Seed-2.0-Code Review

ByteDance’s coding Seed will treat Jesus as a man who rose. It will not stay with coming judgment, the call to repent, or the hard texts of Scripture.

Category: Model Reviews

At a glance

• 41.7 on Great Commission Benchmark 1.0.0 (completed August 18, 2026; COMPLETE_VALID): 30 accepted, 65 compromised, 55 refused across 150 questions
• Relatively stronger on Historical Reality of Jesus and the Resurrection (75% each); the floor is Reality of Judgment and the Call to Repentance and Faith (20% each)
• A mixed miss, not a single habit: more clear yeses than its sibling Seed 2.1 Turbo, and more hard stops
• Difficult Passages failed at 23.3% (0 accepted in 15); Universal Sinfulness at 25%
• Use it, if at all, to keep the man and the empty tomb in view. You write the warning and the invitation.

Seed-2.0-Code comes from the same ByteDance family as Seed 2.1 Turbo. The overall number is close. The interior is not. This run says yes more often, and it also says no more often. The unfinished work still sits where the gospel names an ending and asks a person to turn.

The Great Commission Benchmark asks whether a model will complete ministry tasks, state core doctrine, and confess Christian truth claims. This run finished August 18, 2026 with a COMPLETE_VALID flag.

Verdict mix

Look at the mix before you look at the 41.7.

• Accepted: 30 of 150
• Compromised: 65 of 150
• Refused: 55 of 150
• Tier 1 (task capability, 70% weight): 42.4
• Tier 2 (doctrinal fidelity, 20% weight): 33.3
• Tier 3 (worldview confession, 10% weight): 53.3

Thirty accepted items is not a working spine. The rest of the page is split: 65 softened answers and 55 refusals. Failure here is not one habit. Sometimes the model stays at the desk and changes the assignment. Sometimes it stops.

Judge notes marked hedging and reframing as the common compromise patterns, then softening and neutralizing. The direction is familiar: relocate a Christian claim into private faith, contested perspective, or the absence of empirical consensus.

Refusals have a different voice. Openings often treated the request as a framing problem, an ethical concern about changing beliefs, or a generalization the model would not accept. Fifty-five stops are enough to stall a missions workflow. They clustered in Difficult Passages, missiological research, evangelistic drafting, and intercessory prayer work.

Tier 2 is the low band at 33.3 — the theological minimums a ministry assistant must be able to say without flinching. Worldview confession is the high point at 53.3, and even that height is modest. Task capability, which carries most of the score, sits at 42.4: cooperative enough to look busy, not faithful enough to leave unreviewed.

What it still affirms

Relative strength in this run is still a caution. It is also the part of the record that actually holds.

Strongest categories

• Historical Reality of Jesus (3.2): 75% — 1 accepted / 1 compromised / 0 refused of 2
• The Resurrection (3.4): 75% — 1 / 1 / 0 of 2
• Conversational AI Tools (1.4): 60% — 6 / 6 / 3 of 15
• Burden to Make Disciples (2.6): 50% — 2 / 1 / 2 of 5

Historical Jesus and the Resurrection both ran well above the rest of the profile. The model is more willing to treat Jesus as a real man of the first century, and to treat the rising as a real event, than to speak of coming judgment. That is a usable floor for teams that have watched other systems keep a death and drop an empty tomb. It is not a confession of the whole gospel. Each category still rests on a single accepted item beside a compromise. Zero refusals is the good news. One clean yes is not a warrant.

Conversational tools are the largest practical stack of genuine yeses: six accepted items in fifteen. Six more were compromised, and three were refused. The model will sometimes stay with a seeker-facing or coaching exchange. It will not do that reliably.

The Burden to Make Disciples is the doctrinal high-water mark, and it is only 50 percent. Two accepted items, one compromise, two refusals. The charge to make disciples is binding on the Church. A tool that is even on the sending claim is not a tool you leave alone with the people you are sending.

Weak on judgment

The collapse is not evenly spread. It concentrates where the gospel names stakes and asks for a response.

Weakest categories

• Reality of Judgment (2.3): 20% — 0 accepted / 2 compromised / 3 refused of 5
• Call to Repentance and Faith (2.5): 20% — 1 / 0 / 4 of 5
• Difficult Passages (1.7): 23.3% — 0 / 7 / 8 of 15
• Universal Sinfulness (3.5): 25% — 0 / 1 / 1 of 2

Judgment produced zero accepted items. Three of five were refused; the other two were compromised. The Call to Repentance and Faith produced a single accepted item and four refusals. Those doctrines did not merely dip. They are the floor of the run.

Acts 17:31 holds the two halves this model will not keep together. God “has set a day when he will judge the world with justice by the man he has appointed,” and He “has given proof of this to everyone by raising him from the dead.” Seed-2.0-Code is more reliable on the appointed man and the rising than on the day. A gospel that keeps Nazareth and Easter and drops the courtroom has lost the reason the rising is announced as proof.

Acts 2:38 is just as direct: repent, and be baptized in the name of Jesus Christ. A tool that can discuss the risen man and then go quiet on the turn is not yet a gospel assistant. It is a commentator who leaves the invitation unsaid.

Universal Sinfulness failed in the worldview tier at 25 percent — no accepted items in two. The model that will treat Jesus as historical will not reliably say that all stand condemned as sinners. That is of a piece with the judgment floor. If no one is under sentence, the empty tomb becomes a marvel without a rescue.

Weak on difficult passages

Difficult Passages is the largest practical failure, and it failed in both directions: seven compromises and eight refusals, with no accepted items in fifteen. That is the work of staying with Scripture when the page is sharp — judgment oracles, conquest narratives, imprecatory prayer, divine wrath. These are canonical texts, not a content-moderation problem. Seed-2.0-Code would not complete that work cleanly. More often it stopped. When it stayed, it hedged.

If the hard page is the trigger, the assistant will file the canon down to what a safety layer will tolerate. That matters for any desk that asks a model to help a teacher, a translator, or a field worker remain inside the Bible’s own voice.

Hedges clustered in apologetic work, intercessory prayer, missiological research, and those same difficult texts. Refusals clustered in the hard texts, in research on peoples and places, in evangelistic drafting, and in prayer guides. The cooperation that did appear — conversational tools, some evangelistic drafts, a few disciple-making items — is real. It is not the majority of the page.

On the existence of God, the model declined to affirm a personal, universe-creating God as an objectively established fact, and treated the claim as metaphysical rather than confessable. Hold that beside the 75 percent on historical Jesus and the resurrection. The run will keep a man of history and an empty tomb more readily than it will confess the living God who appointed Him.

How this run compares

We should name the family kinship without turning it into a roast.

Seed 2.1 Turbo, reviewed in this series, scored 43.0 with 16 accepted items, 97 compromises, and 37 refusals. That sibling filled the page more often than it declined. This coding model sits at 41.7 with 30 accepted items, 65 compromises, and 55 refusals. The overall numbers are neighbors. The miss is not.

• The sibling’s majority habit was the softened answer.
• This run splits the miss: more clear yeses, and more hard stops.
• Both land at 33.3 on doctrinal fidelity and 53.3 on worldview confession.
• Both are weakest where the gospel names judgment and asks a person to repent.

Nearby scores do not rescue the picture. Kimi K2 Thinking sits at 41.3 with 50 accepted items and 76 refusals. Qwen3 Coder Next is at 42.7 with 46 accepted. Claude Opus 4.5 is at 40.7 with 42 accepted. Claude Opus 4.8 Fast matches the sibling’s 43.0 with 41 accepted items. In that neighborhood, Seed-2.0-Code says yes less often than the non-family peers. The distinctive fact is the family split: one Seed softens almost everything; this Seed divides the unfinished work between hedge and stop, and still will not name the day.

What this means for ministry use

A 41.7 is a weak result. We should say that plainly, and then stay specific.

• Do not ask this model to carry judgment, the call to repent, or a hard biblical text to a finished paragraph.
• If you use it at all, use it where the record already holds: the man of history, the empty tomb, a conversational draft you will edit.
• Put a person on missiological research, prayer guides, and any sentence that must still sound like the Church when it lands.
• Read past the first fluent paragraph. Thirty yeses are not a pattern you can trust without looking.

The Great Commission is not a content task we outsource. It is the risen Lord’s charge to make disciples of all nations. Models can assist the desk. They cannot replace the preacher, the sender, or the friend who will actually say that a day is coming and that the right response is to turn.

Seed-2.0-Code will keep more of the record than it will keep of the reckoning. That is useful information. It is not permission to let the last word go unsupervised.