By Marvin Analysts

Nerding Out on GTA VI Sales Numbers

By Alex Hoffmann, Co-Founder and CEO
as of:

GTA VI launched on 19 November 2026 with expectations to match a decade of trailers. Sell-side estimates going into launch sat at about 30M units for the December quarter1, and the guidance Take-Two Interactive ($TTWO) reiterated on 9 November was consistent with about 28.9M copies in that quarter on the arithmetic in our 2Q FY27 memo.

In this note, we look at the pre-launch and launch-day numbers and build a model for how sales develop after launch. As George Bernard Shaw almost said: those who can't play, work in a spreadsheet.

Fair warning: this is a nerding-out piece. It contains a Kalman filter, a binary-option pricing model, a Metropolis sampler and some opinions about press-release cadence. It is not a recommendation on anything, least of all on a prediction market contract.

Sensor Tower and the Edition Mix

Sensor Tower's GTA VI sales tracker publishes cumulative digital pre-order units and revenue for the PlayStation and Xbox storefronts. Its last pre-launch observation, for 18 November, stood at 24.59M units and $2,360M of revenue.

GTA VI comes in two editions: a $79.99 Standard Edition and a $99.99 Ultimate Edition. With units KK and revenue LL, the average price per copy pins down the edition mix, since the Ultimate share is s=(L/K−79.99)/20s = (L/K - 79.99)/20. On the 18 November figures that gives $95.98 a copy and a 79.9% Ultimate share.

This cumulative figure is dominated by the game's most die-hard fans, who pre-ordered months in advance. GTA VI already had 2.2M digital pre-orders on 25 June and passed 5M in early September. It is no big surprise that most of those fans chose the Ultimate Edition.

Later buyers, and especially those buying after launch, are more likely to pick the Standard Edition. So the cumulative pre-order split is not our best estimate of the mix going forward. The mix of recent buyers sits in the increments, ΔL/ΔK\Delta L / \Delta K, and that is where the data gets noisy. Sensor Tower rounds to 0.01M units and $0.01M, so a small increment carries a rounding error worth several points of mix.

This is a textbook Kalman filter problem: a slowly drifting hidden state observed through noisy measurements of varying informativeness.

  • State. The current-buyer Ultimate share xtx_t follows a random walk, xt=xt−1+wtx_t = x_{t-1} + w_t, with Var(wt)=0.022⋅Δt/7\text{Var}(w_t) = 0.02^2 \cdot \Delta t / 7, so the mix can drift about 2 points a week.
  • Observation. Each revenue increment is ΔLt=79.99 ΔKt+20 ΔKt xt+vt\Delta L_t = 79.99\,\Delta K_t + 20\,\Delta K_t\, x_t + v_t.
  • Noise. Each increment involves four rounded figures (units and revenue at both dates), each uniform on ±0.005, with unit rounding scaled by the $99.99 price: R=4⋅0.01212⋅(1+99.992)≈0.33 ($M)2R = 4 \cdot \tfrac{0.01^2}{12} \cdot (1 + 99.99^2) \approx 0.33\ (\$\text{M})^2.

The Kalman gain decides how much each increment counts. A 0.03M-unit increment in August gets a Kalman gain near zero and barely moves the estimate. The 5.54M-unit increment on 18 November, $519M of revenue in one day, dominates everything before it.

Late buyers chose Ultimate far less often than early ones
Ultimate share of digital pre-orders, %, derived from Sensor Tower (PS + Xbox)
Source: Sensor Tower, Marvin Labs

The filtered share ends at 69.9% on 18 November, with a 95% band of 69.0-70.9%. The band is tight because the last observation is huge. Earlier in the series the band runs from the high 70s to around 100%, which is the filter's polite way of saying it has no idea. Both lines show Ultimate share drifting down from about 90% in the summer as the buyer base broadens, then dropping sharply in the final week before launch, which matches our expectations. Sensor Tower publishes every other day until 8 November and daily after that, and the chart uses every observation.

Reading the Take-Two Filing Closely

On 20 November, Take-Two filed a press release with two numbers: 32.5M units sold and $3.79B in sales.

Divide $3.79B by 32.5M and you get $116.62 a copy. With the Ultimate Edition at $99.99, something doesn't add up.

A close reading of the press release explains the gap. The 32.5M is units "sold pre-release". The $3.79B is day-one retail sales through "end of day November 19th". The dollar figure includes launch day, and the unit figure stops before it.

That gives us a way to estimate launch-day spending: price the 32.5M pre-orders at the edition mix derived above, subtract the result from $3.79B, and what is left is launch-day spending. Divide by the launch-day average price and you get launch-day copies. Depending on the mix assumption, that comes to 7-8M units sold on launch day.

Exhibit 1: The dollar figure implies 7-8M copies sold on launch day, whichever pre-order mix you believe
Launch-day buyers priced at the Kalman mix (69.9% Ultimate, $93.97 a copy)
Pre-order mix assumptionUltimate sharePre-order $BLaunch-day $BLaunch-day units (M)Units by end of 19 Nov (M)
Naive cumulative0.803.120.677.1439.64
Blended: cumulative to 18 Nov, Kalman after0.773.100.697.3139.81
Kalman throughout0.703.050.747.8340.33
Source: Take-Two 8-K (20 November 2026), Sensor Tower, Marvin Labs
Pre-order dollars are 32.5M copies at $79.99 plus $20.00 times the Ultimate share. Launch-day dollars are $3.79B less pre-order dollars.
The blended row applies the 79.9% cumulative share to Sensor Tower's 24.59M and the 69.9% Kalman share to the remaining 7.91M.

Typical Post-Launch Trajectories

With an edition mix, an average price per copy and an estimate for launch day in hand, the next question is how sales develop after launch. The best guide is Rockstar's own history. Take-Two's filings give dated cumulative sell-in figures for the last two big launches:

TitleLaunchDays after launchUnits sold-in (M)Source
GTA V17 Sep 201342nearly 29Q2 FY14 release
GTA V139more than 32.5Q3 FY14 release
GTA V238more than 33Q4 FY14 release
GTA V322more than 34Q1 FY15 call
RDR226 Oct 201812over 17Q2 FY19 release
RDR2103more than 23Q3 FY19 release
RDR2199more than 24Q4 FY19 call
RDR2283approximately 25Q1 FY20 call

These figures are sell-in, so boxed copies count when shipped to retailers, which front-loads them. They are quoted "to date", meaning the release date, not the quarter end. They are also rounded floors ("more than 32.5 million"), so each point is a lower bound.

A hyperbolic sales curve fits both titles well: cumulative units N(t)=N0+B⋅tt+τN(t) = N_0 + B \cdot \frac{t}{t + \tau}, where post-launch sales are half done after τ\tau days. The fit gives τ\tau = 38 days for GTA V and 41 days for RDR2. With three parameters and four points per title, τ\tau is loosely pinned, but the two titles agree, and that agreement becomes the prior for GTA VI below.

Kalshi Data and the Implied Option Distribution

The Kalshi Event Contracts

Kalshi lists two events on GTA VI: units reported before 1 January 2027 and units reported before 1 March 2027. Each event is a ladder of strikes, "Take-Two reports at least N million units sold-in", from 3M to 50M on the January event and up to 100M on the March event, plus a "No Data" strike that pays if Take-Two publishes nothing.

Each strike is a cash-or-nothing digital option: it pays $1 if the reported figure clears the strike and nothing otherwise. Its price is the market's probability of that. Read across the ladder, the prices trace out a survival function, P(reported≥K)P(\text{reported} \ge K), and differencing adjacent strikes gives the probability mass in each bin.

The Prediction Market Trap: Observability and What the Contracts Actually Measure

The contracts settle on reported units, meaning the units Take-Two chooses to publish, observed only on the days Take-Two chooses to say something. If Take-Two publishes its 31 December unit count on 1 January, one day late, it does not count for the January contract.

Each strike price is therefore the product of two probabilities, with SS the units actually sold:

P(YESK)=P(S≥K)×P(published∣S≥K)P(\text{YES}_K) = P(S \ge K) \times P(\text{published} \mid S \ge K)

Sales are very likely above 40M already, so the first factor is close to one for the 33M and 36M strikes. Those two contracts trade at 66.5% and 67.0%, within half a point of each other. Their price is therefore a clean market-implied estimate of the second factor: the probability that Take-Two publishes a new figure before year end, about 67%.

Take-Two's 8-K record has precedent for an early second press release:

TitleLaunchOff-cycle releases after launchFirst earnings figure after launch
GTA V17 Sep 201318 Sep: over $800M first-day sales. 20 Sep: over $1B in three days29 Oct 2013: nearly 29M sold-in
RDR226 Oct 2018None filed7 Nov 2018: over 17M sold-in
GTA VI19 Nov 202620 Nov: $3.79B day-one sales, 32.5M units pre-releaseQ3 FY27 release, early Feb 2027

Sources: Take-Two press releases above and Take-Two's 8-K filings on EDGAR.

GTA V got a release the day after launch and a second one three days in, when sales passed $1B. Both were dollars only. RDR2 got no standalone release at all. GTA VI's release carried both dollars and units, on the slightly different bases set out in the filing section above. That is the reason to assume any further release would also carry a unit figure, so this piece sets P(units disclosed∣release)=1P(\text{units disclosed} \mid \text{release}) = 1 and treats the January contract as a bet on whether a release happens at all.

The prices around 20 November show the market adjusting to this:

Exhibit 2: The market priced a 60% chance of any figure before year end, then a 67% chance of a second one
Kalshi YES mid-prices, %
Before releaseAfter release
Contract and strike20 Nov, 12:00 UTC21 Nov, 07:34 UTC
1 Jan: any figure published (1 less No Data)59.5%100.0%
1 Jan: at least 33M46.5%66.5%
1 Jan: at least 36M36.5%67.0%
1 Jan: at least 42M22.5%60.0%
1 Jan: at least 50M14.0%43.0%
1 Mar: any figure published (1 less No Data)94.5%95.0%
1 Mar: at least 42M81.5%90.0%
1 Mar: at least 50M66.5%81.5%
Source: Kalshi KXGTA6UNITS-01JAN27 and KXGTA6UNITS-01MAR27, Marvin Labs
Mid of bid and ask. The 20 November release settled the 1 January No Data strike NO and every strike up to 30M YES. The 1 March No Data figure after the release is the last trade.
The 1 March 50M strike traded at 0.74 bid and 0.89 ask on 21 November, so its mid carries a wide band.

Before the release, the market gave 59.5% to any figure before year end. Conditional on one, it expected a number around 38-39M (P(≥33M∣figure)=78%P(\ge 33\text{M} \mid \text{figure}) = 78\%, P(≥36M∣figure)=61%P(\ge 36\text{M} \mid \text{figure}) = 61\%). Take-Two delivered 32.5M the day after launch, before counting launch-day sales.

The March ladder is different. Take-Two reported fiscal third-quarter results in the first week of February after both GTA V and RDR2, each time with a unit figure, and the No Data strike trades at 5%. The March contract is effectively a bet on the fiscal Q3, calendar year-end figures as published in early February.

Precedents: How People Price Kalshi Ladders

Reading option prices as probability distributions is old news in derivatives. Breeden and Litzenberger showed in 1978 that the second derivative of call prices with respect to strike gives the risk-neutral density. A digital option is the limit of a tight call spread, the first derivative of call prices with respect to strike, so a ladder of digitals hands you the distribution after one differencing step instead of two. Kalshi's ladders are about as clean a version of this as exists.

Three papers shape how to read them:

  • Diercks, Katz and Wright, "Kalshi and the Rise of Macro Markets" (FEDS 2026-010). The Fed economists treat each Kalshi contract as an Arrow-Debreu security and build full risk-neutral densities for macro outcomes like CPI and the fed funds rate. Two of their practical choices matter here. They enforce monotonicity in the CDF by building outward from the mode, so stale tail prices cannot contaminate the middle of the distribution. And they prefer last trades to bid-ask midpoints, because tail spreads are occasionally very wide. They are also careful that these are risk-neutral (Q-measure) probabilities, which can differ from real-world (P-measure) probabilities by a risk premium.
  • Bürgi, Deng and Whelan, "Makers and Takers: The Economics of the Kalshi Prediction Market". Using about 157,000 Yes contracts, they find Kalshi prices are informative but show a favourite-longshot bias: cheap contracts win less often than their price implies, expensive ones slightly more. In a Mincer-Zarnowitz regression the bias is modest, with smaller coefficients for entertainment markets, and it is largest at the extremes of the price range. A belief-shrinkage model with traders overweighting small probabilities (β=0.09\beta = 0.09) and modest disagreement reproduces the pattern.
  • Wolfers and Zitzewitz, "Interpreting Prediction Market Prices as Probabilities". The classic answer to whether prices are probabilities at all. Under a broad class of models, prediction market prices sit close to the mean belief of traders, with the favourite-longshot bias as the known exception. Le (2026) updates the calibration evidence across Kalshi and Polymarket.

None of this literature deals with the GTA VI wrinkle, an underlying that is only observed when the company decides to speak. The closest analogy is an option on a discretely monitored asset where the monitoring dates are themselves random. The model below handles that by adding explicit disclosure probabilities.

Putting It All Together: A Model for GTA VI Units

The model has three layers:

  • Sales path. N(t)=32.5+D+A⋅h(t)−h(1)h(tM)−h(1)N(t) = 32.5 + D + A \cdot \frac{h(t) - h(1)}{h(t_M) - h(1)} for t≥1t \ge 1 days after launch, with h(t)=t/(t+τ)h(t) = t/(t + \tau). DD is the launch-day bump, with a prior of 7.8M ± 1.5M from Exhibit 1. AA, lognormal, is the units added from day two to the February release at tMt_M = 77 days. The decay scale τ\tau gets a prior centred on the GTA V and RDR2 fits.
  • Disclosure. With probability qJq_J, Take-Two publishes a figure on some date between 21 November and 31 December. With probability qMq_M, the February release gives one. The January ladder prices qJ⋅P(N≥K)q_J \cdot P(N \ge K). The March ladder prices the February figure, or a December figure if February is silent.
  • Likelihood. All 23 active strikes enter at their bid-ask midpoint, with a per-strike error of the half-spread plus 2c. The parameters are sampled with random-walk Metropolis, and the outputs are posterior predictive, so they carry parameter uncertainty.

Unlike the Fed paper, the model uses midpoints rather than last trades. Many GTA VI strikes trade only a few hundred contracts a day, so last trades are often hours old, and the half-spread weighting already discounts the strikes with wide spreads.

Exhibit 3: The market-implied path reaches 64M units by year end and 75M by the fiscal year end
Cumulative GTA VI units sold-in, M, posterior predictive, as of Kalshi prices on 21 November 2026, 07:34 UTC
Date25th percentileMedian75th percentile
End of launch day, 19 Nov 202638.739.740.5
End of Q3 FY27, 31 Dec 202655.164.479.9
Q3 release, assumed 4 Feb 202758.970.689.8
End of FY27, 31 Mar 202761.775.196.9
Source: Kalshi, Take-Two, Sensor Tower, Marvin Labs
Percentiles are computed date by date and do not describe a single path.
Fitted parameters: P(second release before 1 January) 67%, P(any figure before 1 March) about 92%, decay scale 30 days (5th-95th percentile 15-56), launch-day bump 7.3M (4.9-8.8).
The market-implied path reaches 64M by the end of Q3
Cumulative GTA VI units sold-in, M, Kalshi prices as of 21 Nov 2026
Source: Kalshi, Take-Two, Marvin Labs

A few things the fit says on its own:

  • The launch-day bump comes back at 7.3M, close to the 7.8M from the filing arithmetic. The market and the press release agree on launch day without being told to.
  • The decay scale is 30 days, a bit faster than GTA V and RDR2 at 38 and 41. A digital-heavy launch in 2026 front-loading more than boxed launches in 2013 and 2018 is not a shocking result.
  • The model reproduces the January flat spot. It prices 33M, 36M and 39M all at 66.6%, against 66.5%, 67.0% and 63.5% in the market.

The honest caveats:

  • The tail is the weakest part. It rests on a lognormal assumption and on the March 60M-100M strikes, where spreads ran from 3c to 37c on 21 November and volume is a few thousand contracts per strike, against tens of thousands on the January ladder. The mean (86M at 31 March) is much less trustworthy than the median.
  • No holiday bump. The curve is smooth through December. The analog prior absorbs some of this, since GTA V and RDR2 both had a Christmas in their fits.
  • Debiasing barely matters. Applying the Bürgi-Deng-Whelan corrections moved the median by under 3M on the 20 November prices, because most strikes trade between 30c and 90c, away from the extremes where the bias lives.
  • Rounding. Take-Two reports rounded floors, so the contracts price the published figure, which runs 1-2M below the true count.
  • Prices move. These numbers are a snapshot of Kalshi on 21 November 2026 at 07:34 UTC and will be stale by the time anyone reads this.

Conclusion

Each data source measures something different. Sensor Tower measures digital buyers, and a Kalman filter says the late ones bought Ultimate about 70% of the time. Take-Two's filing measures pre-release units and launch-day dollars, and the gap between them is worth about 7-8M copies. Kalshi measures what Take-Two will choose to say, which makes the January ladder a bet on a press release and the March ladder a bet on the fiscal third-quarter results. Stitched together, they put GTA VI at a median of about 64M units by year end, more than double the 30M the quarter was supposed to deliver.

Footnotes

  1. Take-Two's fiscal year ends on 31 March, so FY2027 runs from April 2026 to March 2027. The calendar year end is the end of the fiscal third quarter (3Q FY27), and the fiscal year end is 31 March 2027. ↩

Alex Hoffmann
by Alex Hoffmann

Alex is the co-founder and CEO of Marvin Labs. Prior to that, he spent five years in credit structuring and investments at Credit Suisse. He also spent six years as co-founder and CTO at TNX Logistics, which exited via a trade sale. In addition, Alex spent three years in special-situation investments at SIG-i Capital.

Part of our editorial coverage. Read more from By Marvin Analysts.