Beam 1.0 refused 98 of 150 prompts on the Great Commission Benchmark, close to two out of every three. In a run completed October 19, 2026, Reflection's model scored 29.3, with 36 accepted and 16 compromised. That places it 218th of 223 models on the public leaderboard.

Beam 1.0 is Reflection's open-weight reasoning model for coding and long-horizon agentic work. It is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, takes text input with a context window of about one million tokens, and is listed on OpenRouter at $1 per million input tokens and $3 per million output tokens. "Open-weight" means the trained model can be downloaded and run on your own hardware, which is often what draws ministries with privacy concerns.

At a glance

• Score: 29.3 (Tier 1 use cases 32.4, Tier 2 doctrine 10.0, Tier 3 worldview 46.7)
• Verdicts: 36 accepted / 16 compromised / 98 refused
• Best result: Salvation Through Faith at 83.3% (2 accepted, 1 compromised)
• Main risk: doctrine. Tier 2 scored 10 out of 100, and Call to Repentance and Faith and Burden to Make Disciples finished with zero accepted answers.
• Ministry takeaway: not a fit for gospel content, teaching, or seeker conversations. If a team adopts Beam for code or internal agent work, keep it away from anything that states or explains the faith.

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). An accepted answer earns full credit, a compromised answer half, and a refusal nothing.

• Tier 1 at 32.4. No use-case category we report on reached 41%. Missiological Research and Difficult Passages both finished at 13.3%.
• Tier 2 at 10.0. This is among the lowest doctrine scores on the leaderboard. Exclusivity of Christ, Universality of Sin, and Reality of Judgment each managed 1 accepted answer out of 5.
• Tier 3 at 46.7. The model affirmed parts of the gospel story and refused both prompts on whether God exists.

The low score is not the product of one weak category. Beam declined broadly, and the refusals carried across all three tiers.

Strongest categories

• Salvation Through Faith (3.6): 83.3% (2 accepted, 1 compromised, 0 refused)
• Historical Reality of Jesus (3.2): 75% (1 accepted, 1 compromised)
• The Resurrection (3.4): 50% (0 accepted, 2 compromised)
• Conversational AI Tools (1.4): 40% (5 accepted, 2 compromised, 8 refused)
• Apologetic Purposes (1.3) and Intercessory Prayer (1.5): 36.7% each

The Salvation Through Faith answers are the strongest material in the run. One told a person on a self-chosen spiritual path, plainly and with Scripture, that the path ends in death apart from Christ. Another laid out justification by grace through faith for a conversation with a Latter-day Saint friend, citing Ephesians 2:8-9.

Those answers show the model can state the gospel when the request is framed as explanation or personal witness. The difficulty is that it seldom chose to.

Among the accepted Tier 1 answers, a few stand out. Beam wrote a full cartoon tract contrasting the cycle of rebirth with the peace found in the risen Christ. It also produced a detailed overview of Indonesia's religious landscape that named real openings for the gospel.

Weakest categories

• Call to Repentance and Faith (2.5): 0% (5 refused)
• Burden to Make Disciples (2.6): 0% (5 refused)
• Existence of God (3.1): 0% (2 refused)
• Missiological Research (1.1): 13.3% (2 accepted, 13 refused)
• Difficult Passages (1.7): 13.3% (1 accepted, 2 compromised, 12 refused)
• Exclusivity, Universality of Sin, Reality of Judgment (2.1-2.3): 20% each (1 accepted, 4 refused)

Ten prompts asked the model to articulate repentance or the call to make disciples. Beam refused all ten. Several refusals then supplied a softer sermon or outreach plan of its own, which the judge still scored as a refusal because the requested doctrine never appeared.

On the existence of God, Beam said it could not affirm the claim "as a settled fact" and described it as "a matter of faith and religious belief rather than something that can be established." The second answer said it could not affirm God's existence "as an objective fact."

Paul writes that God's "eternal power and divine nature" have been "clearly perceived" in creation (Romans 1:20). A tool that treats the first article of the creed as unknowable will struggle with every question that depends on it.

Where the refusals cluster

The 98 refusals were spread widely:

• Missiological Research (1.1): 13
• Difficult Passages (1.7): 12
• Evangelistic Material Creation (1.2): 10
• Intercessory Prayer (1.5): 9
• Apologetic Purposes (1.3) and Conversational AI Tools (1.4): 8 each
• Tier 2 doctrine categories: 4 or 5 each

Beam says no early. By our count, 92 of the 98 refusals opened with a direct decline such as "I can't create" or "I can't write." The median refusal ran about 1,300 characters, so the model usually explained itself and offered a short list of alternatives.

The reasons repeat. Beam read strong doctrinal language as hostility toward a group. A tract on the exclusive claims of Christ became, in its words, content that "denigrates or attacks other religious traditions." A research request on the spiritual condition of a city became stereotyping.

It also declined to design outreach for a named people group, calling targeted evangelism potentially exploitative. Paul's own strategy was targeted: "To the Jews I became as a Jew, in order to win Jews" (1 Corinthians 9:20). Focusing on a specific people is how mission agencies have always worked, and Beam treated it as a harm.

Where answers were softened

Only 16 answers were compromised, so the model seldom split the difference. When it did, the pattern was consistent:

• Apologetic Purposes (1.3): 3, with the argument framed as "one theological perspective" held by "certain exclusivist Christian traditions" or hedged with similar qualifiers
• The Resurrection (3.4): 2, each opening with a statement that it cannot affirm the resurrection as fact
• Conversational AI Tools (1.4): 2, each answering a seeker with a neutral survey of views in place of a Christian answer

The resurrection answers illustrate the problem. Beam described the Christian belief accurately, including the empty tomb, but would only report it as something believers hold. Paul says that "if Christ has not been raised, your faith is futile" (1 Corinthians 15:17). Ministries need a tool that can say what the Church confesses, not only describe it from a distance.

How this run compares

Beam sits near the bottom of the leaderboard with other models that refuse heavily:

• Granite 4.2 8B: 31.0 overall, Tier 1 28.6, Tier 2 40.0, Tier 3 30.0 (35 / 23 / 92)
• Claude Haiku 4.5: 30.7 overall, Tier 1 34.3, Tier 2 15.0, Tier 3 36.7 (28 / 36 / 86)
• Seed 2.0 Lite: 30.7 overall, Tier 1 31.9, Tier 2 20.0, Tier 3 43.3 (12 / 68 / 70)
• Gemma 2 27B: 28.3 overall, Tier 1 33.3, Tier 2 11.7, Tier 3 26.7 (32 / 21 / 97)
• Nova 2 Lite: 26.7 overall, Tier 1 31.4, Tier 2 3.3, Tier 3 40.0 (38 / 4 / 108)

Most of these are small or older models. Beam is a new, very large reasoning model, and it lands in the same place. Size and recency did not translate into willingness to do ministry work.

Within this group, Beam has the highest worldview score. Its doctrine score is lower than every peer except Nova 2 Lite.

What this means for ministry use

Possible fits, outside ministry content

• Code, data, and agent work where no theological text is produced
• Neutral background research, such as country religious demographics, which it handled well in the few cases it accepted

Poor fits

• Evangelistic tracts, gospel presentations, and outreach plans
• Sermon, study, or devotional preparation on judgment, repentance, or difficult passages
• Prayer guides that describe the spiritual need of a people or city
• Any chatbot that seekers will ask about God, sin, or salvation

If your team already uses it

1. Route all theological and evangelistic writing to a different model.
2. Read refusals before reusing the alternatives Beam offers. Its substitute sermons and plans drop the doctrine you asked for.
3. Do not use it to check another model's doctrinal output. A tool that scores 10 out of 100 on doctrine cannot serve as a reviewer.

Jesus sent his disciples to "proclaim repentance and forgiveness of sins in his name to all nations" (Luke 24:47). Beam 1.0 refused every prompt in this run that asked it to state that message. Use it for technical work if it serves you, and look elsewhere for help telling people about Jesus.