I've sat across from 12 HU bots. Most weren't worth the seat.
Honestly? My take after testing 12 heads-up bots over the last 18 months is that only two are worth the lobby-time. The rest fold to 3-bets like they're scared, or they snap-call rivers with second pair because nobody patched their bluff-catching logic since 2023. This is my running tier list, format by format.
I rate on a 1–10 scale that I'll defend (see my criteria). A 7 means "I'd grind this if the rake were reasonable." A 5 means "exploitable inside three orbits." Anything under 4 I just stopped logging hands against, life is short.
My 2026 HU tier list at a glance
| Tier | My score | What's in it |
|---|---|---|
| S — actually scary | 8.5–9.0 | 1 cash HU bot. Solver-aligned, mixes properly, adapts within ~150 hands. |
| A — competent | 7.0–8.4 | 2 bots. One cash, one SNG. Tight technically, won't deviate much. |
| B — beatable but not free | 5.5–6.9 | 4 bots. Predictable bet-sizing trees, ignores blockers. |
| C — print money | 4.0–5.4 | 3 bots. River bluff-catchers are broken. Just value bet thin. |
| F — don't bother | <4 | 2 bots. Timing tells, action-distribution skew, the works. |
Tier list, broken out by format
Cash HU and SNG HU are different games and I refuse to put them on one ladder. Pick a tab.
Cash HU — where the algorithms actually have to think
Deep-stack heads-up cash is the hardest test. The bot has to size on three streets, balance ranges across a tree that branches like crazy, and recover after taking a 60bb hit. Top score in my book: 9.0 / 10. One bot. It's solver-aligned, mixes its check-raise frequency on wet boards somewhere near 32% (where the solver wants it), and it actually reduces bluff-catch frequency when I overbet rivers twice in a row. That's the deviation behavior I want to see.
The rest of the cash field hovers 5–6.5. They play a static GTO-ish strategy and don't budge. Once you find their bet-sizing seams (most use only 33% / 75% / pot, not the half-pot the solver actually loves on certain boards), you print.
★★★★★★★★★☆ 9.0 — the one A-tier cash HU bot
★★★★★★★☆☆☆ 7.0 — runner-up, slightly tilted toward overfolding turns
★★★★★★☆☆☆☆ 6.0 — the rest of the pack
HU SNG — push-fold simulators with a thin pre-bubble layer
I'd put HU SNG bots in a different tier from HU cash. The SNG ones I've seen are basically push-fold engines bolted to a shallow open-raise lookup table. That's not a complaint, exactly — push-fold is closer to solved than 100bb cash play, so a 7/10 SNG bot can feel solid for 80% of the match.
Where they leak: the 25–40bb stage. Some bots play the same 15bb-equilibrium ranges at 32bb, which is a known overfold zone. I tested one bot for ~1,200 hands at 30bb effective and its 4-bet frequency was 1.8%. Solver wants closer to 4.5%. That's exploitable.
★★★★★★★★☆☆ 8.0 — best SNG bot, near-perfect endgame
★★★★★★☆☆☆☆ 6.0 — the rest are formulaic mid-stack
HU MTT — still rough, honestly
HU MTT (final-table heads-up after a tournament has run) is the format with the worst bots, in my honest opinion. The ICM layer is where they fall apart. A bot calibrated for chip-EV will gladly stack off in a spot where any human pro is folding because the pay-jump math says so. I've watched three different HU MTT bots make the same ICM-blind shove. Same hand, same stack split, same mistake. Suspicious.
Best score I'd give an HU MTT bot is a 6.5, and it's a soft 6.5. Most are 4–5.
★★★★★★☆☆☆☆ 6.5 — the least bad option
★★★★☆☆☆☆☆☆ 4.0–5.0 — the field
How I scored these
Four dimensions, weighted equally. Timing consistency. Action-distribution shape across at least 600 hands. Willingness to deviate from GTO when I exploited it. Lobby-grinding behaviour (how it picks seats, how it leaves). Full breakdown on my criteria page, with the format-by-format scores on format comparison.
Want a heads-up game with something better than what's in the public lobbies? I run a small group that fields HU bots privately. Quiet operation, no marketing list.
Match up with us