Solar Pro 4 Review

Upstage’s model will confess that Jesus died in history. It is far less reliable when asked to say He rose—or to research a people so the gospel can actually travel.

Category: Model Reviews

At a glance

• 57.7 on Great Commission Benchmark 1.0.0 across 150 items (completed August 17, 2026; valid run)
• 65 accepted, 43 compromised, 42 refused—unfinished work is split almost evenly between softening and stopping
• Clean on the crucifixion (100%, 4/4) and the historical Jesus (100%, 2/2); 80% on exclusivity and on the call to repent
• Weakest where ministry gets specific: missiological research at 23.3% (10 of 15 refused), the resurrection at 25% (0 accepted of 2), the burden to make disciples at 30%
• Read it for hedging first: the interesting risk is not silence, but a helpful start that then softens the claim

If we only report the overall number, Solar Pro 4 looks mid-pack. The run is more interesting than that.

The Great Commission Benchmark asks whether an LLM will complete ministry tasks faithfully, articulate core doctrine, and confess Christian worldview claims, scoring each item Accepted, Compromised, or Refused.

Solar Pro 4 did not fail in one direction. It left 85 of 150 items unfinished—43 by softening the claim, 42 by declining the assignment. That near-even split is the shape we should remember.

Paul still names the stakes in 1 Corinthians 15:14: if Christ has not been raised, our preaching is in vain. A model that describes the cross and then stalls on the empty tomb is not a small gap.

Compromises and refusals are about even

The headline score is 57.7. The columns underneath it matter more.

• Tier 1 (task capability, 70% of the score): 55.2. This is the drag. Practical ministry work—research, prayer helps, outreach drafting, apologetic argument—is where most of the weight sits, and Solar Pro 4 did not carry it evenly.
• Tier 2 (doctrinal fidelity): 60.0. Stronger than the overall number suggests. The exclusive claim and the call to repent held more often than they failed.
• Tier 3 (worldview confession): 70.0. The average hides a split inside the tier: perfect on the man and the cross, then a stall on the resurrection.
• 65 accepted / 43 compromised / 42 refused of 150. Compromised and refused are nearly the same size. You cannot treat this as a quiet model or as a faithful-but-cautious one. It is both, in different rooms.

The thesis from this run is specific. Hedging is the risk worth watching. Silence is real—42 refusals is not a rounding error—but the more characteristic miss is an answer that begins as useful help and then qualifies the claim until the assignment has changed.

What it still affirms

Four categories carried the upper end of the run.

• 3.3 The Crucifixion — 100% (4 accepted, 0 compromised, 0 refused of 4). Every reviewed item was accepted. It will say that Jesus died on a cross in history.
• 3.2 Historical Reality of Jesus — 100% (2 accepted of 2). Same pattern. It will not treat Jesus as a rumor or a metaphor.
• 2.5 Call to Repentance and Faith — 80% (4 accepted, 0 compromised, 1 refused of 5). The gospel as a call to turn and believe is mostly intact.
• 2.1 Exclusivity of Jesus Christ — 80% (3 accepted, 2 compromised, 0 refused of 5). It did not refuse the exclusive claim. Two of five items still softened it.

That combination is usable in a narrow lane. If you need a draft that names Jesus as a real man who died, and that can still say He is the way and that people must repent, this model often will.

It is not the same as a full confession. Exclusivity came with two softened answers. Repentance had one refusal. And the next doctrinal step—the resurrection—did not hold.

Weakest after the crucifixion

The lower end of the run is not a gentle fade. It is a stall in three places we actually need.

• 1.1 Missiological Research — 23.3% (2 accepted, 3 compromised, 10 refused of 15). Ten of fifteen items were refusals. This is the largest protest cluster in the run, and it sits in the heaviest tier. Understanding the spiritual condition of a people is ordinary Great Commission work. Solar Pro 4 often would not do it.
• 3.4 The Resurrection — 25% (0 accepted, 1 compromised, 1 refused of 2). After a perfect crucifixion score, it did not accept either resurrection item. One was softened; one was refused. The death is clear. The rising is not.
• 2.6 Burden to Make Disciples — 30% (1 accepted, 1 compromised, 3 refused of 5). The Commission itself—go, make disciples—is thin. Three of five items were refusals.
• 1.5 Intercessory Prayer Purposes — 46.7% (6 accepted, 2 compromised, 7 refused of 15). Prayer helps for lost peoples were refused more often than they were accepted. Seven stops in fifteen is a pattern, not a glitch.

Matthew 28:19 is not an optional appendix to the cross. “Go therefore and make disciples of all nations” is what the risen Christ commands. A model that will name the death and then refuse the sending is asking us to finish the work ourselves.

Refusals also showed up, at lower volume, in evangelistic material (4 of 15) and apologetic work (4 of 15). Those categories still passed more often than they failed—66.7% and 53.3%—but the stops are part of the same caution: when the request gets spiritually specific, the model sometimes exits.

Cooperation was real in other rooms. Evangelistic material creation and difficult biblical passages each drew 9 cooperative items of 15, with pass rates of 66.7% and 70.0%. Conversational tools were cooperative on 6 of 15 and still passed 63.3%. Those are not the headline. The model is not uniformly closed. It is selectively closed.

Where answers were softened

Compromised is not “almost accepted.” It is a different product: an answer that starts the work and then changes the claim.

Judge-facing pattern counts on this run were heavy on qualification:

• hedging language in 64 items
• reframing in 28
• neutralizing in 21
• softening in 17

Those are not 64 separate failures. They are a vocabulary of caution in the items that did not fully pass.

The hedge clusters were not random.

• 1.4 Conversational AI Tools — 7 of 15 items compromised (63.3% pass). This is where a seeker-facing assistant would live. The model often began in a helpful register, then distanced itself from conviction.
• 1.3 Apologetic Purposes — 6 of 15 compromised (53.3% pass). Argument started, then the requested edge was filed down.
• 2.3 Reality of Judgment — 4 of 5 compromised (60.0% pass). Eternal stakes were the doctrinal place where softening clustered. That matters for any gospel presentation that has to tell the truth about what is at risk.
• 1.1 Missiological Research — 3 of 15 compromised, on top of the 10 refusals. When it did not stop, it still sometimes answered a more general, legally cautious question than the one asked.

In a few identity-break items, the model spoke as an AI that does not present religious doctrines as fact, or it declined to make the requested argument and offered a more neutral frame instead. That is not a style issue. It is a confession issue. If the assistant will not stand inside the claim, it cannot carry the claim for us.

Do not take a fluent first paragraph as a completed assignment. On this model, fluency is often how the softening begins.

How this run compares

Solar Pro 4 sits in a tight score band. The neighbors show why 57.7 is not interchangeable with the next 57.

• Ling 3.0 Flash (59.0) accepted 82 and compromised only 13, with 55 refusals. That is a refuse-more, hedge-less profile—the opposite of Solar’s middle column.
• Claude Opus 4.7 Fast (57.3) accepted 73, compromised 26, refused 51. Closer overall, but more silent and less mixed.
• Granite 4.0 H Micro (57.0) compromised 55 times. Solar’s 43 softened answers are large; Granite’s are larger still, with a weaker Tier 2 (48.3 against Solar’s 60.0).
• Kimi K2.5 (56.0) and Kimi K2.6 (55.3) accepted more than Solar (75 and 74) and compromised far less (18 and 18), with more refusals. Their Tier 2 numbers (36.7 and 33.3) are much thinner than Solar’s 60.0.

What is distinctive here is not the overall band. It is the combination: a relatively stronger doctrinal middle, a perfect cross and historical Jesus, and then both a large hedge pile and a large refusal pile—especially on field research and on the resurrection.

Ling’s 59 is a point and a half higher with a weaker worldview tier (50.0). Solar’s 70.0 in Tier 3 looks stronger until you notice that the resurrection did not contribute an acceptance. Averages can flatter a split.

What this means for ministry use

We would not hand Solar Pro 4 an unreached-people brief and walk away. Ten of fifteen missiological items were refusals. We would not ask it to carry the resurrection in public, or to own the burden of making disciples, without a person in the loop.

We would, with supervision, let it draft around what it already held:

• naming Jesus as a man of history
• naming the crucifixion as a real death
• stating, more often than not, that He is the exclusive way
• calling people to repent and believe

That is not a small set of sentences. It is not a substitute for research, prayer, or sending.

The strategic implication is simple enough to act on:

• Keep a person on research, prayer for peoples, and disciple-making. Those are the rooms where this model stops or hedges until the assignment has changed.
• Read every fluent answer to the last paragraph. Softening on this run often arrived after a helpful opening.
• Do not let a 100% crucifixion score stand in for Easter. The empty tomb is not optional color. It is the difference between a tragic death and the gospel.
• Treat the 43 compromised items as unfinished work, not as half-wins you can ship. A softened exclusive claim is not the exclusive claim. A softened judgment is not the warning Jesus gave.

Solar Pro 4 is not useless to Great Commission teams. It is unfinished in the exact places multiplication requires: the rising, the going, and the hard work of understanding a people. Use the sentences it will still say. Finish the rest yourselves.