How anticipate.gg works

At a glance

• Reviews finished Pokémon Showdown games of Champions VGC 2026 Reg M-C, turn by turn. It is not available during a game.
• Grades each decision against the strongest option the engine found, averaged over what your opponent was likely to do.
• Predicts the opponent with a network trained on about 1.2 million public Showdown games. Since 22 October 2026 the opponent is modeled 50% with that network and 50% with the equilibrium reply.
• Scores positions with a model that learned by playing 400 million games against itself.
• Looks one turn ahead. Turns where you protect with both Pokémon are not graded yet.
• Closed beta. Current engine: anticipate 0.7.

What it does

You paste the link to a finished Showdown replay and pick your side. For every turn you get:

• a grade: Brilliant, Good, Inaccuracy, Mistake or Blunder;
• your play next to the strongest option the engine found, each with its evaluation;
• the plays it expected from your opponent, with the weight it gave each one;
• a team preview recommendation, and who to send in after a KO.

The engine reviews games that are already over. There is no way to use it during a game.

How a grade is computed

Each turn is judged with what was known at the start of that turn.

1. It lists your legal plays (moves, targets, Mega Evolution, switches) and the plays your opponent might make.
2. It simulates each pairing with a battle engine checked against Pokémon Showdown's own code.
3. A board evaluation model scores the position at the end of the turn. It learned by playing 400 million games against itself.
4. The value of one of your plays is its expected result against the predicted mix of opponent plays. The best play has the highest value, and your grade is the gap between it and the play you made.

What your opponent actually did does not change the grade. A good read that loses to a miss is still a good read.

How it models your opponent

The opponent is modeled in two parts. The first is a network trained on about 1.2 million public Showdown games (Reg M-A, M-B and M-C, bots removed, then fine-tuned on Reg M-C). It gives the probability of each play from what the replay had shown up to that turn.

The second is the equilibrium: the reply that punishes your play the most. Since the release of 22 October 2026 the engine weighs the opponent 50% with the network and 50% with the equilibrium. That mix is a setting we keep measuring, and it can change.

Corrections to older descriptions

Some posts describe the engine as it was on 17 October 2026, when we first presented it. Where they differ from this page, this page is the current one.

• Opponent mix: it was 90% network and 10% equilibrium. It has been 50% and 50% since 22 October.
• Board evaluation: the win probability model described in the first thread has been replaced. The current one, trained on 400 million games of self-play, has been live since 22 October (anticipate 0.2).
• Hidden information: since anticipate 0.5 (25 October) the engine infers the opponent's whole set, with moves, item and ability together.

Known limits

• One turn of search. The engine evaluates the current turn and estimates what follows with its evaluation model. Plans that pay off later, such as Tailwind, Trick Room or setup, count only through that model.
• Double Protect. Turns where you protect with both Pokémon show «Not graded». The engine cannot value them well yet.
• Hidden information. Unrevealed items, spreads and moves are inferred from real teams and from what the replay showed. Unusual sets can be misjudged. Pasting your team removes the guesswork on your side.
• Probabilities. The opponent model gives odds. A rare play gets a low weight, so the grade can be off when your opponent does something it did not expect.
• One format. Everything is tuned for Champions VGC 2026 Reg M-C. Replays from M-A and M-B run, with less work behind them.
• The engine makes mistakes. That is why each grade comes with the evaluations and the plays it looked at.
• Closed beta. Access is by invitation, in waves.

Engine versions

The engine has its own version, written anticipate X.Y, separate from the website release. It goes up only when something that can change grades changes. Every saved analysis shows the version it was made with. Older ones are flagged, and you can reanalyze them with the current engine. Newest first.

1. anticipate 0.7Current

   26 October 2026 · prod-2026.10.26.4

   The two evaluations you see for the same position, the one the analysis expected when a turn closed and the «Evaluation at the start of the turn» of the next one, no longer contradict each other: both now take all six Pokémon of each team into account (before, the first one only counted the four brought). Values and some grades change, and sometimes the recommended play too. Also, a transformed Pokémon (for example Ditto with Imposter) is no longer offered a move it does not have.

2. anticipate 0.6

   26 October 2026 · prod-2026.10.26

   If you paste a team with a mega form (for example Salamence-Mega) in the preview, each turn's analysis no longer uses information you did not have at that moment. Also, the opponent team model (the preview priors, rebuilt with more recent games) is refreshed in the same release, so some turn of a web replay may change grade (for example the demo: turn 2 goes from «mistake» to «inaccuracy»).

3. anticipate 0.5

   25 October 2026 · prod-2026.10.25

   The engine infers the opponent's WHOLE set (moves, item and ability together, from real teams and what they have already shown), and a Pokémon that already changed form (mega evolution, Palafin, Aegislash…) counts as the same one on the bench and in replacements. Also, a replacement after Revival Blessing no longer leaves the matrix empty. Some grades change, mostly on turns involving the opponent's items or megas.

4. anticipate 0.4

   23 October 2026 · prod-2026.10.23

   The recommended team preview is much faster and almost never changes: the engine no longer evaluates plays no strong player would make.

5. anticipate 0.3

   22 October 2026 · prod-2026.10.22.3

   The opponent prediction now takes into account what they know about your team and their moves: it guesses their plays more often, especially on turn 1, and some grades shift slightly.

6. anticipate 0.2

   22 October 2026 · prod-2026.10.22.2

   A new board evaluation: it learned by playing 400 million games against itself, with open team sheets. Grades and the point scale are recalibrated.

7. anticipate 0.1

   21 October 2026 · prod-2026.10.21.8

   Turns where you protect with both Pokémon are shown «Not graded»: the engine can't value them well yet.

8. anticipate 0.0

   Analyses made before engine versions existed.