What actually changed in 2026
The short answer: X ranks the For You feed with a single model, Phoenix, that predicts how likely each person is to engage with a post and scores the post on those predictions. Replies, and especially replies the author takes part in, outweigh likes. Links cost reach because they send people off the platform, and most of a post's ceiling is set in its first thirty to sixty minutes, measured against the account's own recent baseline.
Most published guidance about the X algorithm still describes the system X open-sourced in 2023, Heavy Ranker, SimClusters, RealGraph, and a table of engagement weights. That architecture no longer exists.
In January 2026 xAI published a replacement at github.com/xai-org/x-algorithm, and on 15 May 2026 shipped the largest update yet, the first release where the actual For You ranker can be run locally. The old multi-service stack was collapsed into a single model.
On 13 August 2026 X published the For You timeline and the core ranking engine in full, under the Apache v2 licence. The codebase is reported at roughly ten to fifteen times its previous size, the addition being model configuration and the filter layer rather than new ranking logic.
The Phoenix scoring system can now be trained by researchers outside the company, and pull requests against the algorithm are open. LinkedIn made a structurally similar move in 2026, covered in how the LinkedIn algorithm works.
Read the next sentence before you rewrite anything, because a great deal of advice published this week gets it backwards.
This was a transparency release rather than a ranking change. X was explicit that the systems being opened are the ones already running, and nothing in the announcement alters how a post is scored. If your reach moved this month, the cause is on your side of the line.
One category stayed closed: the systems that use Grok to predict rule violations, withheld specifically so they cannot be gamed. Everything below was already true last week. It is now checkable rather than reported.
The pipeline, as published
Four components, per the repository:
- Home Mixer, the orchestration layer running the pipeline end to end.
- Thunder, an in-memory store of recent posts from accounts you follow.
- Phoenix, a Grok-based two-tower transformer that does both retrieval and ranking. This is the whole brain; it replaced Heavy Ranker, SimClusters and RealGraph together.
- Candidate Pipeline, the framework the retrieval stages are built on.
Roughly 500 million posts a day are narrowed to about 1,500 candidates per user, drawn from both accounts you follow and out-of-network discovery. Phoenix then scores each candidate by predicting how likely that specific person is to react to it.
What Phoenix predicts
The model runs 19 prediction heads, 15 positive signals and 4 negative ones:
| Positive signals predicted | Negative signals predicted |
|---|---|
| Favourites, replies, reposts, quote posts, clicks, profile visits, bookmarks, video views, video completion, photo expansion, shares, dwell time, follows | Not interested, blocks, mutes, reports |
Two things are worth drawing out. Dwell time and video completion are predicted directly, which is why dense posts that hold attention outperform one-liners. And four of nineteen heads are negative, the system is explicitly modelling how likely you are to annoy someone.
On the weights everyone quotes
You will see confident claims that a repost is worth exactly 20 likes, or a reply 13.5, or an author-engaged reply 150. Those numbers are not in the published code.
The numeric constants live in a params module that was not included in the release, in 2026 and in 2023 alike.
Every precise multiplier in circulation is either carried over from the 2023 leak or inferred by a third party. We are not going to repeat them as fact, and anyone presenting them as documented is overstating what is public.
What is reliable is the ordering, and it is consistent across both releases and with what we observe running launches every week:
Reposts and replies are worth substantially more than likes. Conversation the author takes part in is worth more again. Negative signals cost more than any positive signal gains.
That ordering is enough to plan against. The exact ratios would not change a single decision you make.
Velocity, and why the first hour matters
Phoenix scores each post relative to the person seeing it, and the strongest early evidence of engagement probability is engagement already happening. A post earning strong engagement quickly relative to its account’s norm gets tested wider; a slow start is capped.
Across the launches we run, the ceiling is effectively set inside the first thirty to sixty minutes. That is our own observation from campaign data rather than a published figure. X has never documented a window, but it has held consistently enough that we build every launch schedule around it.
Two consequences follow. Because the baseline is relative, small accounts are not disadvantaged in the way people assume, a founder account with 20 followers that suddenly gets real engagement is a stronger anomaly than a large account performing normally, which is how one such account reached 1.1M views on a single amplified article. And because timing is controllable while content quality is not, concentration beats volume.
The practical version of this for a launch is a calendar invite. Send one for the posting minute to every investor, advisor, friend and teammate, 50 to 100 people, and drop the link into it when the post is live.
Fifty real people replying and saving inside the first ten minutes is the anomaly the ranker is built to notice, and it is the cheapest one available.
What costs you reach
- Leading with an external link. Dwell time and on-platform engagement are predicted directly; a link is an instruction to leave. Carry the substance natively.
- Posting cold after a silence. A dormant account has a cold audience and competes against its own inactivity.
- Publishing and leaving. Replies the author engages with sit at the top of the ordering, and they are free.
- Anything that triggers the four negative heads. Engagement bait draws replies and mutes together, and the negatives are the more expensive side of that trade.
- A label you did not know about. Since 13 August 2026 you can check rather than guess. Settings has an "Under the Hood" page, and any account that has posted at least ten times in the past month can download a JSON file of aggregate statistics showing whether labels were applied to the account or to individual posts over the past calendar month. If the file is unreadable, paste it into a model alongside the public repository and ask it to explain what was applied.
- One post, one object. A post, a thread, a follow-up and a reply to a good critique give the ranker four things to evaluate instead of one.
What this means in practice
The 2026 system rewards concentrated early engagement from accounts whose audiences genuinely overlap yours, on content specific enough to generate replies, from an author who shows up in the thread.
Because the hand-engineered features are gone, tactics aimed at gaming them are gone too. What remains is unusually simple to state and hard to execute: be worth replying to, and be in front of the right people fast enough that the model notices. The sequencing that produces that is here.
Questions we get asked
How does the X algorithm work in 2026?
X ranks the For You feed with a single model, Phoenix, that predicts how likely each person is to engage with a post and scores the post on those predictions. Replies, and especially replies the author takes part in, outweigh likes. Links cost reach because they send people off the platform, and most of a post's ceiling is set in its first thirty to sixty minutes, measured against the account's own recent baseline.
What matters most?
Engagement velocity in the first stretch after posting, measured against your own baseline. Strong relative performance earns a wider test; a slow start is capped early and rarely recovers.
Do links reduce reach?
In practice yes. Ranking rewards on-platform engagement and dwell, and a link asks people to leave. Put the substance in the post and the link in a reply.
Does replying to your own audience help?
Substantially. Replies, especially ones the author takes part in, sit above likes in the ordering in both the 2023 and 2026 releases. The exact multipliers have never been public, but the ordering is consistent, and author participation is free.
How long is the window?
Most of it is decided in thirty to sixty minutes. Impressions peak within 48 hours, but the ceiling is set in the first hour.
Did the X algorithm change in August 2026?
No. On 13 August 2026 X published the For You timeline and the core ranking engine on GitHub under the Apache v2 licence, roughly ten to fifteen times more code than the previous release, adding model configuration and the filter layer.
It is a transparency release rather than a ranking change, and the company was explicit that the systems being opened are the ones already running. If your reach moved this month, the cause is not this.