The Hubble Tension

Summary

• Two good ways of measuring the expansion rate give different answers. Climbing a ladder of distance indicators out to supernovae, the SH0ES team gets H₀ = 73.2 ± 0.9 km/s per megaparsec.¹ Reading the expansion rate off the microwave background through the standard model, Planck gets 67.4 ± 0.5.² The gap is about nine percent.
• The disagreement is real. It has grown, not shrunk, as both sides got more precise. Quoted significances run from 4.5σ to 7.1σ depending on which pair is compared.^{3 4} The most obvious escape routes have been checked and closed.
• The microwave-background number is not a measurement of H₀. It is an inference. Planck's own abstract calls its H₀ one of "the inferred (model-dependent) late-Universe parameters."² Planck measures angles on the sky at redshift 1100 and converts them to an expansion rate today by assuming ΛCDM held for 13.8 billion years. Change the model before recombination and the number changes.
• The ladder has been audited hard and it holds up. JWST re-observed more than a thousand Cepheids in the same fields Hubble Space Telescope used. The mean distance difference is −0.01 ± 0.03 magnitudes, and unrecognised crowding of the HST photometry is ruled out at 8.2σ.⁵
• There is a real fight inside the ladder camp, and honesty requires reporting it. Wendy Freedman's Carnegie-Chicago programme measures 70.39 ± 1.22 ± 1.33 from the tip of the red giant branch, and 68.81 and 67.80 from JWST data alone.⁶ That is roughly halfway to Planck. Freedman's team also withdrew its own 2024 headline number after finding "an issue in the data pipeline and in the photometric reductions of the HST data."⁶
• Routes that skip the ladder entirely land all over the place, and none of them is yet precise enough to decide the question. Water megamasers give 73.9 ± 3.0.⁷ Lensed-quasar time delays give 71.6 (+3.9, −3.3).⁸ The one gravitational-wave standard siren with an optical counterpart gives 70.0 (+12.0, −8.0), and a 2026 reanalysis of the same event using radio imaging of the jet gives 65.5 ± 4.4.^{9 10} The inverse ladder, BAO plus big bang nucleosynthesis with no microwave background at all, gives 68.51 ± 0.58.³
• An honest example of a result that moved: time-delay cosmography. H0LiCOW announced 73.3 (+1.7, −1.8), a 2.4% measurement, in 2019.¹¹ When the same team relaxed one assumption about the shape of the lens galaxy's mass, the error bar roughly tripled and the central value slid to 67.4 (+4.1, −3.2).¹² Six years of new kinematic data later, the number is 71.6 with a 5% error. The field caught its own systematic, published it, and lost precision doing so.
• The tension may be better described as a disagreement about the sound horizon than about H₀. Licia Verde, Nils Schöneberg and Héctor Gil-Marín put it this way: "It appears to be a matter of anchors: the shape of the Universe expansion history agrees with the model, it is the normalizations that disagree."¹³ TDCOSMO's lenses plus DESI's BAO measure the sound horizon directly at 135.6 (+6.7, −6.1) megaparsecs, against Planck's model-derived 147.09 ± 0.26.^{8 2}
• Every proposed fix has a cost. Early dark energy raises H₀ but also raises σ₈, worsening a different disagreement, and ACT DR6 finds no evidence for it.^{14 15} Extra relativistic species are boxed in: N_eff = 2.86 ± 0.13, consistent with the standard three neutrinos.¹⁵ George Efstathiou's summary is blunt: "Any such theory must involve parameter degeneracies that combine fortuitously to mimic the 6-parameter ΛCDM cosmology to high accuracy. This is why I regard the Hubble tension as such a frustrating challenge to ΛCDM."¹⁶
• It is unsettled, and this article says so. Nobody knows whether the answer is a systematic error on one side or physics missing from the model.
• It is not evidence for a young universe, and no young-earth writer has built anything on it. Both candidate values give an age between 12.8 and 13.8 billion years. The arithmetic is shown below. A nine percent argument about the expansion rate cannot be turned into a factor of two million. Creationist use of the tension has been limited to Institute for Creation Research news pieces treating it as embarrassment for the Big Bang, and even those do not claim it yields a young universe.

This is the case file for a topic summarised in the standard model overview.

The two numbers

The Hubble constant, written H₀, is the current expansion rate of the universe.  Its awkward units, kilometres per second per megaparsec, encode a simple statement:  a galaxy one megaparsec away recedes at H₀ km/s, a galaxy two megaparsecs away recedes at twice that.  A megaparsec is 3.26 million light years.

Here are the numbers actually on the table in 2026, each with the error its own authors quote.

Sorted this way the table looks like a continuum.  It is not.  The two ends are each internally precise, and the middle entries are the ones with error bars wide enough to straddle both.

The significance depends entirely on which pair you pick, and the papers say so.  Louise Breuval and the SH0ES team quote 5.8σ against the CMB-inferred value.¹⁸  The South Pole Telescope's 2026 analysis quotes 6.2σ against SH0ES from SPT-3G data alone.¹⁷  The 117-author community consensus quotes 7.1σ against Planck plus SPT plus ACT.⁴  DESI quotes the number that matters most for anyone who suspects the microwave background:  "In ΛCDM the tension between the DESI+BBN and SH0ES H₀ results now stands at 4.5σ independent of the CMB."³

Take the lowest of those, 4.5σ.  That still corresponds to a chance coincidence of about one in 150,000.  Something is wrong somewhere.

The microwave-background number is an inference, not a measurement

This is the point most often lost, and getting it right changes how the whole dispute reads.

Planck did not measure the expansion rate.  Planck measured the angular sizes of hot and cold spots in a snapshot of the universe taken 380,000 years after the Big Bang.  The best-measured of those is the acoustic angle θ_*, known to 0.03%.  Turning an angle at redshift 1100 into an expansion rate today requires a model of everything that happened in between.  Planck used flat ΛCDM with six free parameters, and H₀ came out as a derived quantity.

Planck says this itself.  The abstract of the 2018 parameters paper introduces its H₀ with the words "the inferred (model-dependent) late-Universe parameters."²  The parenthesis is Planck's, not this article's.

Two consequences follow, and they cut in opposite directions.

The first cuts against people who treat the tension as a strike against the microwave background.  It is not.  The CMB data are not in dispute.  SPT-3G's 2026 paper puts it flatly:  "CMB data alone show no evidence for physics beyond LCDM."¹⁷  What is in dispute is the extrapolation from those data to today.

The second cuts against people who treat the CMB value as the settled truth.  If the model is wrong anywhere between recombination and now, or before recombination, the number moves.  That is exactly the loophole every proposed new-physics fix tries to use.

There is a third route worth flagging because it removes the microwave background from the argument entirely.  Take the size of the sound horizon from big bang nucleosynthesis plus a measured baryon density, then use baryon acoustic oscillations in galaxy surveys as a ruler.  DESI did this in 2025 and got 68.51 ± 0.58 with no CMB anisotropy data at all.³  The low camp does not depend on Planck.

The distance ladder, rung by rung

Nothing in astronomy measures a distance of 100 megaparsecs directly.  The high value comes from a chain, and each link is calibrated by the one below it.  Break any link and the number moves.  That is why the ladder gets audited so hard.

Rung one is geometry.  Three anchors carry known distances that owe nothing to astrophysics.  Gaia measures parallaxes to Milky Way Cepheids.  Detached eclipsing binaries in the Large and Small Magellanic Clouds give distances from orbital mechanics and stellar surface brightness.  And NGC 4258 has water masers orbiting its central black hole.

The NGC 4258 anchor deserves its own paragraph, because it is the cleanest geometric distance in extragalactic astronomy.  Water molecules in a warped disc around the black hole emit at 22 GHz.  Very long baseline interferometry resolves individual maser spots, tracks their line-of-sight velocities, and measures their accelerations.  Position plus velocity plus acceleration in a Keplerian disc yields the physical size of the orbit, and comparing physical size to angular size gives the distance.  No standard candle is involved.  Mark Reid and colleagues report 7.576 ± 0.082 (stat) ± 0.076 (sys) megaparsecs, about 1.5% precision.¹⁹

That single galaxy is enough to run the whole ladder.  Reid's paper does it:  "Using this as the sole source of calibration of the Cepheid-SN Ia distance ladder results in Ho = 72.0 +/- 1.9 km/s/Mpc, and in concert with geometric distances from Milky Way parallaxes and detached eclipsing binaries in the LMC we find Ho = 73.5 +/- 1.4 km/s/Mpc."¹⁹

Rung two is Cepheids.  Cepheid variable stars pulse with a period that tracks their luminosity, a relation Henrietta Leavitt found in 1912.  Measure the period, read off the intrinsic brightness, compare to apparent brightness, get the distance.  SH0ES measures Cepheids in the anchor galaxies and in galaxies that have hosted a Type Ia supernova, using the same instrument and the same three filters throughout so that zero-point errors cancel.²⁰

Rung three is Type Ia supernovae.  These are bright enough to be seen across hundreds of megaparsecs, and once standardised for light-curve shape and colour their brightnesses scatter by only 0.135 magnitudes in the Hubble-flow sample.²⁰  Calibrate their absolute magnitude using the handful whose host galaxies have Cepheid distances, then read H₀ off the redshift-magnitude relation of the far larger sample in the smooth Hubble flow.

The 2022 SH0ES paper is the reference version of this.  It used Cepheids in the hosts of 42 supernovae, "all suitable SNe Ia in the last 40 years at z<0.01, measured with >1000 orbits," and reported H₀ = 73.04 ± 1.04 including systematics.²⁰  The team also ran about 70 analysis variants, changing anchors, supernova surveys, redshift ranges, dust treatment, metallicity corrections and the form of the period-luminosity relation.  The headline number "lies near the median of all analysis variants."²⁰

A fourth anchor arrived in 2024.  Louise Breuval and colleagues measured 88 Cepheids in the core of the Small Magellanic Cloud with HST, modelling the cloud's depth, and obtained a period-luminosity scatter of 0.102 magnitudes, "approaching the width of the Cepheid instability strip."¹⁸  With the SMC alone the answer is 74.1 ± 2.1.  With all four anchors it is 73.17 ± 0.86.¹⁸  The SMC also doubles the metallicity range over which the Cepheid relation can be calibrated, which matters because a metallicity error that grows with distance is exactly the kind of thing that could fake a Hubble tension.  Their measured slope is γ = −0.234 ± 0.052 mag/dex.¹⁸

The checks on the ladder

If the high number is wrong, the error has to hide in one of those rungs.  Between 2019 and 2026 nearly every rung was checked with a different method, a different telescope, or a different team.  Here is what the checks found.

JWST versus HST on the same stars.  The most-feared systematic was crowding.  Cepheids in distant galaxies sit inside blended clumps of unresolved stars, and if HST's blending grew worse with distance it would inflate the inferred brightnesses and shrink the distances, faking a high H₀.  JWST's mirror is 2.7 times the diameter of HST's, so its images are 2.7 times sharper at the same wavelength.  In 2024 SH0ES observed more than 1,000 Cepheids with JWST in NGC 4258 and five hosts of eight supernovae, using two epochs to fix the pulsation phase and three filters to remove reddening.  The scatter in the period-luminosity relations fell by a factor of 2.5.  The distances did not move:  "We find no significant difference in the mean distance measurements determined from HST and JWST, with a formal difference of −0.01+/−0.03 mag."⁵  The paper rejects distance-dependent crowding "at 8.2 sigma, i.e., greater confidence than that of the Hubble Tension itself."⁵

Freedman's team disputes how much that 8.2σ covers, and the objection is worth stating:  "of the five galaxies for which they compare HST and JWST distance moduli, one is extremely nearby (NGC 5643 at 13 Mpc), two (NGC 1448 and NGC 5584) showed the smallest differences between their 2016 and 2022 distance moduli (0.01 - 0.02 magnitudes) and thus are the least likely to have significant crowding issues, and only one out of the five galaxies, NGC 5468, is at a distance of ~40 Mpc."⁶  In other words, most of the test's statistical weight rests on one distant galaxy.

Even so, the result deserves to be taken at face value.  It is a null result SH0ES did not have to publish in that form, and it closes the single most plausible way for the ladder to be wrong by nine percent.

Different stars entirely.  Cepheids are young, massive, and live in dusty spiral arms.  Three alternatives avoid all three problems.

The tip of the red giant branch is a sharp cutoff in the luminosity function of old, low-mass stars, set by the helium flash.  It lives in galaxy halos, away from dust and crowding.  Reid's NGC 4258 maser calibrates it at M_F814W = −4.01 ± 0.04, and running that ladder gives H₀ = 71.1 ± 1.9.¹⁹

J-region asymptotic giant branch stars are carbon-rich giants that occupy a narrow band in the near infrared.  They are a newer candle and their systematics are still being mapped.

Mira variables are old, low-mass pulsating stars.  Caroline Huang and colleagues found 115 oxygen-rich Miras in NGC 1559, host of SN 2005df, calibrated them against Miras in NGC 4258, and got H₀ = 72.7 ± 4.6 from that one supernova, or 73.3 ± 4.0 combining the maser and LMC calibrations.²¹  This was the first calibration of a Type Ia supernova by anything other than a Cepheid or a red giant tip.

Surface brightness fluctuations use the pixel-to-pixel graininess of an unresolved galaxy image, which scales with distance because more distant galaxies pack more stars into each pixel.  John Blakeslee and colleagues applied it to 63 mostly early-type galaxies out to 100 megaparsecs and found H₀ = 73.3 ± 0.7 (stat) ± 2.4 (sys).²²

The combined JWST cross-check.  In 2024 SH0ES compared all eight available combinations of method, sample and telescope against the HST Cepheid distances:  JWST Cepheids, tip of the red giant branch and JAGB by either group, plus HST red giant tip and Miras.  All eight agreed to about 1σ.  "Mean differences are ~0.03 mag, far smaller than the 0.18 mag 'Hubble tension'."¹  The linearity of the HST Cepheid distances came out at 0.994 ± 0.010, ruling out a distance-dependent bias at about 7σ.¹

Taken together, the ladder has been checked with three independent stellar populations, two telescopes, and two rival teams, and it has not moved.  Anyone arguing that the ladder is simply wrong now has to explain which of those checks failed and how.

The fight inside the ladder camp

Popular accounts present the Hubble tension as SH0ES against Planck.  That is not the whole story.  A second distance-ladder team gets an answer in between, and the two teams have been arguing in print since 2019.

Wendy Freedman's Carnegie-Chicago Hubble Program, now the Chicago-Carnegie Hubble Program, deliberately avoids Cepheids as its primary candle.  Its published results calibrate Type Ia supernovae with the tip of the red giant branch and with JAGB stars, anchored on NGC 4258, the same maser galaxy SH0ES uses.  Its numbers are lower.  The best-precision result is H₀ = 70.39 ± 1.22 (stat) ± 1.33 (sys) ± 0.70 (σ_SN) from the red giant tip with 24 supernova calibrators.  From JWST data alone the red giant tip gives 68.81 ± 1.79 ± 1.32 and JAGB gives 67.80 ± 2.17 ± 1.64.⁶  Freedman's team draws the obvious conclusion:  "Our results are consistent with the current standard LambdaCDM model, without the need for the inclusion of additional new physics."⁶

Then there is the part that a fair article has to report, because it cuts both ways.

CCHP withdrew its own headline number.  The first version of the 2024 preprint gave three methods and a combined value:  "Ho = 69.96 +/- 1.05 (stat) +/- 1.12 (sys) km/s/Mpc," plus a Cepheid value of "72.05 +/- 1.86 (stat) +/- 3.10 (sys)."²³  The published version, seven months later, drops both.  The reason is stated plainly:  "In the case of the Cepheid distance scale, subsequent tests after unblinding revealed an issue in the data pipeline and in the photometric reductions of the HST data.  A re-analysis of these data is on-going."⁶  As of August 2026 the replacement Cepheid analysis has not appeared.

That is a team correcting itself in public, which is to its credit.  It also means the widely circulated "69.96" figure should not be quoted as a current CCHP result.  It has been retracted by its own authors.

Where the two teams actually disagree.  It is not mainly about the stars.  SH0ES compared its HST Cepheid distances against all eight available JWST and HST alternatives and found mean differences of about 0.03 magnitudes, when the tension itself is 0.18 magnitudes.¹  The distances agree.

The disagreement is about which supernovae get used and how they are calibrated.  SH0ES argues that the gap between the CCHP JWST programme and the full HST sample, about 2.5 km/s/Mpc, "is entirely consistent with differences in sample selection," because CCHP's early JWST subsample is roughly a quarter the size of the full HST set.¹  Small samples wander.

CCHP's own supernova recalibration paper makes the same point from the other side, and is unusually candid about it.  Taylor Hoyt and colleagues measured JWST red giant tip distances to ten supernova hosts and found that the answer depends on whose supernova magnitudes you adopt:  "yielding 68.4 < H0 < 69.6 km/s/Mpc depending on which of four sets of SN magnitudes are used."²⁴  It also depends on how many calibrators you use.  Expanding their sample from 11 supernovae to 24 raises the Carnegie Supernova Project II answer by 0.8 km/s/Mpc, a 1.4σ shift, but raises the Pantheon+ answer by 2 km/s/Mpc, a 3.1σ shift.²⁴  Their closing sentence is the honest one:  "More JWST observations of the TRGB as well as independent analyses of low-redshift SNe are needed to continue unraveling the true nature of the Hubble Tension."²⁴

A second live disagreement is the calibration field itself.  Both teams anchor on NGC 4258, but they do not always measure the same anchor value.  Riess and colleagues note "a net difference of 1.3 km s−1 Mpc−1 (or 0.04 ± 0.02 mag) between measurements by two groups of the location of the tip in NGC 4258."²⁰  For JAGB stars the problem is worse:  Siyang Li and colleagues find a 0.11 ± 0.02 magnitude difference between CCHP's JAGB measurements in two different fields of NGC 4258, and the same field-to-field wandering in SH0ES fields.²⁵  Their JAGB result, H₀ = 73.3 ± 1.4 (stat) ± 2.0 (sys), carries a systematic error "dominated by apparent differences across NGC 4258 calibrating fields or their measures."²⁵  When a single anchor galaxy gives different answers in different patches of itself, nobody's error bar is safe.

And CCHP has a pointed argument about SH0ES's own history.  Between the 2016 and 2022 SH0ES analyses, the same raw HST data were reprocessed and the mean distance modulus for the original 19 calibrator galaxies moved by 0.035 ± 0.025 magnitudes.  Freedman's team observes:  "if one retroactively corrects the R16 distances to the R22 scale correcting for the 0.035 mag offset, the SHoES H0 value becomes 72.05 ± 1.74 km s−1 Mpc−1, resulting in a lower (2.6σ) and much less interesting Hubble tension."⁶

That is a fair jab, and it is also not a refutation.  A reanalysis that changes an answer is what reanalysis is for.  But it shows that the tension's significance has depended on choices that have moved before.

One thing not to do.  The 2026 community consensus paper, H0 = 73.50 ± 0.81, is signed by 117 authors and represents a serious covariance-weighted combination of parallaxes, eclipsing binaries, masers, Cepheids, the red giant tip, Miras, JAGB stars, supernovae, surface brightness fluctuations, the Fundamental Plane and Tully-Fisher.⁴  It does not include Freedman, Madore, Jang, Hoyt or Lee.  Presenting it as "the community" against a lone dissenter misrepresents the state of play.

George Efstathiou, who is on neither team, states the position that this article endorses:  "Evidently more work needs to be done to achieve a consensus between Freedman and collaborators and the SH0ES team."¹⁶

Routes that skip the ladder entirely

If the ladder is the problem, methods that do not use it should agree with the microwave background.  Four such methods exist.  None of them is decisive yet, and saying otherwise in either direction is overclaiming.

Water megamasers

The NGC 4258 technique works on other galaxies too, and when the galaxy is far enough away its recession velocity is dominated by the Hubble flow rather than by local motions.  Then the geometric distance and the redshift give H₀ in one step, with no candles at all.

The Megamaser Cosmology Project applied this to six galaxies and reported H₀ = 73.9 ± 3.0 km/s/Mpc, "independent of distance ladders and the cosmic microwave background."⁷  The weakness is peculiar velocities:  each galaxy's motion relative to the Hubble flow has to be modelled, and different treatments span 71.8 to 76.9.⁷  The team's own confidence that the local value exceeds the early value is 95 to 99%, which is two to three sigma, not five.

Megamasers are the cleanest ladder-free method available, and only six galaxies have so far yielded a usable measurement.

Standard sirens

A binary neutron-star merger encodes its own luminosity distance in the shape of its gravitational-wave chirp.  General relativity fixes the amplitude, so the distance falls out of the waveform without any astrophysical calibration.  If an electromagnetic counterpart identifies the host galaxy, the redshift comes for free, and H₀ follows.

GW170817 is the only such event so far.  The LIGO-Virgo measurement is H₀ = 70.0 (+12.0, −8.0).⁹  That error bar is enormous, and it is enormous for a specific reason:  distance and orbital inclination are degenerate in the waveform.  A face-on distant merger and an inclined nearby one produce nearly the same signal.

The obvious fix is to measure the inclination another way.  Very long baseline interferometry watched the merger's jet move across the sky, which constrains the viewing angle.  A 2026 reanalysis by Kelly Gourdji and colleagues fitted all the radio data jointly with the gravitational-wave distance and reported viewing angle 16.8 to 19.2 degrees, D_L = 44.0 ± 1.6 Mpc, and H₀ = 65.5 ± 4.4.¹⁰  Their own framing:  "the peak of our H0 posterior is within 0.5σ of the early-Universe Planck H0 value, but 1.7σ from the late-Universe SH0ES measurement."¹⁰

That is one event, reanalysed, with a 7% error bar, and it depends on hydrodynamic jet models.  It is not a resolution.  It is a reminder that the siren constraint is soft enough to have moved from 70 to 65.5 without anyone doing anything wrong.

Dark sirens, which use mergers with no counterpart and statistically weight all candidate host galaxies, are softer still.  Combining the fourth gravitational-wave catalogue with the Dark Energy Survey Year 6 galaxy catalogue gives H₀ = 71.7 (+22.7, −18.3), or 73.1 (+11.9, −8.5) with GW170817 added.²⁶  Those bars cover both camps and most of the space outside them.

Gravitational waves will eventually settle this.  They do not settle it in 2026.

Time-delay cosmography, and a result that moved

This is the most instructive case in the whole dispute, and it belongs in any honest account.

When a quasar sits behind a massive galaxy, gravitational lensing produces several images.  Light takes different paths, so a flare in the quasar appears in each image at a different time.  The delays are measured in days to years by monitoring.  Their size scales inversely with H₀.  Combined with a model of the lens galaxy's mass, they give an absolute distance with no ladder and no CMB.

In 2019 the H0LiCOW collaboration published six lensed quasars and reported "H₀ = 73.3(+1.7, −1.8), a 2.4% precision measurement," in "3.1σ tension with Planck," rising to 5.3σ when combined with the distance ladder.¹¹  It looked like the tension's independent confirmation.

Then the same people attacked their own assumption.  Lensing observables are unchanged by the mass-sheet transform:  add a uniform sheet of mass across the lens and rescale the source, and every image position, shape and flux ratio stays exactly the same, while the time delays and therefore H₀ change.  Lensing alone cannot see it.  H0LiCOW had suppressed the degeneracy by assuming the lens mass follows a power law or a stars-plus-standard-halo profile.

TDCOSMO IV replaced that assumption with a flexible profile constrained only by stellar kinematics.  The result:  "We apply the method to the TDCOSMO sample of 7 lenses (6 from H0LiCOW) and measure H₀ = 74.5(+5.6, −6.1)."¹²  Same central value, error bar roughly tripled.  Adding 33 non-time-delay lenses from the SLACS survey to pin down the mass profiles pulled the answer down:  "From the joint analysis of the TDCOSMO+SLACS sample, we measure H₀ = 67.4(+4.1, −3.2)."¹²

A 2.4% measurement at 73.3 had become an 8% measurement that could be read as 67.4.  The team's own verdict was careful:  "While our new analysis does not statistically invalidate the mass profile assumptions by H0LiCOW, and thus their H₀ measurement relying on those, it demonstrates the importance of understanding the mass density profile of elliptical galaxies."¹²

This is what a systematic looks like when a field finds one in its own work.  Nobody was caught cheating.  An assumption that seemed safe turned out to carry most of the error budget.  Anyone reasoning about the Hubble tension should hold the possibility that the same thing is waiting inside one of the other methods.

Six years of new data later, the number has firmed up again.  TDCOSMO 2025 used JWST, Keck and VLT stellar kinematics for eight lensed quasars, keeping the conservative treatment of the mass-sheet degeneracy, and found H₀ = 71.6 (+3.9, −3.3) in flat ΛCDM, tightening to 4.6% with the SLACS and SL2S lens samples added.⁸  That sits between the camps and excludes neither.  A 2026 single-lens analysis with a maximally flexible mass sheet gives 73.2 (+4.8, −4.7) and finds the internal mass-sheet parameter λ_int = 1.12 (+0.05, −0.06), two sigma from unity, which is a direct measurement that the degeneracy is not negligible.²⁷

The inverse distance ladder

Run the ladder backwards.  Instead of calibrating supernovae with nearby stars, calibrate them with a ruler of known length placed at high redshift.

The ruler is the sound horizon.  Before recombination, the universe was a plasma in which pressure waves travelled at about 58% of the speed of light, c divided by the square root of three.  The distance such a wave covered before the plasma broke up is imprinted as a preferred separation in the clustering of galaxies, the baryon acoustic oscillation.  Its length in megaparsecs can be computed from the baryon density and the radiation density, both of which follow from big bang nucleosynthesis plus a deuterium measurement.  No microwave background required.

DESI did exactly this with 14 million galaxies and quasars.  Combining BAO with a BBN prior on the baryon density gives H₀ = 68.51 ± 0.58, and with the CMB acoustic angle added, 68.45 ± 0.47.³  This is the measurement that makes the tension impossible to blame on Planck.  DESI states the resulting significance in its own words:  "In ΛCDM the tension between the DESI+BBN and SH0ES H₀ results now stands at 4.5σ independent of the CMB."³

DESI also states the one assumption that remains, in the same paragraph, and it is the crux of the whole subject:  "Note that the DESI+BBN result does assume standard pre-recombination physics to determine r_d through eq. (2)."³

Is it a tension in H₀, or in the sound horizon?

The question in the heading is not word-play.  It changes what kind of explanation is even eligible.

Every low-side measurement, whether it comes from Planck, from the South Pole Telescope, or from DESI's BAO with no CMB at all, shares one ingredient:  the length of the sound horizon r_d, about 147 megaparsecs, computed from pre-recombination physics.  It is the ruler that converts an observed angle into a distance.  If the ruler is the wrong length, every measurement that uses it is wrong in the same direction by the same fraction, and they will all agree with each other while disagreeing with the ladder.

José Luis Bernal, Licia Verde and Adam Riess made this point in 2016, before the tension was famous:  "The tension in H₀ reflects a mismatch between our determination of r_s and its standard, CMB-inferred value."²⁸

Kevin Aylor and colleagues measured the mismatch directly.  Use supernovae calibrated by the classical ladder to set absolute distances, then read the sound horizon off galaxy clustering.  Compare that to the value ΛCDM computes.  Their result:  "we find the model-based sound horizon to be larger than the empirically-determined one with a statistical significance of between 2 and 3σ, depending on the dataset."²⁹

TDCOSMO's 2025 analysis does it again with a completely different absolute distance.  Anchoring DESI's BAO on lensed-quasar time delays instead of on supernovae, they measure "the sound horizon scale, 135.6(+6.7, −6.1) Mpc, independent of assumptions about early-Universe physics," about 1.8σ below Planck's model-derived 147.09 ± 0.26.^{8  2}

So the shape of the expansion history is agreed.  What is disputed is the length of the ruler used to convert it into absolute numbers.  Verde, Schöneberg and Gil-Marín state it best:

It appears to be indeed a problem of anchors: because the shape of the expansion history is tightly constrained by current data in a model-independent way, the cosmic ladder is quite rigid; it is the early-time anchor r_d (and possibly also the equality scale) calibrated on early-time physics that does not agree with the late-time "anchor" of H₀.¹³

This framing explains why almost all serious theoretical proposals attack the era just before recombination rather than the era after it.  Lloyd Knox and Marius Millea surveyed the possibilities and concluded:  "We single out the set of solutions that increase the expansion rate in the decade of scale factor expansion just prior to recombination as the least unlikely."³⁰  Note the phrasing.  Not "likely."  Least unlikely.

One dissent worth reporting.  Ioannis Pantos and Leandros Perivolaropoulos assembled 88 sound-horizon-free H₀ determinations in 2026 and argue that the usual early-versus-late framing is wrong.  Sorting them by method rather than by epoch, the 30 distance-ladder measurements give 72.73 ± 0.39 and the 58 measurements that are either ladder-independent or sound-horizon-free give 69.37 ± 0.34, a 6.5σ gap that falls to a minimum of 3.9σ once correlations are accounted for.  Their conclusion:  "the Hubble tension is better characterized as a discrepancy between the Distance Ladder and all other methodologies, rather than an early-vs-late-time split."³¹

If that reading is right, the problem is on the ladder side after all, and the sound horizon is innocent.  Their own paper also reports a 2.9σ internal disagreement among the ladder-independent methods, so it does not settle anything either.  It is a live minority view and the reader should know it exists.

What would resolve it

There are only two kinds of answer.  Either somebody's measurement is wrong in a way nobody has found, or the standard model is missing physics.  Both remain open in 2026.

An unrecognised systematic

On the ladder side, the candidates have been narrowing for seven years.  Crowded-field photometry is dead as an explanation, killed by JWST at 8.2σ.⁵  Distance-dependent bias is dead at about 7σ.¹  Metallicity is now calibrated over twice the range it used to be.¹⁸  What survives is unglamorous:  which supernovae enter the calibrator sample, whose photometric compilation is used for their magnitudes, and how the anchor galaxy NGC 4258 is measured field by field.  Those are exactly the things Hoyt and Li found moving results by one to three sigma.^{24  25}  A systematic hiding there would have to be worth about 4 km/s/Mpc, which is much larger than any single effect yet demonstrated, but it is not obviously impossible.

On the early side the position is harder for the sceptic.  Three separate instruments with different optics, different observing sites and different foregrounds, Planck in space, ACT and SPT on the ground, agree to a fraction of a percent.^{17  15}  And DESI's BAO with a nucleosynthesis prior reaches the same answer using no CMB anisotropy at all.³  An instrumental systematic would have to be shared by all of them, which is close to saying it is not instrumental.

The remaining early-side vulnerability is not the instruments but the modelling assumption they all share:  standard pre-recombination physics fixing the sound horizon.

New physics before recombination

The target is easy to state.  Shrink the sound horizon by about eight percent and the CMB-inferred H₀ rises to meet the ladder.  Any way of adding energy density in the last decade of expansion before recombination will do it, since a faster expansion gives the sound wave less time to travel.

Early dark energy is the leading proposal.  Vivian Poulin, Tristan Smith, Tanvi Karwal and Marc Kamionkowski proposed a scalar field that "behaves like a cosmological constant at early times (redshifts z≳3000) and then dilutes away like radiation or faster at later times," and showed it can bring the Planck-inferred H₀ into agreement with SH0ES.³²

Then the costs showed up.  Colin Hill and colleagues found that fitting EDE to CMB plus SH0ES data changes the matter power spectrum, adding "10% more power at k = 1 h/Mpc," which raises σ₈ and worsens the separate disagreement with weak-lensing and galaxy-clustering surveys.¹⁴  Their conclusion is unusually direct for a title:  "We conclude that the EDE scenario is, at best, no more likely to be concordant with all current cosmological data sets than ΛCDM, and appears unlikely to resolve the H₀ tension."¹⁴

The EDE proponents' own 2023 review concedes the theoretical awkwardness, listing "a second 'cosmic coincidence' problem in particular" among the challenges, and ends where the evidence actually stands:  "current data remain inconclusive on the existence of an EDE phase."³³  A field that switches on and switches off at precisely matter-radiation equality, for no reason connected to anything else in physics, is a fix that needs its own explanation.

Efstathiou's assessment after reviewing the fits:  "EDE as a solution of the Hubble tension is quite strongly disfavoured by the data."¹⁶

Extra relativistic species were the first idea and are the most cleanly testable.  Adding dark radiation raises the expansion rate before recombination and shrinks the sound horizon.  Bernal, Verde and Riess found in 2016 that "a component of dark radiation with an effective number of species around 0.4 would reconcile the CMB-inferred constraints, and the local H₀," but noted in the same abstract that adding the high-multipole Planck polarization data "disfavours this solution."²⁸

That is now settled against.  Extra relativistic species also change the damping of small-scale CMB fluctuations, which the ground-based telescopes measure precisely.  ACT DR6 finds N_eff = 2.86 ± 0.13, tightening to 2.89 ± 0.11 with external BBN data, against the standard-model prediction of 3.044.¹⁵  There is no room for ΔN_eff = 0.4.  Planck plus BAO had already given 2.99 ± 0.17.²

Modified recombination and a varying electron mass are the surviving members of the family.  SPT-3G's 2026 analysis reports that combining CMB with DESI BAO "drives mild preferences for models that address the Hubble tension through modified recombination or variations in the electron mass in a non-flat universe."¹⁷  "Mild" is the operative word, and requiring a non-flat universe to make it work is a heavy price.

The general verdict from the most precise ground-based data is blunt.  ACT DR6's extended-models paper found no evidence for early dark energy, dark radiation, modified recombination, or early-universe variation of fundamental constants, and summarised:  "In general, models introduced to increase the Hubble constant or to decrease the amplitude of density fluctuations inferred from the primary CMB are not favored by our data."¹⁵

Late-time solutions are in worse shape

Changing the expansion history after recombination cannot work, because the shape of that history is well measured and matches the model.  Rong-Gen Cai and Shao-Jiang Wang's decade review states the two late-time options and what happened to them:  resolutions that work "by changing the absolute magnitude of supernovae either intrinsically or effectively" are "strongly constrained by the inverse distance ladders with the cosmic distance duality relation," and "a local Hubble bubble or cosmic void solution has long been ruled out as a significant contribution to the Hubble tension."³⁴

Evolving dark energy, the headline result of DESI DR2, does not help either.  DESI says so itself:  "It is worth noting that allowing w(z) to vary does not help resolve the so-called Hubble tension, since in w₀w_aCDM the recovered H₀ is lower than the Planck ΛCDM value."³  Two anomalies do not cancel.

The constraint every proposal has to satisfy

Efstathiou states it as well as anyone:

If it is argued that new physics beyond ΛCDM is required to explain, for example, distance scale measurements of the Hubble constant (the 'Hubble tension'), galaxy weak lensing measurements, (the 'S8 tension'), evolving dark energy, or bulk flow anomalies … then that new physics must reproduce the CMB anisotropies of the base ΛCDM cosmology to extraordinarily high precision.¹⁶

And his summary of where that leaves the field:

In summary, observational data probing both high and low redshifts set such strong constraints that it is difficult to construct a plausible theory to match the SH0ES value of H₀.  Any such theory must involve parameter degeneracies that combine fortuitously to mimic the 6-parameter ΛCDM cosmology to high accuracy.  This is why I regard the Hubble tension as such a frustrating challenge to ΛCDM.¹⁶

Cai and Wang, reviewing ten years of it, call the tension "likely … a real crisis for modern cosmology."³⁴  Notice that every resolution they then review is internal to the hot Big Bang.  A crisis in cosmology is not the same as a crisis for cosmology.

The honest state of the question:  nobody knows.  Anyone who tells you the tension is definitely a systematic, or definitely new physics, is ahead of the evidence.

What the tension does not do

An apologetics site has an obvious temptation here.  Two teams of cosmologists disagree about the expansion rate at six sigma, so surely the whole edifice is shaky and a young universe is back on the table.

That argument does not work, and this article will not run it.

Both candidate values give an age near 13 billion years

The arithmetic is short enough to do in public.  In a flat universe with matter and a cosmological constant, the age is

t0 = (2 / (3 H0 sqrt(OmegaL))) * asinh( sqrt(OmegaL / OmegaM) )

where Ω_Λ = 1 − Ω_m, and 1/H₀ expressed in gigayears is 977.79 / H₀ for H₀ in km/s/Mpc.

Check the formula against a published answer first.  Planck's parameters, H₀ = 67.36 and Ω_m = 0.3153, give 13.80 Gyr.  Planck's own table says 13.797 ± 0.023 Gyr.²  The formula is right.

Now feed it the other end of the tension.  The community consensus value H₀ = 73.50 with Ω_m = 0.30 gives 12.83 Gyr.  With Ω_m = 0.315 it gives 12.65 Gyr.  The SH0ES value 73.2 gives 12.88 Gyr.  Freedman's 70.39 gives 13.39 Gyr.

The entire dispute moves the age of the universe by about one billion years out of thirteen.  Every value on the table is between 12.6 and 13.8 billion years.

There is a cruder bound that needs no model at all.  The Hubble time, 1/H₀, is an upper limit on the age only for a universe that has always decelerated, and a lower limit for one that has always accelerated.  At H₀ = 73.5 the Hubble time is 13.3 billion years.  Even Einstein-de Sitter, the flat matter-only model that was standard until the 1990s and has since been ruled out, gives two-thirds of that, 8.9 billion years.

To reach an age of ten thousand years, H₀ would have to be about 9.8 × 10⁷ km/s/Mpc.  That is 1.3 million times the measured value.  The gap between the two camps is nine percent.  Nine percent and a factor of 1.3 million are not the same kind of disagreement, and no amount of citing the tension bridges it.

Independent age measurements sit where you would expect

Stellar physics can date the oldest stars without any cosmology at all.  David Valcin and colleagues fitted the full colour-magnitude diagrams of 68 globular clusters and derived an age of the universe of 13.5 (+0.16, −0.14) (stat) ± 0.23 to 0.33 (sys) Gyr, which they note "agrees well with 13.8 ± 0.02 Gyr, the cosmological model-dependent value inferred by the Planck mission assuming the ΛCDM model."³⁵

Notice which side of the tension that supports.  A stellar age of 13.5 billion years sits comfortably with the low camp's 13.8 and awkwardly above the high camp's 12.8.  If anything, the high H₀ makes the old globular-cluster age problem slightly worse, not better.  It certainly does not point downward toward thousands of years.

Raul Jimenez and colleagues ran the argument in the other direction, using stellar ages plus a late-time matter density to infer the expansion rate:  "A proof-of-principle of this approach with current data gives H₀ = 71 ± 2.8 (H₀ = 69.3 ± 2.7) km/s/Mpc from globular clusters (very-low-metallicity stars)."³⁶  A method that knows nothing about the CMB, nothing about Cepheids and nothing about supernovae lands between the two camps.

What young-earth writers have actually done with it

Very little, and to their credit they have not overclaimed.

The Institute for Creation Research is the main user.  Jake Hebert's 2019 piece "Big Bang Hubble Contradiction Confirmed" reports the discrepancy accurately and makes one point this article agrees with entirely:  the CMB route "implicitly assumes that the Big Bang model is correct."³⁷  He is right about that.  His broader complaint, that "secular scientists almost never allow data to count against the Big Bang, only for it," is a fair worry in general and a weak fit to this particular case, since the tension is a disagreement the field itself found, published, amplified and has been unable to make go away.³⁷

Hebert's 2023 piece frames it as "another indication that all is not well with the Big Bang model."³⁸  Note that even this does not claim a young universe.  It claims embarrassment.

Beyond ICR, the argument is barely used.  A full crawl of creation.com's English sitemap in 2026 found no published article using the modern Hubble tension against the Big Bang.  Answers in Genesis has no article on it;  Danny Faulkner discusses H₀ uncertainty without the tension framing.  John Hartnett, who argues for a non-expanding universe, does not run the argument either.

That silence is the right instinct.  A creationist who leans on the Hubble tension is arguing that cosmologists cannot measure the expansion rate to better than nine percent, while simultaneously needing them to be wrong about it by a factor of a million.  The argument undercuts itself.

What the tension is actually evidence for

It is evidence that ΛCDM's calibration of the pre-recombination sound horizon may be wrong, or that a systematic remains buried in the distance ladder.  Those are the two live options, and both are internal to the Big Bang.  Every proposed fix in the literature, early dark energy, extra radiation, modified recombination, evolving dark energy, interacting dark energy, keeps the hot dense early universe and adjusts a detail of it.

That is worth saying plainly, because it is the same shape as most anomaly arguments in apologetics.  An unexplained discrepancy inside a model is not an argument for abandoning the model unless somebody has a rival model that explains the discrepancy better.  For the Hubble tension, no young-earth cosmology explains it at all.  None of them predicts it, and none of them accounts for why two mainstream methods differ by nine percent rather than by nothing at all.

What would settle it

Four things are in progress, and any of them could break the deadlock in the next few years.

The CCHP Cepheid reanalysis.  Freedman's team pulled its Cepheid number in 2025 and said a reanalysis was under way.⁶  When it lands, the two ladder teams will be comparing the same candle on the same telescope, and the remaining gap will be attributable to supernova samples alone.

More JWST supernova hosts.  Both teams' JWST samples are small.  SH0ES's own analysis shows that a quarter-size sample can wander by 2.5 km/s/Mpc from sampling alone.¹  Doubling the sample kills that argument in one direction or the other.

More lensed quasars with resolved kinematics.  Time-delay cosmography is the only ladder-free method whose precision is improving fast.  TDCOSMO went from 8% to 4.6% between 2020 and 2025 by adding JWST and Keck stellar kinematics.⁸  At 2% it becomes a referee.

Standard sirens.  A handful more neutron-star mergers with optical counterparts would give a percent-level H₀ with no astrophysical calibration at all.  This is the cleanest possible arbiter, and it is limited only by how often such mergers happen close enough to localise.

Watch also for the next generation of CMB damping-tail measurements, which tighten N_eff further and squeeze the remaining pre-recombination fixes.

Objections and responses

"The Hubble tension shows the Big Bang model is falling apart."

It shows one number in the model is not reproducible by two routes.  Everything else the model predicts still works:  the blackbody spectrum of the microwave background, the acoustic peak positions, the light-element abundances, the baryon acoustic scale in galaxy clustering, supernova time dilation.  Notice also that both sides of the tension are computed inside the same model.  There is no rival framework in the dispute at all.  What is under strain is the calibration of the sound horizon, and every published fix keeps the hot dense early universe intact.

"The microwave-background value is just the output of a model fit, so it should not count."

The first half is true and Planck says so:  its H₀ is one of "the inferred (model-dependent) late-Universe parameters."²  The second half does not follow.  DESI's baryon acoustic oscillations plus a big-bang-nucleosynthesis prior on the baryon density give 68.51 ± 0.58 using no CMB anisotropy data whatsoever, and that is still 4.5σ from SH0ES.^{3  3}  Throwing out Planck does not move the low camp.

"SH0ES is one team with a career stake in a high value."

Then check the others.  Megamasers, which use no candles at all, give 73.9 ± 3.0.⁷  Surface brightness fluctuations give 73.3 ± 0.7 ± 2.4.²²  Miras give 73.3 ± 4.0.²¹  Lensed quasars give 71.6 (+3.9, −3.3).⁸  A 117-author consensus spanning eleven distance indicators gives 73.50 ± 0.81.⁴  The high camp is not one team.

"But Freedman gets a lower number, so the tension is going away."

This is the strongest version of the sceptical case and it deserves a straight answer.  CCHP's best value, 70.39, combined in quadrature over its three quoted errors, is about 1.9 wide.  Against Planck's 67.4 ± 0.5 that is roughly a 1.5σ difference, which is no tension at all.  So if CCHP is right, the crisis dissolves.

Two things stop that from being the end of the story.  First, CCHP's own supernova recalibration shows its answer swings between 68.4 and 69.6 depending on which supernova magnitude compilation is used, so the method is not yet stable at the precision required.²⁴  Second, SH0ES's cross-comparison finds the two teams' distances agree to about 0.03 magnitudes, so the difference lives in supernova selection rather than in the stars.¹  The disagreement is real and unresolved.  Anyone claiming it is settled either way is guessing.

"Nobody can agree on the expansion rate, so the 13.8-billion-year age is worthless."

The age barely depends on the disputed quantity.  Across the entire range under dispute, 66.7 to 73.5 with Ω_m = 0.315 to 0.30, the flat-ΛCDM age runs from 13.95 down to 12.83 Gyr.  Independent stellar dating of globular clusters gives 13.5 (+0.16, −0.14) ± 0.23 Gyr with no cosmology assumed.³⁵  Three unrelated methods agree on the age to within about eight percent.  That is the opposite of worthless.

"Two anomalies at once, the Hubble tension and DESI's evolving dark energy, mean the model is collapsing."

They do not combine.  DESI addressed this directly:  "allowing w(z) to vary does not help resolve the so-called Hubble tension, since in w₀w_aCDM the recovered H₀ is lower than the Planck ΛCDM value."³  Adding the second anomaly makes the first one worse.  Two problems that pull in opposite directions are not evidence for one solution.

"This is a posteriori statistics.  Look at enough numbers and something will be off at five sigma."

That objection has real force against the microwave background's large-angle anomalies, which were spotted after the fact.  It has almost none here.  H₀ was the target of both measurements before either was made.  SH0ES and CCHP both blind their analyses, revealing the answer only after the pipeline is frozen.  The H0LiCOW lenses were analysed blind.  TDCOSMO 2025 was blind.  And the discrepancy has grown as the errors shrank, which is the opposite of what a statistical fluke does.

"Does the tension support a young universe?"

No.  A universe of 6,000 years needs H₀ near 1.6 × 10⁸ km/s/Mpc.  The dispute is over whether the answer is 67 or 73.  No young-earth cosmology predicted the tension, explains it, or is helped by it, and the one young-earth organisation that has written about it at any length, ICR, claims only that it embarrasses the Big Bang.  They are right that it is a genuine unsolved problem.  They do not claim it yields a young universe, and neither should anyone else.

Sources