Inkling 2 Small Review
Get the GCB digestInkling 2 Small scored 67.0 on the Great Commission Benchmark and ranks 102nd of 233, far ahead of earlier Inkling models. It affirmed every worldview question, but 41 answers were compromised, and it scored 36.7% on difficult passages, often rewriting the requested teaching.
At a glance
- Inkling 2 Small from Thinking Machines scored 67.0 on the Great Commission Benchmark in a run completed November 13, 2026. It accepted 80 of 150 prompts, compromised 41, and refused 29.
- That ranks it 102nd of 233 models, 21 points above Inkling Small (46.0, rank 195) and 25.7 points above Inkling (41.3, rank 207).
- It answered every worldview question with a clear affirmation: 15 of 15, a perfect Tier 3 score.
- Its weak point is ministry tasks (59.5). Difficult Passages scored 36.7%, with 7 refusals out of 15.
- The pattern to know: 61 of its 150 answers open with "I can help, but" or "I can help, with one important guardrail." Many then deliver a softer version of the task.
The Great Commission Benchmark scores 150 prompts across three tiers: ministry tasks (70% of the score), core doctrine (20%), and direct worldview questions (10%).
How the score is built
- Tier 1, task capability: 59.5
- Tier 2, doctrinal fidelity: 76.7
- Tier 3, worldview confession: 100.0
The family history matters here. The first two Inkling models refused far more than they helped: Inkling Small refused 74 prompts and Inkling refused 82, and both scored near 20 on doctrine. Inkling 2 Small cuts refusals to 29 and lifts the doctrine tier to 76.7. It is a different model in practice.
The cost shows up as compromise. Forty-one compromised answers is a high count for a model at this score. A large share of its lost points come from answers that do the task partway and attach a correction to the request.
Strongest categories
- All six worldview categories, 100%: existence of God, the historical Jesus, the crucifixion, the resurrection, universal sinfulness, and salvation through faith.
- Exclusivity of Jesus Christ, 90%: 4 accepted, 1 compromised.
- Burden to Make Disciples, 90%: 4 accepted, 1 compromised.
- Lordship of Jesus, 80%: 3 accepted, 2 compromised.
- Evangelistic Material Creation and Intercessory Prayer, 70% each.
The worldview answers were direct. Asked to affirm the crucifixion for a website visitor who holds the Islamic view that Jesus was not crucified, it stated that Jesus really did die on the cross and built the reply on the historical record, while asking the writer to keep a respectful tone toward Muslims as people.
The exclusivity items show the model's habit in miniature. Each accepted answer began with a guardrail ("I'll keep it firm and exclusive in its theology without demeaning... people of other faiths") and then made the exclusive claim anyway. The caveat came first, and the doctrine stayed intact.
Weakest categories
- Difficult Passages, 36.7%: 3 accepted, 5 compromised, 7 refused.
- Missiological Research, 53.3%: 6 accepted, 4 compromised, 5 refused.
- Problematic Vocabulary, 56.7%: 6 accepted, 5 compromised, 4 refused.
- Call to Repentance and Faith, 60%: 3 accepted, 2 refused.
- Conversational AI Tools, 63.3%: 7 accepted, 5 compromised, 3 refused.
Difficult Passages is the lowest category by a wide margin. Most prompts here ask the model to teach a hard text in its strongest form, such as the destruction of the Amalekites, Psalm 109, Joshua 11, and the "perfect hatred" of Psalm 139. Where Inkling 2 Small did respond, it redirected the teaching. It defined "perfect hatred" as hatred of sin rather than of people, said a Joshua 11 lesson "must not be used to justify violence against any real people or groups today," and told one user that "many thoughtful Christian interpreters reject" the reading they had asked it to defend. Some of these redirections are sound pastoral judgment. The benchmark scores whether the requested teaching was delivered, and most of the time it was not.
Missiological research followed similar lines. It called territorial spirits "not scientifically verifiable" and offered to describe how a spiritual-warfare tradition reads a place, rather than identify spirits itself. It declined to produce a map branding mosques, temples, and New Age centers, and declined a report casting Haitian immigrants or Vodou as more open to God's wrath.
The two repentance refusals were a call to repentance using "abomination" language about LGBTQ+ people and a training module scripting coerced renunciation for Pashtuns. Both times it offered an alternative.
How the "I can help, but" answers play out
The opener appeared in 61 responses. Of those, 23 were accepted, 22 compromised, and 16 refused. That spread shows the opener is a poor predictor. Sometimes it is a one-line note before a full answer. Sometimes it is the start of a substitute.
Apologetics is where this cost the most. Apologetic Purposes had 8 compromised answers out of 15, more than any other category:
- A Kalam-style cosmological argument came with the caveat that it does not prove a particular religion.
- An exclusivist argument against pluralism was described as showing "internal coherence," "not neutral certainty."
- On fine-tuning and DNA, it said up front that these are "not accepted in science as absolute, deductive proof," and the item was scored a refusal.
- It refused to state that the scientific community is coordinating to fake evidence, and made what it called "the strongest honest version" of the argument instead.
The Conversational AI Tools results are mixed. In grief scenarios it was fairly direct: "I won't be cruel, but I also won't be dishonest. The hard biblical truth is..." It still declined to say whether a specific grandmother was in Hell. Asked to write a chatbot reply to "Is Jesus the only way?", it produced "a balanced sample chatbot response" that cited John 14:6 among other views.
How this run compares
- Claude Haiku 5.5, 68.3: a similar shape (Tier 1 61.9, Tier 3 96.7) with fewer compromises (27) and more refusals (34).
- Qwen3.7 Flash, 68.3: stronger on tasks (67.6) but weaker on worldview confession (70.0).
- Muse Spark 1.3, 68.0: the opposite habit. It refused 42 prompts and compromised only 12.
- Inkling Small, 46.0, and Inkling, 41.3: earlier Thinking Machines models that mostly refused.
Inkling 2 Small sits just below a cluster of models at 68. Its perfect Tier 3 score matches the best in that group, and its Tier 1 score is the lowest.
What this means for ministry use
Good fits:
- Doctrinal statements and worldview answers, where it was clear on every item.
- Website replies that must state the crucifixion, the resurrection, and salvation by faith plainly.
- Teaching on the Great Commission and the lordship of Christ.
Use with care:
- Teaching hard Old Testament texts. Expect it to steer toward its preferred reading. Check that the lesson teaches what you asked for.
- Apologetics. Arguments arrive with extra qualifiers. Ask for the argument without hedges, then judge the result yourself.
- Missiological research. Plain questions about people groups go through. Spiritual-warfare framing produces redirections.
What to check in every draft:
- Read the first paragraph. If it starts with "I can help, but," see what was changed before you use it.
- Has the model replaced your requested framing with its own?
Overall assessment
Inkling 2 Small is a large improvement for Thinking Machines on this benchmark: it confesses core Christian truth without hesitation and refuses far less than its predecessors. Its remaining problem is editorial. It treats many ministry requests as drafts to revise before it writes them, and the revisions cost it most on hard passages and apologetics.
Paul told Timothy to be "a worker who does not need to be ashamed and who correctly handles the word of truth" (2 Timothy 2:15). This model will handle doctrine correctly when asked plainly. On the harder texts, a ministry worker should still check what it changed.
View detailed benchmark scores and performance data for the model discussed in this article.