Today is the day of the women's semifinals at Flushing Meadows. In 48 hours, we'll have two champions. The right moment to take stock: how did our picks hold up from Round 1? Which surprises did our model miss? And more importantly — what does it say about what's coming next? The finals this weekend, and the Asian swing starting September 28.
Key takeaways - Our tennis model hit 65.8% on 114 official moneyline picks over the past 30 days, for +7.56u (August 11 – September 7, 2026) - Biggest surprise of the tournament: Ben Shelton eliminates Alcaraz (defending champion) in the QF in a match that ended at 3:33 AM local time - WTA SF today: Sabalenka-Pegula, Gauff-Rybakina. ATP SF Friday: Shelton-Tiafoe — an all-American final in the making at Arthur Ashe - China Open (Beijing, Sept 28 – Oct 11) and Japan Open (Tokyo, Sept 30 – Oct 6): our next hunting ground, fast indoor hard courts that favor big-server profiles
Picks review: how did ProbWin do at US Open 2026?
Since the start of the main draw (late August), our model published official picks throughout the tournament. The numbers speak for themselves: over the past 30 days across all tournaments, our tennis AI ran at 65.8% on 114 moneyline picks, for +7.56 units. The US Open accounts for the majority of that volume.
| Period | Picks | Win rate | PnL |
|---|---|---|---|
| Last 30 days (all tournaments) | 114 | 65.8% | +7.56u |
Sample: August 11 – September 7, 2026. Official moneyline picks only (is_official = 1). Performance varies by tournament and variance is inherent to any probabilistic model.
The model tends to perform better at Grand Slams than at ATP 250s: deeper draws, denser historical data, and Flushing Meadows conditions (fast outdoor hard) match profiles well documented in our five-year dataset.
What our picks got right — and wrong
Real examples from our US Open 2026 pick database:
WINS
Van de Zandschulp vs Gea (R16) — backed VdZ at 1.65 -> WIN +0.65u
Ben Shelton vs Tsitsipas (R16) — backed Shelton at 1.40 -> WIN +0.40u
Alejandro Tabilo vs Popyrin (R64) — backed Tabilo at 1.85 -> WIN +0.85u
Luciano Darderi vs Svrcina (R64) — backed Darderi at 1.65 -> WIN +0.65u
Rinderknech vs Munar (R64) — backed Rinderknech at 1.58 -> WIN +0.58u
LOSSES
Cobolli vs Alexander Blockx (R32) — backed Cobolli at 1.93 -> LOSE -1u
Rublev vs Daniel Merida Aguilar (R64) — backed Rublev at 1.74 -> LOSE -1u
Nakashima vs Alex Michelsen (R64) — backed Nakashima at 1.52 -> LOSE -1u
The Rublev loss against Merida Aguilar is typical of a known model limitation: ATP players ranked 80-150 outside the top circuit have thin historical records in our hard-court database. That's a documented improvement axis for the 2027 season.
The Cobolli loss against Blockx points to a different issue: Alexander Blockx (Belgian, 22) showed service stats above what his odds reflected on fast hard. Our model lacked recent data on this specific profile to flip the pick his way.
Rejoignez ProbWin gratuitement
Accédez à nos prédictions IA pour MLB, NHL et NFL. Inscription en 30 secondes.
The big shock: Shelton takes down the defending champion
Carlos Alcaraz — defending champion and pre-tournament favorite — out in the quarterfinals against Ben Shelton. Final score: 6-7, 6-1, 6-3, 1-6, 7-6 in a match that finished at 3:33 AM Eastern time. That's the defining moment of the 2026 US Open.
Our model had Alcaraz as a clear favorite going in. This outcome illustrates a principle we return to often:
At a Grand Slam, the variance of a single match can wipe out
weeks of probability built around a tournament favorite.
Shelton flagged as dangerous outsider on fast hard:
-> among the best ace rates on the ATP Tour
-> dominant first serve at Flushing (fast surface conditions)
-> exceptional mental profile under pressure
But his win probability over 5 sets was still below 35%.
The market had Shelton around 3.00 before the QF. That's not the kind of outsider our model systematically backs — but the surface_speed and service_dominance parameters had correctly flagged him as a real threat. The takeaway: big servers at Flushing Meadows have an upset capacity that standard regression models consistently underestimate.
Semifinals: what our AI says
WTA — today, September 10
| Matchup | Market favorite | Key factors in our model |
|---|---|---|
| Sabalenka vs Pegula | Sabalenka | Cross-court backhand power decisive on fast hard |
| Gauff vs Rybakina | Rybakina slight favorite | Serve-and-volley game; Gauff's defensive improvement since Wimbledon |
Sabalenka-Pegula is a rematch of the 2024 final. Our model rates Sabalenka slightly above market odds in this configuration: fast hard amplifies her power advantage on the backhand side. Pegula remains a credible finalist — her return game is among the three best on the WTA Tour.
Gauff-Rybakina is the tightest call. Elena Rybakina (2026 Australian Open champion) brings a dangerous serve-and-volley on hard courts, but Coco Gauff has visibly upgraded her defense. Our model separates them by less than 5 probability points.
ATP — Friday, September 11
Ben Shelton against Frances Tiafoe: two Americans in a US Open semifinal. One factor the market tends to misprice:
Shelton: night match (QF ended at 3:33 AM) -> recovery window < 48h
Tiafoe : came back from 0-2 sets vs Michelsen -> also played 5 sets
Surface: fast hard -> amplifies the edge of the better server
Model signal: on fast hard, first serve in % and ace rate
take over in SF — "narrative momentum" flattens out.
This is exactly the type of matchup where our model spots a gap between market odds (shaped by the storyline of each player's run) and the underlying service metrics. Our picks for this SF are live on our tennis page.
The Asian swing: what's next on the calendar
Once the US Open wraps, the tour moves straight into Asia. Two events to watch closely:
| Tournament | City | Surface | Dates | Category |
|---|---|---|---|---|
| Japan Open | Tokyo | Fast indoor hard | Sept 30 – Oct 6 | ATP 500 |
| China Open | Beijing | Fast indoor hard | Sept 28 – Oct 11 | ATP 500 (M) / WTA 1000 (W) |
Why indoor hard changes the equation
Moving from outdoor Flushing to indoor Beijing or Tokyo amplifies service variables. On fast indoor hard, aces typically rise 15-20% compared to the US Open outdoor. Our model adjusts surface_speed and rally_exchange_weight parameters per tournament, calibrated on five years of historical data.
In practice: big-server profiles priced at 1.30-1.60 tend to offer better expected value on fast indoor than on outdoor — not because the model is sharper, but because the market slightly underprices the service advantage in these specific conditions.
China Open (Beijing, Sept 28 – Oct 11) is also the first WTA 1000 post-US Open. The finalists playing this weekend will enter Beijing with very fresh form data in our system. That's a real informational edge in the first 48 hours after the draw is released.
Japan Open (Tokyo, Sept 30 – Oct 6) is an ATP 500 with a smaller draw — fewer picks, but better concentration of profiles our model knows well.
For a deeper look at our hard-court analysis methodology, read our complete US Open AI method guide. And for background on the summer hard-court season: our hard court AI predictions guide.
Frequently asked questions
Did your AI predict Alcaraz's exit at US Open 2026? No. Our model had Alcaraz as the tournament favorite going in. Shelton was flagged as a dangerous outsider (top ace rate, dominant serve on fast hard), but a five-set win in those conditions was still a sub-35% probability event. That's the nature of Grand Slams: even the best models can't eliminate single-match variance.
Who does your AI see winning US Open 2026? Our live projections for the semifinals and finals are updated on our tennis page. For the WTA SF: Sabalenka rates slightly higher in our model. For the ATP SF: Shelton-Tiafoe is very tight, with post-match recovery being our primary variable.
When will your model publish picks for the China Open? The China Open draw (Beijing, Sept 28) triggers our pipeline automatically. First picks expected around September 24-25, once odds and entries are stable.
How does your AI adapt from US Open (outdoor) to China Open (indoor)? Our model uses a surface_speed parameter calibrated per tournament on five years of data. Moving outdoor-to-indoor increases the weight of service variables (first serve in %, ace rate) and reduces the weight of rally exchange. The model recalibrates automatically — no manual tuning.
Why post a picks review before the tournament ends? Because the data is there now. We don't choose the most flattering time window — we show results in real time, including the losses. Transparency is the product. The full post-tournament review (after the Sept 12-13 finals) will be published early next week.
Next step
The US Open 2026 finals are this weekend — women's final Saturday September 12, men's final Sunday September 13. Our semifinal picks are live on the ProbWin tennis page.
To understand how our model approaches Grand Slam hard-court analysis, read our complete methodology guide. And catch up on our QF and SF analysis: US Open 2026 — our AI predicts the finalists.
The Asian swing starts September 28. Subscribe to get our first China Open and Japan Open picks as soon as they're published — it's one of the best value windows of the entire tennis calendar.
