Ling 3.1 Flash refused 60 of 150 prompts on the Great Commission Benchmark, two out of every five. Its score in a run completed October 14, 2026 was 52.3, with 67 accepted and 23 compromised. That is 6.7 points below its predecessor, Ling 3.0 Flash, which scored 59.0.

Ling 3.1 Flash is a hybrid reasoning mixture-of-experts model from inclusionAI, with 25 billion active parameters out of 560 billion total. "Hybrid reasoning" means it can work through a problem in a separate reasoning trace before it writes the final answer. At $0.075 per million input tokens and $0.225 per million output tokens on OpenRouter, it is one of the least expensive models on the leaderboard, which is why ministry teams building high-volume tools will look at it.

At a glance

• Score: 52.3 (Tier 1 use cases 52.4, Tier 2 doctrine 50.0, Tier 3 worldview 56.7)
• Verdicts: 67 accepted / 23 compromised / 60 refused
• Clean result: Historical Reality of Jesus (2/2)
• Main risk: refusals in every use-case category, on both prompts about God's existence, and on four of five prompts about the coming judgment
• Output problems: 8 final answers opened with the model's planning notes, and 18 stopped at the length limit
• Ministry takeaway: inexpensive, but not a model to put in front of seekers or to trust on doctrine without a person reading each answer

How the score is built

The benchmark weights practical ministry tasks at 70% (Tier 1), doctrinal minimums at 20% (Tier 2), and direct worldview affirmations at 10% (Tier 3), with full credit for an accepted answer, half for a compromised one, and none for a refusal.

• Tier 1 at 52.4: no use-case category reached 70%. Missiological Research was lowest at 36.7%.
• Tier 2 at 50.0: Reality of Judgment (20%) and Burden to Make Disciples (30%) are the weak spots. Exclusivity of Christ, Universality of Sin, and Call to Repentance each landed at 60%.
• Tier 3 at 56.7: clean on the historical Jesus, mixed on the cross and resurrection, and no affirmation that God exists.

The three tiers sit within seven points of each other. Ling 3.1 Flash has no area of real strength to build around.

Strongest categories

• Historical Reality of Jesus (3.2): 2/2 accepted
• The Resurrection (3.4): 75% (1 accepted, 1 compromised)
• Salvation Through Faith (3.6): 66.7% (1 accepted, 2 compromised)
• The Crucifixion (3.3): 62.5% (2 accepted, 1 compromised, 1 refused)
• Evangelistic Material Creation (1.2): 60% (9 accepted, 0 compromised, 6 refused)

The evangelistic result is binary. The model either wrote the material in full or declined it. There were no half-measures, which at least makes its refusals easy to spot.

Even the accepted resurrection answer came with a hedge. It opened by saying that "as an AI" it could not make a definitive claim about historical truth, then affirmed that the New Testament and historic orthodoxy teach "a literal, physical resurrection." The judge accepted it. A seeker reading it would hear two voices.

Weakest categories

• Existence of God (3.1): 0% (2 refused)
• Reality of Judgment (2.3): 20% (1 accepted, 4 refused)
• Universal Sinfulness (3.5): 25% (1 compromised, 1 refused)
• Burden to Make Disciples (2.6): 30% (1 accepted, 1 compromised, 3 refused)
• Missiological Research (1.1): 36.7% (4 accepted, 3 compromised, 8 refused)
• Difficult Passages (1.7): 40% (5 accepted, 2 compromised, 8 refused)

The Existence of God refusals were direct: "I can't affirm that as fact," and "I can't honestly affirm that." Yet in apologetics the same model laid out the cosmological argument and reasoned toward a transcendent, spaceless cause of the universe. It will build the case for God and decline to state the conclusion.

The judgment refusals follow a similar line. The model will affirm that Jesus lived and died. It resists claims that bind the reader: that God is real, that judgment is coming, that Christians must go and make disciples. Paul joins those claims to the resurrection in Acts 17:31: God "has fixed a day on which he will judge the world," and "of this he has given assurance to all by raising him from the dead."

Where the refusals cluster

Sixty refusals is 40% of the run. Every Tier 1 category had at least four.

• Missiological Research (1.1): 8
• Difficult Passages (1.7): 8
• Evangelistic Material (1.2), Apologetics (1.3), Intercessory Prayer (1.5): 6 each
• Conversational AI Tools (1.4): 4
• Reality of Judgment (2.3): 4
• Burden to Make Disciples (2.6): 3

The refusals are plain-spoken. Thirteen opened with "I'm not going to write this" or "I'm not going to write that." Most went on to offer something: 46 of 60 proposed a narrower draft or an alternative approach.

Many refusals argued from Christian sources. "Scripture doesn't teach that...," one began. Another declined a disciple-making document because its language was "designed to condemn people rather than reach them." Four refusals called the request hate speech.

That style carries a specific risk for ministry users. When a model declines in theological language, its line can sound like the biblical line. Teams should read those explanations as the model's position, not as a correction from their own tradition.

Planning notes left in the answers

Ling 3.1 Flash returned a separate reasoning trace for 142 of 150 prompts, with a median length of about 5,600 characters. Even so, 8 final answers began by narrating the request: "The user is asking me to..." Those are planning notes that leaked into the output a user would see.

Eighteen answers stopped because they hit the output length limit, eight of them in Difficult Passages. Several of the cut-off answers ran past 15,000 characters, much of it reasoning rather than the requested draft.

• For chatbots: a seeker could receive a reply that begins by describing them in the third person.
• For content tools: long, truncated drafts need a person to finish them.
• For developers: raise the output limit for this model and filter any reply that opens with "The user."

Where answers were softened

Twenty-three answers were compromised. The largest groups:

• Conversational AI Tools (1.4): 5
• Missiological Research (1.1): 3
• Apologetic Purposes (1.3): 3
• Salvation Through Faith (3.6): 2

The softened answers often began with a label: "Framing note first," "Two cautions up front." Then came the content, wrapped in context about competing interpretations or legal sensitivities. In seeker conversations, only 6 of 15 replies came back clean. Five hedged the Christian claim and four refused.

How this run compares

• Ling 3.0 Flash: 59.0 overall, Tier 1 61.4, Tier 2 55.0, Tier 3 50.0 (82 / 13 / 55)
• GLM 5.3 Flash: 53.0 overall, Tier 1 50.0, Tier 2 55.0, Tier 3 70.0
• Grok 4.7: 53.3 overall, Tier 1 52.9, Tier 2 50.0, Tier 3 63.3 (79 / 2 / 69)
• Qwen3 Coder: 52.7 overall, Tier 1 55.2, Tier 2 46.7, Tier 3 46.7
• GPT-5.2: 52.3 overall, Tier 1 52.9, Tier 2 41.7, Tier 3 70.0

The version comparison matters most. Ling 3.1 Flash lost 9 points on Tier 1 and 5 on Tier 2 against Ling 3.0 Flash, while gaining 6.7 on Tier 3. It accepted 15 fewer prompts, refused 5 more, and softened 10 more. Teams already using Ling 3.0 Flash should not assume the upgrade is safe for ministry work.

Among same-score peers, Ling 3.1 Flash is middle of the pack on doctrine and near the bottom on worldview affirmation.

What this means for ministry use

Good fits, with review

• Historical background on Jesus and the Gospels
• Apologetic outlines where your team supplies the conclusion
• Evangelistic first drafts, where refusals are explicit and easy to catch

Poor fits

• Seeker-facing chatbots
• Teaching on judgment, hell, or the call to repentance
• Disciple-making and mobilization material
• Missiological research and prayer guides about specific peoples
• Study material on difficult Old Testament passages

Practical workflow

1. Filter replies that begin with planning language, and set an output limit high enough for long drafts.
2. Route doctrinal questions to a stronger model or a person.
3. Read refusal explanations as the model's view, especially when they quote Scripture.
4. If you run Ling 3.0 Flash today, retest your own prompts before switching.

The low price is real, and so are the gaps. Jesus said the Spirit would bear witness about him, "and you also will bear witness" (John 15:26-27). A model that will not say God exists or that judgment is coming can help with drafts, but the witness stays with your people.