrust. Nuclear electric propulsion first makes electricity, then uses it to accelerate a gas. SR-1 Freedom belongs to the second class, so its defining feature is a long, frugal burn rather than a brief shove.

That choice changes what the spacecraft has to carry. A chemical engine carries both fuel and oxidizer; a fission reactor supplies energy but still needs a working gas for the thruster. NASA notes that nuclear systems can provide solar-independent power for years, along with electricity for instruments and communications. The design bet is that a years-long low-thrust cruise can consume less propellant than a faster chemical departure. As sunlight weakens with distance from the Sun, that trade becomes more attractive. Those choices also directly affect spacecraft mass, shielding, thermal design and mission timeline.

That bargain appears in NASA’s 12-kilowatt Advanced Electric Propulsion System thruster. A 2025 technical paper reports about 600 millinewtons of thrust and a specific impulse near 2,800 seconds at full power (NASA Glenn Research Center, 2025). The exhaust velocity implied by that specific impulse is v_e = Isp * g_0, roughly 27,500 metres per second. Yet the propellant flow is tiny: mass flow = F / (Isp * g_0) gives about 0.079 kilograms per hour. Applied to a 12,000-kilogram spacecraft, one thruster produces an acceleration near 0.00005 m/s2 and adds only about 4.3 m/s over a full day. The craft is not trying to leap. It is trying to spend almost nothing while it goes.

The vehicle inventory on NASA’s SR-1 page lists a 12,000-kilogram mass, high-assay low-enriched uranium fuel, a 20-kilowatt-electric closed Brayton converter, a 48-kilowatt power-and-propulsion bus, and a 12-kilowatt Hall thruster (NASA SR-1 mission page). Those power numbers describe different pieces and should not be added into one reactor output. One labels the converter; the other belongs to the bus. The electrical design therefore has to reconcile a 20-kilowatt fission source with a bus capable of 48 kilowatts. That gap makes power management, radiator sizing and the operating sequence central engineering questions before the first Mars departure.

The converter is where nuclear physics turns into machinery. In a closed Brayton cycle, a working gas is compressed, heated through a heat exchanger coupled to the reactor, expanded through a turbine to drive a generator, then cooled and returned to the compressor. “Closed” means the working fluid circulates. The reactor fuel and working gas need not touch, which keeps the power-conversion loop separate from the radioactive core. For an ideal Brayton engine, thermal efficiency follows eta = 1 - 1 / r_p^((gamma - 1) / gamma), where r_p is compressor pressure ratio and gamma is the gas ratio of specific heats. Real hardware loses efficiency to expansion, pressure drops and heat leakage.

Heat also has to leave. If reactor thermal input is Q_in and the converter efficiency is eta, the waste heat is Q_reject = Q_in * (1 - eta). At a useful 20 kilowatts, even a hypothetical 30 percent converter would have to shed roughly 47 kilowatts. That is why the artist concept shows radiator panels: the cycle cannot close without a way to dump heat in space. At 20 kilowatts of output, the converter is a heat-rejection problem as much as an electricity problem. Radiator area, fluid temperature and reactor thermal limits have to be solved as one system.

Safety has shaped the schedule as much as propulsion. NASA has ruled out critical reactor operation during launch; startup waits until the spacecraft is beyond Earth orbit. The agencies also apply interagency reviews, regulatory compliance and failure-tolerance analysis, while a NASA chief safety officer provides continuing oversight (NASA, Oct. 20, 2026). That separation reduces one class of accident: if the launcher fails, there is no active reactor to scatter. The remaining work is substantial. Fuel must be produced and qualified; the converter must survive vibration and years in space; the reactor must start remotely; and ground teams must rehearse commands across a communications delay that grows with distance.

The science payload gives the engineering test a reason to keep flying. SR-1 will carry SkyFall, three Mars helicopters derived from NASA’s Ingenuity design. They are to be released during atmospheric descent, fly to the surface and use ground-penetrating radar, imagery and atmospheric sensors to scout terrain and look for buried ice. NASA plans an initial Mars flyby in 2029 and a landing after a second approach in autumn 2030 (NASA SkyFall mission page). The helicopters can scout landing sites that a rover would take much longer to reach, and their data can be transmitted directly through an orbiter rather than a relay rover.

SR-1’s late-2028 launch also has a schedule built into it: NASA has targeted the 2028 Mars launch window, with arrival in 2029 before deployment (NASA SR-1 mission page). If the reactor, converter, bus and thrusters work together, the mission will establish an operating fission system in deep space. It will also test a supply chain that NASA wants to reuse for Lunar Reactor-1, the surface power system intended for its Moon Base (NASA SR-1 mission page).

Success will be a chain of systems behaving together: the reactor starts, the converter holds a stable electrical output, the radiators reject waste heat, the thrusters accelerate the craft and the payload separates cleanly. Each link can be tested on Earth. Their behavior together, after years of unattended exposure and a communications delay measured in minutes, cannot. If that chain holds, the useful outcome is larger than three helicopters. It is a flight-proven method for carrying watts instead of sunlight into places where the Sun is weak, absent or inconvenient.

Spaceflight systems tend to fail at their interfaces. A converter can meet its bench target and still strain the reactor; thrusters can work while the bus changes modes; the payload can separate but its radio link can fail. A useful sortie would give engineers measured limits, failure modes and workarounds for those seams. That knowledge is what a lunar reactor, cargo tug or outer-planet probe will need. The helicopters are visible; the operational inheritance is the point.

NASA’s Artemis III operations concept graphic has the Moon above the Earth and three rockets lined up beneath it: a New Glenn carrying a Blue Origin lander test article, the Space Launch System carrying Orion, and a Starship carrying a second test article. All three fly in low Earth orbit, and none goes to the Moon. That is the plan Administrator Jared Isaacman, Associate Administrator Amit Kshatriya and Artemis Program Manager Jeremy Parsons presented at Kennedy Space Center on October 23, and it reverses the promise that defined the program for a decade. Artemis III, the mission built to put astronauts back on the surface, is now a two-week low Earth orbit flight targeted for mid-2027, with four astronauts docking first with a Blue Origin test lander and then with a SpaceX Starship. The first landing since Apollo 17 moves to Artemis IV in early 2028.

Artemis II cleared the way, launching on April 1, 2026 and flying four astronauts around the Moon on a free-return trajectory, 406,771 kilometres from Earth at closest, before splashing down in the Pacific ten days later (Wikipedia, Artemis II). The mission worked, but the arithmetic behind the schedule never changed. The European Service Module, the Airbus-built element that powers Orion and pushes it around, lacks the delta-v to return from low lunar orbit; in Apollo the lunar module was the crew’s lifeboat, and in Artemis the lander must do that job and more (Wikipedia, European Service Module).

That lifeboat arrangement had a staging point, the Lunar Gateway, and NASA cancelled it in March 2026. A month before Artemis II flew, Isaacman had already restructured the flight sequence so Artemis III would stay in Earth orbit and Artemis IV would attempt the landing (Wikipedia, Artemis III). The pressure sits on the parts NASA does not build: both commercial landers are behind schedule, as is Axiom’s spacesuit, and the FY2026 budget request proposed ending SLS and Orion after Artemis III, citing a launch cost of about $4 billion (Wikipedia, Artemis III).

The flight plan is specific enough to be built. Blue Origin goes first, with a lander test article lifted off New Glenn from Launch Complex 36A carrying a crew module, avionics, partial life support and a reaction control system but no main propulsion. It can loiter in orbit for weeks, which is why it launches first, and Blue Origin would spend about two weeks checking it out before Orion flies (Wikipedia, Artemis III). Orion follows on an SLS Block 1 from Pad 39-B with an inert structural spacer in place of the Interim Cryogenic Propulsion Stage, since the mission never leaves low Earth orbit and does not need that stage’s RL10 engine (Wikipedia, Artemis III and Interim Cryogenic Propulsion Stage).

Orion is the chaser in both encounters, holding station while the passive target keeps its orientation, with time for another attempt if the first docking fails. Once mated, the crew stays about two days, equalises pressure in the Blue Moon vestibule with Orion’s atmosphere, opens the hatch and goes inside to check cabin habitability and the interfaces between the lander and Axiom’s AxEMU suit (Wikipedia, Artemis III). After undocking, a standard Starship v3 with a passive docking fitting on its nose cone launches from Pad 39-A, and Orion docks with it briefly without the crew leaving the capsule, because that vehicle has no crew cabin and no life support. The point is control: verifying that Starship can hold the attitude of a combined stack with Orion attached. Orion comes home on an upgraded heat shield and splashes down off San Diego for Navy recovery.

What the rehearsal is for became clearer in the briefing charts. Two landing architectures are on the table and NASA has not chosen between them. In the Starship version, Orion docks with a Starship Human Landing System in low Earth orbit and Starship pushes both vehicles to the Moon, which needs a propellant depot and several tanker flights to move cryogenic methane and oxygen in orbit, a manoeuvre never demonstrated at scale (SpacePolicyOnline, October 23, 2026). In the Blue Moon version, three transfer stages on New Glenn push the lander to lunar orbit while Orion flies straight to the Moon. Each company must land uncrewed first, and NASA intends “to wait as long as possible” before choosing the provider that carries the first crew back (CBS News, October 23, 2026). NASA will also accept a lighter “sortie” suit with about two hours on the surface (SpacePolicyOnline, October 23, 2026).

The destination is unsettled. Kshatriya conceded that NASA’s knowledge of the south pole is thin, while Isaacman called it “probably the most advantageous area” for ice in permanently shadowed terrain, and a National Academies report in mid-October identified four non-polar sites with scientific value (SpacePolicyOnline, October 23, 2026). NASA’s Aerospace Safety Advisory Panel objected a day earlier to requiring only one uncrewed landing per provider and to putting cryogenic propellant on the Moon for the first time (SpacePolicyOnline, October 22, 2026). China is aiming for a crewed landing by 2030 (CBS News, October 23, 2026).

The propulsion problem underneath is arithmetic. Orion’s service module carries 8,600 kg of usable MON-3/MMH propellant and flies one AJ10 main engine at 26.6 kN, eight R-4D thrusters and 24 smaller reaction control units (Wikipedia, European Service Module). A direct rendezvous profile asks that stack to leave Earth, arrive at the Moon and come home: roughly 3,100 m/s for the translunar injection, a few hundred for lunar orbit insertion, about 900 m/s to come back, close to 4,800 m/s in total. The rocket equation says what that costs: Δv = v_e * ln(m_0 / m_f), with v_e = Isp * 9.80665. At a generous 300 seconds of specific impulse, exhaust velocity is about 2,940 m/s, so 4,800 m/s demands a mass ratio near 5, and Orion would have to burn four-fifths of its own mass.

The cheap way to buy delta-v is not a bigger spacecraft but a cryogenic stage, because hydrogen’s chemical energy per kilogram is roughly three times that of the hypergolic mix and empty tanks are cheap. The Interim Cryogenic Propulsion Stage that flew Artemis II shows the arithmetic: a single RL10 at 110.1 kN, specific impulse 462 seconds, gross mass 32,748 kg, empty mass 3,800 kg, so Δv = 462 * 9.80665 * ln(32,748 / 3,800) comes to about 9,700 m/s, roughly double the whole cislunar budget (Wikipedia, Interim Cryogenic Propulsion Stage). NASA is now working with United Launch Alliance on an extended Centaur V to fly with the SLS core stage, so the launch vehicle can inject Orion to low lunar orbit and leave the service module’s small tank for the lunar leg and the return (Ars Technica, October 24, 2026).

The orbit where the vehicles meet matters as much as the propulsion. NASA’s charts show Orion and the Blue Moon lander joining in an elliptical lunar orbit with an apolune near 8,500 km, which lets a lander with limited propellant reach the surface; Parsons said a low lunar orbit a few hundred kilometres up would cut the lander’s propellant requirement substantially (Ars Technica, October 24, 2026; Wikipedia, Artemis IV). Docking adds its own problem: Orion is the active vehicle, the other side passive, and with Starship the mass difference is roughly an order of magnitude, so a closing speed of a few centimetres per second is the difference between a soft mate and a bent structure. Isaacman’s summary of the fix was blunt: reaching low lunar orbit means “solving propulsion and spacecraft integration problems that were left unaddressed for far too long” (Ars Technica, October 24, 2026).

The most useful thing NASA did on October 23 was turn a landing into a list of things that can be checked, in an orbit where a failed rendezvous ends in a re-entry within hours instead of a rescue measured in days. What Artemis III will not settle is the hardest part: the test articles are stripped of the systems that make the real flight hard, with no main engine on the Blue Origin vehicle and no life support aboard Starship. Those are the 2028 problems, and they will be solved, if at all, by the two companies racing each other to be ready. The long-lead hardware keeps renewing in the meantime: on October 21 an RS-25 fired for 500 seconds on the Fred Haise Test Stand at Stennis, reaching 111 percent of the thrust needed to launch SLS (NASA, October 21, 2026). Apollo was a state program with one rocket and one schedule. Artemis is a procurement program that has to fit itself around two private launch cadences, and the first way NASA tests that arrangement is to launch three rockets and see whether they can find each other.

In a testbed at the Johns Hopkins Applied Physics Laboratory in Laurel, Maryland, a rotorcraft that has never left Earth crossed a landscape that does not exist. It accelerated over simulated dunes, scanned the ground beneath it, picked out a flat patch, turned, and settled down. On October 23, 2026, NASA described that run as the most thorough rehearsal yet of the moment Dragonfly lifts off for the first time, on Titan in 2034. The test ran in real time on flight-like avionics, with simulated data fed into the navigation cameras, inertial measurement units and lidar the spacecraft will carry (NASA, Oct. 23, 2026). The software did more than fly the vehicle. It searched for a landing site and chose one, and on Titan that choice belongs to the vehicle alone.

Dragonfly is a nuclear-powered octocopter about the size of a small car: roughly 953 kg, 3.85 m across and 1.75 m tall, with four pairs of coaxial 1.35 m blades, a 147 ampere-hour battery and a Multi-Mission Radioisotope Thermoelectric Generator supplied by the Department of Energy through NASA’s radioisotope power program (NASA Dragonfly fact sheet, Sept. 2026). Its five instruments run from a drill that feeds surface material to a mass spectrometer, to a camera suite, a gamma-ray and neutron spectrometer, and a geophysics and meteorology package. Launch is no earlier than 2028, the cruise takes about six years, and the surface mission runs a little over three years.

The world waiting for it is the reason flying makes sense there. Titan’s surface pressure runs about 60 percent above Earth’s, the air is roughly 95 percent nitrogen and 5 percent methane, and the ground sits near -179 degrees Celsius (NASA Titan facts). Sunlight takes about 80 minutes to arrive and lands roughly 100 times fainter than it does at Earth, so solar panels would be close to useless; the radioisotope generator is what keeps the rotors turning. The moon’s pull is weak, with a surface gravity of 1.35 m/s2, about 0.138 times Earth’s (NASA Glenn, 2022). Thick air and weak gravity are the two conditions that turn a spacecraft into an aircraft.

Everything about the recent test follows from that last point. The team fed the flight software simulated sensor data: imagery from the two navigation cameras installed at APL in August 2026, inertial readings, and lidar returns for terrain height and slope (NASA, Oct. 23, 2026). Lidar was the missing piece. The campaign had grown in steps, from spinning a single rotor to a short repositioning flight to a partial scouting flight, and adding lidar finally made a complete first flight reproducible in simulation. A custom camera emulator rendered each synthetic Titan scene on graphics processors with exposure times of roughly 20 to 70 ms, matching what the real cameras will see in flight. Because the run happened on flight-like hardware, the team could watch information move between processes at the millisecond pace it really moves.

Zach Putnam, the mobility subsystem lead engineer at APL, frames the difficulty plainly: “The cool thing about first flight is that it’s our most challenging flight.” It is the case that “uses every piece of flight sensor hardware on the lander” (NASA, Oct. 23, 2026). Leave out a sensor there and you leave out the ability to make a decision.

The part that cannot be rehearsed in software is the moment of release. Dragonfly arrives inside a backshell hanging under a parachute, and has to lower itself clear of that shell before separation, with its rotors already turning. Engineers at Lockheed Martin in Denver, where the cruise stage and entry, descent and landing hardware are being assembled, have been testing that “pose” mechanism: pushing and pulling on it with representative mass to check the loads it must take, lowering it to see how smoothly and quickly it deploys and whether it locks, and tilting the whole rig to reproduce the swing of a backshell beneath a parachute (NASA, Oct. 23, 2026). The pose structure links the backshell to an aluminum interface attached to the lander, and it has to work once, unrepeatably, about 1.6 billion kilometres from home.

That hardware has already been through a strange battery of ground tests. In May and June 2026, engineers suspended the fuselage a few inches above the floor on long bungee cords and measured how vibration at the rotor mounts travelled through the frame, then lowered the vehicle onto protective padding to see how it would sit on its skids. They also pressurized the assembled structure to hunt for gaps, because most planetary hardware is built for vacuum and Dragonfly will sit in air 60 percent denser than Earth’s (NASA Dragonfly blog, July 9, 2026). A slotted foam high-gain antenna, 87.4 cm across, folds onto a motorized arm before each flight so it cannot resonate against the vehicle.

The physics of the machine is worth doing roughly, because it explains how a tonne of hardware can hover on another world. Hover needs thrust equal to weight, T = m * g, so 953 kg on Titan asks for about 1.29 kN. Spread across eight rotor discs of 1.35 m diameter, the total swept area is roughly 11.4 m2, giving a disk loading near 112 N/m2. With air density around 5.7 kg/m3, worked out from Titan’s pressure and temperature and about 4.6 times the density of air at sea level, the induced velocity v_i = sqrt(σ / (2ρ)) comes out near 3.1 m/s. The same machine in Earth’s air would need about 6.8 m/s, so ideal hover power P = T * v_i falls from roughly 8.7 kW to about 4.0 kW. Dense air is a gift: the rotors push against more mass of gas for the same thrust. Real numbers run higher once profile drag and rotor efficiency are counted, and the electrical budget comes from a generator and a 147 ampere-hour battery, which leaves no room for waste.

None of that saves a vehicle that cannot decide for itself. Signals from Titan take the better part of an hour and a half to reach Earth, which is why NASA’s engineers talk about autonomy as a requirement rather than a feature; the rule Putnam repeats is that “everything we do on Titan has to happen autonomously” (NASA, Oct. 23, 2026). So the next round of tests is as much about failure as about flight. The team will run the first flight on the actual flight system instead of the testbed, fly other profiles, and deliberately inject faults to see whether the vehicle notices and recovers on its own.

Spread across 2026 and 2027, that work is the quiet way a 2034 landing gets decided. Dragonfly will spend its first years on Titan moving between dune fields and impact craters, drilling and analyzing samples for the chemistry that runs ahead of life rather than life itself (NASA Dragonfly fact sheet, Sept. 2026). If the autonomy holds and the pose mechanism behaves in flight, the thing that looked like an oversized drone in a Maryland clean room becomes the first powered aircraft to work on another world, and the missions designed after it will look different for that reason.