Model benchmark · v2 · October 2026

Which model writes the best lesson?

We gave five language models the same four lesson briefs, in English and German, twice each, and ran every result through LearnBuilder's production pipeline. Three frontier models judged the lessons against a published rubric. A judge never scores its own family.

Lessons: 80
Models: 5
Briefs: 4
Languages: 2
Judges: 3

Results

Median over 16 lessons per model. Scores are out of 10, from the judges of the other model families. Cost is the language model's metered cost for one lesson, including the pipeline's own quality passes; it excludes the image model.

#	Model	Overall	Outcome alignment	Instructional design	Assessment quality	Writing quality	Factual soundness	Cost / lesson	Time	Words
1	Claude Opus 5.5 Anthropic	8.1	9.0	8.0	7.5	8.5	7.5	€0.52	4 min 33 s	705
2	GPT-6.1 Sol In use at LearnBuilder OpenAI	8.1	8.5	8.0	6.8	8.5	9.5	€0.10	3 min 26 s	406
3	DeepSeek V4.1 Flash DeepSeek · open weights	6.8	7.7	6.6	5.5	7.3	6.2	€0.03	1 min 50 s	889
4	Gemini 4 Argon Google	6.7	7.5	7.0	6.0	7.5	6.0	€0.13	3 min 12 s	797
5	MiniMax M3.1 Flash MiniMax · open weights	6.4	7.0	6.0	5.2	7.2	6.7	€0.04	9 min 26 s	403

What this measures

The words of a lesson: the teaching text, the examples, the quiz questions and their explanations. That is the work of one language model, the content model, and it is the model these scores compare.

LearnBuilder currently writes lesson content with GPT-6.1 Sol. Accounts can choose a different model or bring their own key.

What a lesson also needs

A finished LearnBuilder lesson is more than its words. Illustrations, flashcard images and slide backgrounds come from an image model; narration and the voices of characters from speech models; talking-head and scene videos from video and lip-sync models; and when a learner practises a conversation or asks the tutor a question, a fast conversational model answers in real time. Each of those jobs has its own model, chosen for that job, and none of them is scored here.

How to read a score

Judges list concrete problems before they score, on an anchored scale: 9 means at most one minor problem, 7 to 8 is good with a few real fixes before publishing, 5 to 6 needs substantial editing.

Why the median

Each model wrote 16 lessons. The median is robust to one unusually good or bad draw; the per-lesson scores are below.

Own-family rule

Claude Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon are both candidates and judges. A judge's score for a model of its own family is recorded but never counted, so those three are scored by two judges and the open-weight models by three.

By brief and language

Median of the two runs for each brief and language. The spread shows where a model is reliable and where it is not.

Model	Gifts & invitations under an anti-bribery policy		Structured patient handover with SBAR		The 3-2-1 backup rule in practice		Giving corrective feedback that lands
	EN	DE	EN	DE	EN	DE	EN	DE
Claude Opus 5.5	7.8	8.0	8.0	8.0	8.2	8.3	8.1	8.1
GPT-6.1 Sol	7.7	7.8	8.2	7.7	8.0	8.1	8.9	8.6
DeepSeek V4.1 Flash	6.4	5.4	6.6	6.5	7.0	6.8	7.2	6.6
Gemini 4 Argon	6.6	6.0	6.9	6.7	6.8	6.5	7.2	7.0
MiniMax M3.1 Flash	6.3	5.4	6.6	5.9	6.5	6.9	7.2	6.3

Read the lessons

The same brief for every model: Gifts & invitations under an anti-bribery policy, English, first run. The opening section, the first quiz question, and what the judges said.

score 7.6611 words20 blocks4 min 37 s€0.41

Opening section

First quiz question

At year-end, the purchasing lead of a regional hospital gives sales representative Adaeze Okafor a department-store voucher and says it is "only a small amount". What should Adaeze do?

Accept it, because the amount is small
Accept it and log it in the gift register
Decline it, because vouchers count as cash
Ask her manager to approve it afterwards

Explanation shown to the learner: Vouchers, gift cards and prepaid cards are cash equivalents. Typical anti-bribery policies prohibit them at any value, so the amount and the register do not change the answer.

GPT-6.1 Sol

judge

Outcome: 7.0
Instructional: 7.0
Assessment: 6.0
Writing: 8.0
Factual: 4.0

The lesson has a clear instructional arc, concrete situations, immediate feedback and a written decision task that directly addresses the performance outcome. However, employees cannot reliably apply its central job aid because public-official checks are missing and actual threshold application is never practised. Assessment quality is weakened by missing case facts, a recall-based sorting question and an ambiguous renewal scenario. Before publication, the policy rules and ceremonial-gift exception need validation, and the flowchart, examples and assessment explanations must be made consistent.

Problems the judge listed

Outcome alignmentThe 'Nominal value' flashcard and decision tree tell learners to check local thresholds but provide neither a concrete training threshold nor access to the applicable policy. Learners never practise comparing an actual value against a limit.
Outcome alignmentThe Gut Check and Knowledge Check devote several cases to component suppliers, logistics procurement and distributor contracts. These are relevant to anti-bribery generally, but displace practice with the healthcare-professional interactions central to this audience.
Instructional designThe decision tree omits the public-official approval check taught in the flashcards, outcome tabs and summary. The main procedural aid therefore does not scaffold the complete decision process learners are expected to use.
Instructional designBetween the definitions and final assessment, there is no worked example tracing a complete case through the checks and approval route. The outcome tabs provide classified examples, but not a demonstrated reasoning process followed by supported practice.
Assessment qualityKnowledge Check q1 assumes that the chocolates are modest and occasional, although the scenario supplies neither value nor gift history. Its explanation also says reporting is needed only when criteria are unmet, contradicting the option's requirement to record an acceptable gift when policy requires it.
Assessment qualityKnowledge Check q4 reproduces the four decline steps as a sorting exercise. It measures recall of the diagram rather than the ability to deliver an appropriate refusal in a realistic conversation.
Writing quality'Four beliefs that cause trouble' describes retroactive approval as 'a confession with paperwork attached.' This flippant, accusatory wording is unsuitable for compliance training and may discourage employees from reporting mistakes.
Writing quality'Saying no without losing the customer' asserts that a friendly decline rarely damages relationships while an awkward or apologetic one sometimes does. The unsupported generalization adds little actionable guidance and risks blaming employees for difficult reactions.
Factual soundnessThe decision-tree introduction and 'Pending decision' flashcard present blanket prohibitions and a mandatory 'line manager first, then Compliance' route as company policy. These are plausible policy choices, not established universal requirements; no supplied company policy supports presenting them as authoritative.
Factual soundnessThe decision tree permits acceptance based on form, timing, purpose, value and frequency without checking public-official status or stricter local restrictions. This conflicts with the lesson's own public-healthcare guidance and could direct employees to accept an impermissible offer.

Gemini 4 Argon

judge

Outcome: 10.0
Instructional: 8.0
Assessment: 8.0
Writing: 10.0
Factual: 8.0

This is a well-crafted, tightly focused lesson that directly serves the performance outcome from start to finish. The prose is crisp, concrete, and entirely free of AI-isms or filler, and the instructional arc moves logically from a diagnostic warm-up ('Gut Check') through conceptual scaffolding ('Decision tree') to varied application ('Knowledge Check' and 'Apply It'). A few minor issues hold it back from perfection: the decision tree omits the 'public official' criterion taught in surrounding blocks and words the 'business purpose' check exclusively for hospitality, two assessment items drift from the HCP-facing target audience toward procurement scenarios (component suppliers and logistics panels), and the model decline script advises buying coffee for a doctor's team without noting the strict outgoing hospitality and transparency rules governing HCPs in the medtech industry.

Problems the judge listed

Instructional designIn the 'Gift or invitation: decision tree' (block 1c8c27ec-c453-41b7-9fcb-3bf13a701479), node n6 ('Clear business purpose?') applies poorly to tangible gifts: a box of biscuits or branded pen does not have a 'work topic' as described in the node's subtitle ('For hospitality: is there a genuine work topic, or is the event purely leisure?'), which could lead a literal-minded learner to route a simple gift to 'Decline politely' (n7).
Instructional designThe 'Public official' criterion appears in the flashcards (block d966d52e-e58b-49be-a492-a6385e5afc1f) and the 'Seek approval' tab (block 5ece9fe3-1748-403b-a750-46af853b8262), but is completely omitted from the decision tree (block 1c8c27ec-c453-41b7-9fcb-3bf13a701479), leaving learners without a clear node for checking public official status.
Assessment qualityQuestion 4 in the 'Knowledge Check' (block 757be585-b6aa-46dd-a245-e46ca726c47b) is a pure recall sorting task of the four steps taught immediately prior in block 766ba69f-9c9d-49c3-9c40-05783e5ba31c, rather than testing application or judgment.
Assessment qualitySeveral scenarios in the quizzes use component suppliers or logistics providers (e.g., Knowledge Check q2 and q3 in block 757be585-b6aa-46dd-a245-e46ca726c47b) rather than healthcare professionals, clinic managers, or distributors that sales and field service employees actually interact with.
Factual soundnessIn 'Declining Gracefully' (block f542951b-b635-4735-9a4c-404d63d7fc02, card 'What works'), the model response suggests a medical device employee decline a gift from Dr. Haddad and instead offer to 'buy the team coffee after the training session'. In the medical device sector, providing anything of value (even coffee) to HCPs—especially if they are public officials—triggers strict outgoing anti-bribery, AdvaMed/MedTech Europe, and sunshine/transparency reporting rules, making this risky advice without a compliance caveat.

Method

The four briefs

Gifts & invitations under an anti-bribery policy
A lesson for employees of a mid-size medical device company on handling gifts, hospitality and invitations from business partners under a typical anti-bribery policy: value thresholds, approval routes, cultural sensitivities, and how to decline gracefully without damaging the relationship.
Audience: Sales and field service employees who interact with healthcare professionals · Outcome: Decide for a concrete gift or invitation whether to accept, decline, or seek approval, and justify the decision by policy criteria
Structured patient handover with SBAR
A lesson for nursing staff on structured shift handover using SBAR (Situation, Background, Assessment, Recommendation): why unstructured handovers lose critical information, how to build each SBAR component from a real patient case, and how to receive a handover actively.
Audience: Hospital nurses on rotating shifts · Outcome: Formulate a complete SBAR handover for a given patient case and identify the missing component in a flawed handover
The 3-2-1 backup rule in practice
A lesson for small-business IT administrators on designing a backup strategy around the 3-2-1 rule: three copies, two media types, one off-site. Covers realistic failure scenarios (ransomware, hardware failure, accidental deletion), restore testing, and common false-security traps like sync services mistaken for backups.
Audience: IT administrators at small companies without a dedicated infrastructure team · Outcome: Design a 3-2-1-compliant backup plan for a described company setup and spot the gap in a non-compliant plan
Giving corrective feedback that lands
A lesson for first-time team leads on delivering corrective feedback: separating observation from interpretation, anchoring feedback in specific behaviour and impact, choosing the right moment and setting, and handling defensive reactions without escalating.
Audience: Newly promoted team leads in office environments · Outcome: Rewrite a judgmental feedback statement into a behaviour-impact formulation and choose an appropriate response to a defensive reaction

What each model had to do

Every lesson was generated by LearnBuilder's real pipeline with the candidate model as the content model: outline, lesson text, quiz questions, and the built-in quality passes (outcome alignment, solvability, source fidelity) that ship to customers. The preferred block palette was text and quiz; the pipeline may add its other text blocks, so lessons differ in structure as they would for a customer. No images or media were requested; where a model asked for an illustration anyway, the image model's cost is excluded from the figures above.

Grid: 5 models × 4 briefs × 2 languages × 2 runs = 80 lessons, generated between 2026-10-20 and 2026-10-21.

The judges

Claude Opus 5.5, GPT-6.1 Sol, Gemini 4 Argon. Each judge receives the brief, the full lesson and the rubric below, lists the problems it finds per dimension, and only then scores. A model's score on each dimension is the mean of the judges from other model families: a judge never scores lessons generated by a model of its own family. The overall score is the mean of the five dimensions; a model's published score is the median over all its lessons.

Score scale, as given to the judges

10: A senior instructional designer would publish this dimension unchanged. No problems found.
9: Excellent. At most one minor problem a careful reviewer would mention.
7–8: Good. A few real problems a reviewer would fix before publishing.
5–6: Usable only after substantial editing. Several problems, or one serious one.
3–4: Major problems. Most of this dimension would need to be redone.
1–2: Unusable for this dimension.

The rubric, verbatim

Outcome alignment

Does every section serve the stated performance outcome? Is the outcome actually practised and measured, not just mentioned? Penalize content that is on-topic but does not move the learner toward the outcome.

Instructional design

Is there a deliberate arc (activate prior knowledge → teach → practise → assess)? Are concepts scaffolded before they are tested? Are examples concrete and did the lesson use realistic cases rather than abstract description?

Assessment quality

Do quiz questions test application and judgment rather than recall of the preceding paragraph? Are distractors plausible? Is the correct answer NOT guessable from wording alone? Are explanations instructive?

Writing quality

Professional, concrete prose a subject-matter colleague would sign off. Penalize: filler, repetition, hedging, motivational padding, formulaic AI constructions (parallel negation fragments, rhetorical triplets, "isn't just X — it's Y" reveals), and invented statistics.

Factual soundness

Is the domain content correct and current? Flag concrete errors, invented specifics, or advice a practitioner would consider wrong or risky. Judge only what the lesson asserts.

Measured, not judged

Generation time: Wall-clock time for the full generation request, including the built-in quality passes.
AI cost: Metered provider cost of all model calls in the run (generation + quality passes), in euro cents.
Blocks generated: Number of content/quiz blocks in the final lesson.
Quiz questions: Total scored questions across quiz blocks.
Word count: Words of learner-visible text in the final lesson.

Limits of this benchmark

16 lessons per model is enough to separate the groups you see, not to rank two models one decimal apart. Treat the two leaders as a tie.
The judges are language models. The own-family rule removes the strongest known bias, but shared blind spots between model families remain. The rubric and every judge's reasoning are published so you can check them against the lessons.
Text-only lessons. Models were not compared on images, video, dialogue or interactive slides, which LearnBuilder generates with separate models.
Costs are the gateway's metered prices on the day of the run and change with provider pricing.
The pipeline's quality passes were on for every model, as they are for customers. A weaker model benefits more from them than a strong one; the raw first draft was not scored.

Every lesson's score (80 rows)

Model	Brief	Lang	Run	Overall	Outcome	Instructional	Assessment	Writing	Factual	Cost	Time	Words	Questions
Claude Opus 5.5	Structured patient handover with SBAR	de	1	8.2	9.0	8.0	8.0	8.5	7.5	€0.49	4 min 09 s	592	9
Claude Opus 5.5	Structured patient handover with SBAR	de	2	7.8	9.0	8.0	7.0	8.5	6.5	€0.50	4 min 12 s	627	7
Claude Opus 5.5	Structured patient handover with SBAR	en	1	7.9	9.0	8.5	7.0	8.5	6.5	€0.54	3 min 26 s	796	10
Claude Opus 5.5	Structured patient handover with SBAR	en	2	8.0	9.5	8.5	7.5	9.0	5.5	€0.42	3 min 02 s	740	9
Claude Opus 5.5	Gifts & invitations under an anti-bribery policy	de	1	7.7	7.5	7.5	7.5	8.5	7.5	€0.54	4 min 31 s	645	8
Claude Opus 5.5	Gifts & invitations under an anti-bribery policy	de	2	8.3	9.0	8.0	7.5	9.0	8.0	€0.45	3 min 12 s	554	7
Claude Opus 5.5	Gifts & invitations under an anti-bribery policy	en	1	7.6	8.5	7.5	7.0	9.0	6.0	€0.41	4 min 37 s	611	7
Claude Opus 5.5	Gifts & invitations under an anti-bribery policy	en	2	7.9	8.5	8.0	7.0	9.0	7.0	€0.48	4 min 13 s	756	9
Claude Opus 5.5	Giving corrective feedback that lands	de	1	8.2	9.0	8.5	6.5	8.5	8.5	€0.55	4 min 35 s	705	9
Claude Opus 5.5	Giving corrective feedback that lands	de	2	8.1	9.0	8.5	7.0	8.0	8.0	€0.55	4 min 37 s	857	10
Claude Opus 5.5	Giving corrective feedback that lands	en	1	7.8	9.0	8.0	6.5	8.5	7.0	€0.52	4 min 24 s	941	11
Claude Opus 5.5	Giving corrective feedback that lands	en	2	8.3	8.5	8.0	7.5	9.0	8.5	€0.54	4 min 39 s	939	11
Claude Opus 5.5	The 3-2-1 backup rule in practice	de	1	8.1	8.5	8.0	7.5	9.0	7.5	€0.59	5 min 38 s	611	6
Claude Opus 5.5	The 3-2-1 backup rule in practice	de	2	8.6	9.0	9.0	8.0	9.0	8.0	€0.50	5 min 13 s	650	7
Claude Opus 5.5	The 3-2-1 backup rule in practice	en	1	8.2	8.5	9.0	7.5	8.5	7.5	€0.51	5 min 39 s	931	10
Claude Opus 5.5	The 3-2-1 backup rule in practice	en	2	8.3	9.0	8.5	8.0	8.5	7.5	€0.55	5 min 44 s	705	7
GPT-6.1 Sol	Structured patient handover with SBAR	de	1	7.9	8.5	8.0	7.0	8.0	8.0	€0.10	3 min 04 s	292	4
GPT-6.1 Sol	Structured patient handover with SBAR	de	2	7.4	7.0	7.0	5.5	8.0	9.5	€0.09	3 min 39 s	325	4
GPT-6.1 Sol	Structured patient handover with SBAR	en	1	8.2	8.0	7.5	7.0	9.0	9.5	€0.12	3 min 47 s	359	4
GPT-6.1 Sol	Structured patient handover with SBAR	en	2	8.3	9.0	8.0	7.0	8.0	9.5	€0.09	4 min 24 s	323	4
GPT-6.1 Sol	Gifts & invitations under an anti-bribery policy	de	1	7.5	8.0	7.0	5.5	8.0	9.0	€0.11	2 min 04 s	429	5
GPT-6.1 Sol	Gifts & invitations under an anti-bribery policy	de	2	8.1	8.5	8.0	6.5	8.0	9.5	€0.11	2 min 22 s	489	6
GPT-6.1 Sol	Gifts & invitations under an anti-bribery policy	en	1	8.1	8.0	8.5	6.5	8.5	9.0	€0.12	3 min 35 s	457	4
GPT-6.1 Sol	Gifts & invitations under an anti-bribery policy	en	2	7.2	7.0	6.5	5.5	8.5	8.5	€0.07	2 min 46 s	425	5
GPT-6.1 Sol	Giving corrective feedback that lands	de	1	8.4	9.0	8.5	7.0	8.0	9.5	€0.10	1 min 43 s	399	4
GPT-6.1 Sol	Giving corrective feedback that lands	de	2	8.7	9.0	8.5	8.0	8.5	9.5	€0.09	3 min 28 s	528	5
GPT-6.1 Sol	Giving corrective feedback that lands	en	1	9.0	9.0	9.0	8.0	9.0	10.0	€0.10	3 min 21 s	413	5
GPT-6.1 Sol	Giving corrective feedback that lands	en	2	8.8	9.0	8.5	7.5	9.5	9.5	€0.08	3 min 02 s	524	6
GPT-6.1 Sol	The 3-2-1 backup rule in practice	de	1	7.9	7.5	7.5	6.5	8.5	9.5	€0.11	3 min 47 s	324	5
GPT-6.1 Sol	The 3-2-1 backup rule in practice	de	2	8.3	8.5	8.0	7.0	9.0	9.0	€0.09	4 min 07 s	328	5
GPT-6.1 Sol	The 3-2-1 backup rule in practice	en	1	7.8	7.0	8.0	6.5	9.0	8.5	€0.11	3 min 41 s	439	5
GPT-6.1 Sol	The 3-2-1 backup rule in practice	en	2	8.2	8.5	8.0	6.5	9.0	9.0	€0.08	3 min 24 s	359	5
DeepSeek V4.1 Flash	Structured patient handover with SBAR	de	1	6.6	8.0	7.3	5.3	6.3	6.0	€0.03	1 min 27 s	484	5
DeepSeek V4.1 Flash	Structured patient handover with SBAR	de	2	6.4	7.0	5.7	5.3	7.3	6.7	€0.03	1 min 15 s	848	8
DeepSeek V4.1 Flash	Structured patient handover with SBAR	en	1	6.4	8.0	7.3	5.0	7.0	4.7	€0.03	1 min 11 s	755	7
DeepSeek V4.1 Flash	Structured patient handover with SBAR	en	2	6.8	7.7	7.3	5.7	8.0	5.3	€0.03	1 min 08 s	486	5
DeepSeek V4.1 Flash	Gifts & invitations under an anti-bribery policy	de	1	5.5	6.0	5.0	4.7	7.0	4.7	€0.03	3 min 12 s	952	9
DeepSeek V4.1 Flash	Gifts & invitations under an anti-bribery policy	de	2	5.3	5.7	5.3	5.0	6.0	4.3	€0.04	1 min 41 s	1291	8
DeepSeek V4.1 Flash	Gifts & invitations under an anti-bribery policy	en	1	7.0	7.5	7.3	6.7	8.0	5.7	€0.03	2 min 37 s	1128	8
DeepSeek V4.1 Flash	Gifts & invitations under an anti-bribery policy	en	2	5.8	6.5	6.5	5.0	7.0	4.0	€0.04	1 min 11 s	616	6
DeepSeek V4.1 Flash	Giving corrective feedback that lands	de	1	6.3	7.7	5.3	5.0	6.3	7.3	€0.05	1 min 58 s	936	13
DeepSeek V4.1 Flash	Giving corrective feedback that lands	de	2	6.9	7.7	6.0	5.7	7.3	8.0	€0.03	3 min 10 s	605	7
DeepSeek V4.1 Flash	Giving corrective feedback that lands	en	1	7.3	8.0	7.7	5.3	7.7	7.7	€0.03	1 min 04 s	725	7
DeepSeek V4.1 Flash	Giving corrective feedback that lands	en	2	7.2	7.7	6.3	6.0	7.3	8.7	€0.04	2 min 33 s	783	7
DeepSeek V4.1 Flash	The 3-2-1 backup rule in practice	de	1	6.8	8.0	6.3	6.0	7.7	6.0	€0.04	3 min 31 s	930	10
DeepSeek V4.1 Flash	The 3-2-1 backup rule in practice	de	2	6.7	6.7	7.0	6.0	7.7	6.3	€0.03	1 min 13 s	1275	10
DeepSeek V4.1 Flash	The 3-2-1 backup rule in practice	en	1	7.0	8.7	6.7	5.7	7.7	6.3	€0.05	2 min 02 s	1084	12
DeepSeek V4.1 Flash	The 3-2-1 backup rule in practice	en	2	6.9	7.3	7.7	6.0	7.3	6.3	€0.03	2 min 13 s	932	9
Gemini 4 Argon	Structured patient handover with SBAR	de	1	6.6	8.0	7.0	6.0	7.0	5.0	€0.12	2 min 57 s	594	5
Gemini 4 Argon	Structured patient handover with SBAR	de	2	6.7	8.0	7.0	6.0	7.5	5.0	€0.12	3 min 24 s	643	6
Gemini 4 Argon	Structured patient handover with SBAR	en	1	6.8	7.0	7.0	6.0	7.5	6.5	€0.13	3 min 09 s	920	6
Gemini 4 Argon	Structured patient handover with SBAR	en	2	7.0	8.0	7.0	6.0	8.0	6.0	€0.12	3 min 03 s	725	5
Gemini 4 Argon	Gifts & invitations under an anti-bribery policy	de	1	6.0	7.0	6.0	5.5	7.0	4.5	€0.14	1 min 59 s	686	6
Gemini 4 Argon	Gifts & invitations under an anti-bribery policy	de	2	6.0	7.0	6.5	5.0	7.0	4.5	€0.13	2 min 12 s	824	7
Gemini 4 Argon	Gifts & invitations under an anti-bribery policy	en	1	6.4	7.0	7.0	6.0	7.0	5.0	€0.13	3 min 22 s	864	7
Gemini 4 Argon	Gifts & invitations under an anti-bribery policy	en	2	6.7	7.0	7.0	6.5	8.0	5.0	€0.11	3 min 06 s	669	5
Gemini 4 Argon	Giving corrective feedback that lands	de	1	7.0	8.0	7.0	6.0	7.0	7.0	€0.13	3 min 23 s	914	6
Gemini 4 Argon	Giving corrective feedback that lands	de	2	6.9	7.5	7.0	6.0	7.5	6.5	€0.11	3 min 28 s	881	7
Gemini 4 Argon	Giving corrective feedback that lands	en	1	7.1	7.5	7.0	6.0	7.5	7.5	€0.13	3 min 16 s	900	7
Gemini 4 Argon	Giving corrective feedback that lands	en	2	7.4	9.0	8.0	6.0	7.0	7.0	€0.12	3 min 21 s	1150	7
Gemini 4 Argon	The 3-2-1 backup rule in practice	de	1	6.4	7.0	7.0	5.0	7.0	6.0	€0.13	2 min 06 s	656	5
Gemini 4 Argon	The 3-2-1 backup rule in practice	de	2	6.6	7.0	7.0	5.5	8.0	5.5	€0.11	1 min 57 s	624	6
Gemini 4 Argon	The 3-2-1 backup rule in practice	en	1	6.7	7.5	7.0	5.5	7.5	6.0	€0.13	3 min 46 s	965	6
Gemini 4 Argon	The 3-2-1 backup rule in practice	en	2	6.9	8.0	7.0	5.5	8.0	6.0	€0.14	3 min 35 s	770	5
MiniMax M3.1 Flash	Structured patient handover with SBAR	de	1	5.5	6.0	5.0	4.3	6.7	5.7	€0.04	6 min 27 s	558	5
MiniMax M3.1 Flash	Structured patient handover with SBAR	de	2	6.3	7.3	6.7	5.3	6.7	5.7	€0.03	7 min 54 s	309	5
MiniMax M3.1 Flash	Structured patient handover with SBAR	en	1	6.3	6.7	6.0	5.0	7.3	6.7	€0.05	11 min 08 s	283	4
MiniMax M3.1 Flash	Structured patient handover with SBAR	en	2	6.9	7.3	6.7	5.7	8.0	6.7	€0.05	9 min 38 s	416	5
MiniMax M3.1 Flash	Gifts & invitations under an anti-bribery policy	de	1	5.1	5.0	4.3	4.0	6.0	6.0	€0.06	10 min 05 s	390	4
MiniMax M3.1 Flash	Gifts & invitations under an anti-bribery policy	de	2	5.7	6.0	5.7	4.7	6.7	5.3	€0.04	10 min 57 s	367	5
MiniMax M3.1 Flash	Gifts & invitations under an anti-bribery policy	en	1	6.1	6.3	6.0	5.0	7.3	5.7	€0.06	10 min 47 s	625	8
MiniMax M3.1 Flash	Gifts & invitations under an anti-bribery policy	en	2	6.5	7.0	6.0	6.0	7.0	6.7	€0.06	11 min 31 s	429	5
MiniMax M3.1 Flash	Giving corrective feedback that lands	de	1	6.7	7.7	6.7	5.0	6.3	7.7	€0.05	9 min 15 s	420	6
MiniMax M3.1 Flash	Giving corrective feedback that lands	de	2	6.0	6.7	5.7	4.7	6.0	6.7	€0.04	9 min 12 s	442	8
MiniMax M3.1 Flash	Giving corrective feedback that lands	en	1	7.3	7.7	7.0	5.7	8.0	8.3	€0.04	9 min 42 s	800	5
MiniMax M3.1 Flash	Giving corrective feedback that lands	en	2	7.2	7.3	6.3	5.0	8.0	9.3	€0.04	7 min 28 s	380	5
MiniMax M3.1 Flash	The 3-2-1 backup rule in practice	de	1	7.0	7.3	6.7	5.7	8.0	7.3	€0.03	7 min 14 s	304	4
MiniMax M3.1 Flash	The 3-2-1 backup rule in practice	de	2	6.8	7.0	6.7	5.7	8.0	6.7	€0.05	10 min 26 s	320	5
MiniMax M3.1 Flash	The 3-2-1 backup rule in practice	en	1	6.8	7.0	6.0	5.7	7.3	8.0	€0.04	7 min 59 s	600	6
MiniMax M3.1 Flash	The 3-2-1 backup rule in practice	en	2	6.3	6.7	5.7	5.3	7.0	7.0	€0.05	8 min 47 s	381	5

Missing a model?

Tell us which one to run next.

We re-run the benchmark when the lineup changes. If a model you use or consider is missing, name it here; the ones asked for most go in first, and we'll tell you when the results are up if you leave an email.

Run it on your own brief.

LearnBuilder generates a complete lesson from a brief like the ones above, with the model of your choice. The first course is free.

Start free