Introducing Gen-1 Slides: AI Model Built for Work
For two years, Genspark has built on the world’s best AI models. Today, we’re introducing our own: Gen-1 Slides, the first in a family of post-trained models built for the knowledge worker. Gen-1 Slides turns a single request into a boardroom-ready presentation, with frontier-level performance against benchmarks and impressive cost-efficiency.
Starting today, it will power Standard mode in Genspark AI Slides, delivering significantly higher quality decks while lowering token costs for users. Trained with Fireworks AI, Gen-1 Slides ranks first on average across internal and external benchmarks, ahead of Claude Opus 5 at roughly 1/17 the input-token price. It also effectively tied with Claude Fable 5.1 for the highest overall quality in our comparison.
With the current wave of frontier models, the race has played out in coding benchmarks, math competitions, and reasoning tests. Frontier labs optimize for general capability. Open-model releases chase the same leaderboards while closing the price gap. The result is that far less attention has gone to the practical work most knowledge workers actually do all day: decks, spreadsheets, documents.
At 120,000 decks on a peak day, we see exactly where general-purpose models fall short and what people actually want from the finished presentation. We had the scale, the feedback, and the reason to build something better. So we trained our own model.
For Genspark, this is an important milestone. We'll continue to orchestrate the best models for every task, while now adding the ability to train our own where our scale, product experience, and user feedback give us an advantage. Slides are the first example. We are already applying the same approach to spreadsheets, documents, and research.
This new capability gives us another way to improve the quality of work we deliver to users and move faster on the work we understand best. Gen-1 Slides is live today: if you use Genspark AI Slides, it is already the default in Standard mode and you can give it a try.
Below, we'll explain why we did it, exactly how we measure it, and where it still falls short.
Highlights
For users: a major upgrade to decks at the same price. Gen-1 Slides is now the default in Standard mode, trained specifically to create better presentations. On 200 de-identified slide tasks from consumer users who have AI Data Retention enabled, it scores above Claude Opus 5 on our internal grader and ranks first on visual design, with no change in price for users. Opus-tier quality is now the default, not the premium tier.
For the industry: an application company can now build a frontier-tier model that runs at 1/17th the input-token cost. Gen-1 Slides starts with the open-weight MiniMax M3, trained jointly with Fireworks AI, inside the same system Genspark uses to produce real presentations. We disclose our base model openly — in our view, provenance transparency is part of the product, not something to avoid. The result matches Claude Opus 5 on deck quality while costing $0.30 per million input tokens, compared with $5 for Opus 5. Measured per finished deck, a team producing 1,000 decks a month goes from about $4,200 to about $440, and the budget that used to serve one user at frontier quality now serves roughly ten. As the Gen-1 model family extends to other deliverables, frontier quality becomes the floor rather than the upsell.
For the technical reader: strong results on benchmarks you can run yourself. We tested Gen-1 Slides on three datasets, each paired with its own grader: our internal real-task set with our internal grader, UltraPresent with PPTEval, and UniPPTBench with UniPPTEval. It ranked first in eight of the nine evaluations and in the top two in all nine, with an average rank of 1.11 compared with 2.44 for Kimi K3 and 2.56 for Claude Opus 5.
Available now as the default in AI Slides Standard mode as well as from the gsk CLI. OpenRouter listing in progress. Early API access: [email protected].
Figure 1. Aggregate internal-grader score vs. input-token list price: eleven models, one harness.
Vertical axis: aggregate internal grader score on the real tasks completed by all eleven models. Horizontal axis: list input-token price, log scale. Gen-1 Slides sits alone in the upper-left corner: no model is both cheaper and better. It shares a price with the off-the-shelf M3 it was trained from; the arrow marks what post-training added at that price (0.576 to 0.827). We compare primarily against Opus 5 because we have served it at scale on this task; Fable 5 / 5.1 are shown for reference only. Full axis and sample notes in Appendix D.
Why slides need their own model specifically
Our users were paying frontier-model prices for decks that still came back with broken layouts, fabricated citations, and claims the source material did not support. The strongest general-purpose models in the world are remarkable at many things, but producing a finished, deliverable slide deck has not been one of their priorities. That is understandable: frontier labs are optimizing for general capability, and open-model releases are measured against the same benchmarks. A model built specifically for this task was unlikely to come from anywhere else, so we built it ourselves.
Slide generation may look like a writing task, but it is really an engineering task. The model has to plan the narrative, stay faithful to the source materials, write front-end code for every page, render it, inspect the result, fix problems, and repeat, for dozens of turns. General-purpose models are strong at some of these steps and weak at others, and the weaknesses are stable enough to form signatures in our evaluation:
Claude Opus 5 is the strongest at task completion and near the top on content, but it runs the fewest self-check loops, so the finished deck carries visible flaws. For a user, that means paying a high price and still fixing pages.
GPT-5.6 Sol produces the sparsest decks, with little on each page. It ranks last on visual design among the general-purpose models, and last of all five on UniPPTEval's density, layout, consistency, and hierarchy dimensions across every dataset.
Kimi K3 is the most balanced of the three, but it ranks third on the internal grader's visual design across all three datasets.
Slide-making sits closer to composing and laying out a canvas than to solving a math problem. It is hard, but it is tractable if you train for it directly from what users actually experience. Running 120,000 decks a day gives us three things a general-purpose lab does not have on this task: we know what people actually ask a deck tool to do, and where the general models fail them. We know what a good deck looks like to the person who requested it, because we see, in aggregate, which decks get a thumbs-up, get downloaded, or get sent back for edits. We can compute the reward on the finished deliverable rather than on a transcript of text.
How we evaluate
The quality of a deck depends heavily on how you measure it. We run slide generation at scale every day, so we built our own grader, and we would rather explain how it works and how we validated it than ask you to trust it. Three signals, three datasets, five models.
Three signals:: automated graders, human annotation, production feedback.
Internal grader: our scoring pipeline. Its judge, an agent driven by Claude Fable 5, scores the rendered deck on four dimensions: task completion, content quality, visual design, and process quality. Penalties (layout defects, fabrication, misrepresentation, ignored instructions) and a clean-deck bonus are applied on top. This grader also supplied the RL reward for Gen-1 Slides, so its Table 1 scores are not independent of training; the two public graders below are. Throughout this post, grader means a scoring pipeline, judge means the multimodal LLM inside a grader, and reward means the scalar training signal computed from the grader's output.
Public graders: we use the two graders published by the benchmark authors. PPTEval, from the PPTAgent project by the UltraPresent authors, scores content, design and coherence. UniPPTEval, defined in the same paper as UniPPTBench, scores ten base dimensions plus visual integrity. Both run on all three datasets, not just their own, and neither was ever used as a reward during training. We run PPTEval on a Claude Fable 5 judge and UniPPTEval on the Gemini judge its paper specifies.
People: Product managers and designers annotated 4,063 pages across 77 slide-generation tasks. Separately, 1.57 million production tasks gave us star ratings, thumbs, and download signals.
Three datasets: We used 200 de-identified tasks from consumer users who have AI Data Retention enabled, plus two public benchmark sets: UltraPresent, published by the PPTAgent authors, of which we use the 128-task validation split; and UniPPTBench, 126 tasks. Throughout this post, UltraPresent and UniPPTBench name the datasets, and PPTEval and UniPPTEval name their official graders.
Five models ran every task in the same agent harness, with the same tools, rendering loop, prompts, and production thinking-effort settings: Gen-1 Slides, Claude Opus 5, Kimi K3, GPT-5.6 Sol, and the off-the-shelf MiniMax M3 (pre-RL) checkpoint as the no-post-training baseline. Figure 1 includes more models for reference, but these are the models we have run at meaningful scale in production, so they are the comparison we trust.
Quality
Same task, three models
The same quarterly business review task, rendered by three models. Each image shows the corresponding overview slide from the generated deck.
On our 200 real tasks, Gen-1 Slides has the highest aggregate score of the five models and the highest visual-design score. Claude Opus 5 leads on task completion and content quality. The two sit in the same tier overall, and the gap between them is smaller than the gap to everything else.
Table 1. Internal grader, 200 real production tasks (n=200).
| Model | Aggregate | Task completion | Content quality | Visual design | Process quality | USD / deck |
|---|---|---|---|---|---|---|
| Gen-1 Slides | 0.821 | 0.780 | 0.737 | 0.756 | 0.719 | 0.44 |
| Claude Opus 5 | 0.810 | 0.849 | 0.782 | 0.688 | 0.695 | 4.16 |
| Kimi K3 | 0.723 | 0.840 | 0.784 | 0.557 | 0.748 | 2.00 |
| GPT-5.6 Sol | 0.668 | 0.820 | 0.743 | 0.479 | 0.748 | 2.01 |
| MiniMax M3 (off-the-shelf) | 0.563 | 0.741 | 0.668 | 0.494 | 0.608 | 0.34 |
Cost is measured LLM spend per finished deck in the same harness. Bold marks the best in each column. Aggregate is not the average of the four columns; see Appendix C.
Does it hold on graders we didn't build?
The internal grader reflects Genspark's understanding of this task, so the fair question is whether these gains still hold up on public benchmarks. Figure 2 compares the same five models on three datasets under three independent grader families. Bars start at zero; some gaps look small because the strongest models sit near the top of each scale, not because the axis is compressed.
Figure 2. Five models, three datasets, three grader families.
Each grader on its own scale (internal 0–1, PPTEval 1–5, UniPPTEval 0–10). 200 from internal real tasks, 128 from UltraPresent's validation split, 126 from UniPPTBench. Production reasoning settings: adaptive thinking, high effort, skipped where unsupported.
The three graders mostly agree. Gen-1 Slides ranks first in eight of the nine grader × dataset combinations. The one it loses is UniPPTEval on UniPPTBench, where Kimi K3 narrowly takes the lead. It is also the only model that never drops out of the top two, with an average rank of 1.11; Kimi K3 is 2.44, Opus 5 is 2.56, the M3 base is 4.22, and GPT-5.6 Sol is 4.67.
The gains generalize to out-of-distribution task sets. Gen-1 Slides improves on the base model across all nine comparisons. Scores on the public graders are also tight at the top: across those six comparisons, the difference between first and second never exceeds 1.7% of the full scale. Gen-1 Slides still ranks first in five of the six. Full per-dimension results are in Appendix C.
Does the grader agree with people?
A grader is only worth trusting if it agrees with human judgment. We checked ours against two independent signals: fine-grained annotation by our product and design team, and feedback from production users.
What the annotators found. Human evaluation covered 77 tasks (76 usable for the side-by-side comparison), one deck per model, 4,063 pages in total, blind-reviewed by our product and design team, with 20 pages double-annotated to measure inter-annotator agreement. It focused on the three dimensions a reader notices first: layout, source faithfulness, and aesthetics. Information delivery and narrative structure were not human-annotated, and tasks involving personally identifiable information were excluded before annotation.
Layout: comparable to Opus 5, slightly worse. 88% of Gen-1 pages had some layout issue vs. 81% for Opus 5, but the difference narrowed for serious problems: 33% of Gen-1 pages and 34% of Opus 5 pages were rated severe or worse. When reviewers compared the same tasks side by side, Gen-1 won 24, lost 26, and tied 26. Source faithfulness: slightly better than Opus 5. Across 9,141 atomic claims drawn from 30 decks(6 tasks × 5 models), Gen-1's fabrication rate was 13.3% vs. 17.7% — a 4.4-point gap. Re-annotation leaves about three points of judgment noise, so we say 'slightly better,' not 'significantly better.' Aesthetics: no difference detected. In the designer's pairwise comparisons, Gen-1 Slides and Opus 5 win 49% and 51% of the style judgments; the craft channel shows no significant difference either.
Table 2. Human evaluation, Gen-1 Slides vs. Claude Opus 5.
| Dimension / Metric | Gen-1 Slides | Claude Opus 5 |
|---|---|---|
| Layout: pages with any issue (%) | 88 | 81 |
| Layout: pages severe or worse (%) | 33 | 34 |
| Layout: unusable pages (%) | 7 | 15 |
| Layout: same-task pairing (W : L : T) | 24 : 26 : 26 | — |
| Source faithfulness: atomic claims fabrication rate (%) | 13.3 | 17.7 |
| Aesthetics: style-channel win rate (60 pairs each) (%) | 49 | 51 |
Layout rows: scoring on ~42–48 pages per model, with roughly ±14 points of confidence interval, so two- or three-point gaps are not interpreted. Pairing row: two-sided binomial test, not significant.
Human and grader calls line up on all three dimensions: annotators kept 88% of the grader's page-level layout calls in a 229-page audit, and a three-judge majority matched the human gold standard on 48 of 49 sampled fabrication points. Where they disagree, the pattern is consistent: the grader is too strict on slightly empty pages and misses small text. Humans also disagree with each other, which is the caveat behind any "alignment" claim. Annotation examples are in Appendix E.
What users did. During the staged release from August to September, across 1.57 million tasks, Gen-1 Slides and Claude Opus 5 are indistinguishable on average rating, low-rating share, thumbs ratio, and download rate. The untrained M3 base lags clearly: 0.45 stars lower on average rating, a low-rating share roughly five times higher, less than half the thumbs ratio — only download rate is close (30.4%).
Table 3. Production user feedback.
| Model | Avg. rating (★) | Low ratings (≤2★) | Up / down ratio | Download rate |
|---|---|---|---|---|
| Gen-1 Slides | 4.25 | 3.6% | 19.6 | 33.1% |
| Claude Opus 5 | 4.23 | 3.4% | 18.8 | 31.5% |
| MiniMax M3 (off-the-shelf) | 3.80 | 18.0% | 8.8 | 30.4% |
Live user feedback from Genspark AI Slides production traffic, August–September 2026. Average rating is the mean star rating out of 5; low-rating share is ratings of 2 stars or less; up/down ratio is thumbs-up to thumbs-down; download rate is the share of finished decks downloaded. Bold marks the best in each column.
Cost
Gen-1 Slides is priced at its base model's list rates: $0.30 per million input tokens and $1.20 per million output tokens, with cached input tokens (prompt-cache hits) at $0.06. Claude Opus 5 lists at $5.00 per million input tokens. On the number anyone can look up, Gen-1 Slides costs roughly 1/17 of Opus 5's list price.
List price is not what a deck costs, though. A model that renders, inspects, and revises its own pages writes more tokens per deck than one that skips those steps, and Gen-1 Slides does more of that than any model in the comparison. So we also report what a finished deck actually cost in our harness, on the same 200 production tasks.
Table 4. Measured cost per finished deck.
| Model | USD / deck | USD / 1,000 decks | Decks per $1 of Opus 5 budget | Aggregate score |
|---|---|---|---|---|
| Gen-1 Slides | 0.44 | 435 | 9.5 | 0.821 |
| Claude Opus 5 | 4.16 | 4,155 | 1.0 | 0.810 |
| Kimi K3 | 2.00 | 1,999 | 2.1 | 0.723 |
| GPT-5.6 Sol | 2.01 | 2,006 | 2.1 | 0.668 |
| MiniMax M3 (off-the-shelf) | 0.34 | 338 | 12.3 | 0.563 |
Measured LLM spend on the 200-task run; all models share the same harness. Post-training adds about 29% to per-deck cost over the off-the-shelf M3 checkpoint while lifting the aggregate score by 0.26.
Measured that way, a finished Gen-1 deck costs $0.44 against $4.16 for Opus 5, roughly 9.5 times cheaper. That is narrower than the 17× list-price gap because Gen-1 spends more of its budget checking its own work. Post-training added about 29% to per-deck cost over the off-the-shelf M3 and lifted the score into the Opus 5 tier. We report both numbers because they answer different questions. List price tells you what the model costs to call. Measured spend tells you what it costs to get a deck you can ship.
Two things stand out. The only model whose internal-grader score exceeds Gen-1 Slides is Claude Fable 5.1, by a margin of 0.003, a difference too small to call, at roughly 14× the per-deck cost. And every general-purpose model sits either below Gen-1 on quality or above it on price. Nothing in the comparison is both cheaper and better.
That gap is why this is more than a model release. It makes Opus-tier slide quality cheap enough to be the default for every Genspark user rather than a premium tier. It changes the economics of running slides at scale: a team producing 1,000 decks a month goes from about $4,200 to about $440 with no measurable loss in quality. And it means the budget that used to serve one user at frontier quality can now serve roughly ten. As the Gen-1 family extends to other deliverables, that is the shape of the business we are building: frontier quality as the floor.
How it was trained
Gen-1 Slides was trained with RL on roughly 2,000 internally constructed slide tasks. Rollouts ran on the Genspark AI Slides cluster; parameter updates ran on the Fireworks platform.
After an extensive exploratory phase, the final RL run completed in one week on 96 NVIDIA B300 GPUs. Over the run, the mean training reward rose from 0.271 at the first step to an EMA of 0.765 at release (Figure 3). There was no reward collapse and no checkpoint rollback at any point in the run.
Figure 3. Reward during training.
Light points are the mean rollout reward at each optimizer step; the dark line is an EMA with decay 0.9; the shaded band is ±1σ over a 15-step window. The three background tints mark the three phases (256k GSPO → 512k GSPO → 512k GSPO + KL guard). Dashed lines are the training-time reward of Claude Opus 5 (0.721) and the MiniMax M3 base (0.324) on the 200-task held-out set, shown for reference only; training reward is computed on training tasks and is not directly comparable to them.
We will publish the full technical report soon. Compared with a conventional RL post-training recipe, this run improves on it in the following ways.
The online grader evolves with the policy to counter reward hacking. Slide quality is unusually easy to game. Over the run, the policy learned, in turn, to import a reference deck and pass it off as the finished product, to write "sources verified" into its report without checking anything, and to shrink fonts until overflow detection stopped firing. Each of these satisfies a check and hands the user a worse deck. Reward hacking therefore cannot be eliminated once and for all, and as RL scale and model capability grow, hacks become more complex and harder to detect. So rather than trying to get the grader right in one shot, we update it online and let it evolve with the policy, and every update is judged against signals independent of the grader. Specifically:
- A watcher agent monitors training routinely. It continuously analyzes rollout processes and outputs; each of the patterns above was caught this way and closed promptly by an online grader update.
- Human designers run ongoing reviews. Their verdicts and suggestions are aligned back into the online grader and serve as the acceptance test for every grader version.
Learning from comments: OPSD with grader comments. During training we observed that a slides trajectory spans dozens of turns and hundreds of thousands of tokens, yet ends with a single scalar reward telling the policy whether the deck is good. That signal carries too little information: the model learns how well it did, not what went wrong or how to fix it. The grader's comments contain exactly that information, such as "body text on page 4 runs off the canvas" or "page 2 has only an image and no content", but the model never sees them, and can only improve by rolling out again and again in the state space, by trial and error.
So Gen-1 Slides tries to learn directly from grader comments. When the grader gives a concrete suggestion, we internalize that comment into the policy: teacher and student are the same model, the teacher's context additionally contains the comment, and we run on-policy self-distillation (OPSD) with a token-weighted KL from the student to the teacher, computed on the student's own samples. Experiments show that a handful of gradient updates is enough to internalize a grader comment into the policy, a large gain in sample efficiency.
"AI slop": an adversarial discriminator. Users pass the same verdict on many AI-generated decks: they read as AI-made, and audiences now file that look under "AI slop". The tells are consistent but hard to capture with a finite set of rules. Borrowing from GAN-style adversarial reward learning, we treat the policy as the generator and train a small discriminator to tell its slides from human-designed decks, used as an internal grader signal and updated online. Re-scoring every checkpoint with the final, frozen discriminator, the share of Gen-1 Slides' output that reads as AI-made fell from 0.749 to 0.417 over training.
A curriculum over reward components shapes the optimization path. In multi-objective RL, conflicts between metrics decide how efficiently the model learns. Sustained single-metric experiments on smaller models revealed a clear conflict between visual design and layout defects. Based on that, we reward complete, polished design early in training and raise the weight on layout defects step by step once the style stabilizes; the weighting and switch points were chosen from training-run metrics and rolled out through online grader updates.
Training stability:
-
The training environment is the production environment. Rather than copying the harness onto the training cluster, rollouts run directly in the live serving stack, removing the train–serve distribution shift.
-
GSPO replaces GRPO: sequence-level instead of token-level importance ratios. The reward is trajectory-level, while GRPO's per-token ratios give a high-variance gradient estimate over trajectories tens of thousands of tokens long. GSPO uses the length-normalized sequence likelihood ratio, the geometric mean of the token ratios, computed over model-generated tokens only, with tool outputs masked. Optimization granularity matches reward granularity, and the curve climbs smoothly.
-
Asynchronous RL with bounded staleness. The median episode takes 15 minutes and the longest over an hour, so rollouts and training run continuously in parallel. The sampler may be at most one policy version behind the trainer; weights are synced to it after every step, and samples older than one version are discarded.
-
Truncated importance sampling (TIS) corrects the training–inference mismatch; KL is only a guard. The FP8 inference engine and the BF16 trainer disagree slightly on every token's log-probability. This is a separate, token-level ratio, distinct from GSPO's sequence-level ratio; we correct it with truncated importance sampling at a threshold of 2.0, so no single token can dominate a batch's gradient. There is no reference-KL term. Before each update, the KL between the current policy and the rollout policy is estimated with the unbiased, low-variance k3 KL estimator; a batch above the threshold is dropped, and KL never enters the gradient.
Limitations
We'd rather say these ourselves before someone else does.
- It learned design, and pushed it too hard. Gen-1 Slides makes denser pages than any other model in the comparison: smaller type, more elements, and more pages that look top-heavy. If your audience reads slides on a phone, it might be a real weakness today.
- Content and task completion trail Opus 5. The base model started far behind on both. Gen-1 Slides closed most of that gap on content quality and a good part of it on task completion, but has not caught up on either.
- It is built for slides and is not a general-purpose model. That is the only task we have evaluated and recommended it for. We have seen gains carry over to document generation and coding. On spreadsheets or research, expect performance close to its base model, which itself trails frontier models.
How to use it
- Genspark AI Slides. Standard mode now uses Gen-1 Slides by default. Ultra mode continues to offer frontier models side by side in the model picker.
- Command line and coding agents. Install the Genspark CLI, log in, and create a deck with
gsk task create slides. Standard mode, and therefore Gen-1 Slides, is the default. In coding agents such as Claude Code, the ACP gsk-slides agent drives multi-turn generation and editing. - API. OpenRouter listing is in progress under the Genspark brand. Early access: [email protected].
- For enterprise workspaces. No customer content was used to train or evaluate Gen-1 Slides, and your data is never shared with the base model's developer. Existing workspace model restrictions apply to it unchanged: if your organization limits AI Slides to specific providers, Standard mode continues to use an approved model as before. Contact your account team to enable or disable Gen-1 Slides for your workspace.
What's next
Gen-1 is a family name, not a one-off release. The same recipe — a task-specific environment plus a reward computed on the delivered artifact — extends to spreadsheets, documents, and research. Models for those tasks are already in training.
On slides, this checkpoint is a starting point. We will keep training against the same production harness and the same three signals, with the goal of matching and then surpassing the leading frontier models on every dimension, not just on average. As newer checkpoints clear that bar, they will move up the tiers, and each release lifts the floor of Standard mode behind it.
Next up is native PowerPoint (OOXML) output, so a deck the model designs survives a round-trip through PowerPoint intact.
If you use Genspark AI Slides, there is nothing to set up. Open Standard mode, and your next deck is already on Gen-1. We can’t wait to see what you make with it.
Try Gen-1 Slides in Genspark AI Slides →
Acknowledgements
Gen-1 Slides is trained jointly with Fireworks AI. It is built on MiniMax M3, an open-weight, natively long-context MoE base designed for agentic tool calling, which let us skip pre-training and spend our entire budget on the slide loop. Without such an open base, Gen-1 Slides could not have been built in a matter of weeks. We thank its creators for releasing its weights openly, and the authors of the PPTAgent project and UniPPTBench for open-sourcing their benchmarks and graders. The model runs on US infrastructure, with no data flowing to the base model's developer.
Appendix
A. Evaluation Setting
We picked the four models that matter most in our production slides stack: Claude Opus 5, Kimi K3, GPT-5.6 Sol, and MiniMax M3, plus our trained Gen-1 Slides — five models under test in total. On the datasets side, we drew from both internal and external sources: 200 internal tasks and 254 tasks from two public benchmarks, for 454 tasks in total. On the grader side, we combined our internal grader with the two graders shipped alongside the public benchmarks. The result is a 5 models × 3 datasets × 3 graders evaluation grid. Every deck comes out of the same Genspark slides harness, so the model is the only thing that changes. Each of the 45 cells carries a score on every dimension the grader defines, plus its overall score.
B. Dataset and Grader
We use three datasets: Real tasks, UltraPresent, and UniPPTBench
Real tasks (200, ours): de-identified tasks from consumer users who have AI Data Retention enabled; Team and Enterprise data is never used. It is the only one of the three that reflects the real user intent distribution. UltraPresent(128): arXiv 2602.22839. Each task pairs a natural-language brief with hard constraints such as slide count and aspect ratio, in English and Chinese. We use the full validation split. The same authors created the popular PPTAgent, which introduced PPTEval; we score UltraPresent with PPTEval, the grader from that earlier paper. UniPPTBench (126): arXiv 2605.17356, the paper defines both the task set and its evaluator, UniPPTEval, covering four input settings: vague prompt, long document, multimodal document, and multi-source. We use all 126 tasks.
For graders, in addition to our internal grader, we use the official graders released alongside the two public benchmarks.
Internal grader v3.2 — scores on four dimensions (task completion, content quality, visual design, process quality) on a 0–1 scale. It is aligned with the real user experience of our system and its score also serves as the reward for Gen-1 Slides' RL training. The multimodal LLM judge is claude-fable-5. PPTEval(PPTAgent's grader) — three dimensions, content, design and coherence, on a 1–5 scale. We replaced the paper's now-dated GPT-4o judge with claude-fable-5, leaving the rubric unchanged. UniPPTEval (UniPPTBench grader) the official grader released with UniPPTBench; 10 base dimensions plus a visual-integrity dimension, 10-point scale. When scoring the Real tasks and UltraPresent datasets, we leave the scenario dimension unscored (N/A) because those datasets have no corresponding scenario specification. Internally we use the paper-specified multimodal LLM judge, gemini-3-flash-preview.
C. Full per-dimension tables
How to read these tables. The three graders use different scales (0–1, 1–5, 0–10), so we never average scores across graders, only ranks within a column. Dataset sizes are labeled at full size (200 / 128 / 126). Thinking-effort settings match production and are listed in "How we evaluate."
C1. Internal grader v3.2 scores (0–1 scale)
Real tasks (n=200)
| Model | Score | Task completion | Content quality | Visual design | Process quality |
|---|---|---|---|---|---|
| Gen-1 Slides | 0.821 | 0.780 | 0.737 | 0.756 | 0.719 |
| Claude Opus 5 | 0.810 | 0.849 | 0.782 | 0.688 | 0.695 |
| Kimi K3 | 0.723 | 0.840 | 0.784 | 0.557 | 0.748 |
| GPT-5.6 Sol | 0.668 | 0.820 | 0.743 | 0.479 | 0.748 |
| MiniMax M3 (off-the-shelf) | 0.563 | 0.741 | 0.668 | 0.494 | 0.608 |
UltraPresent (n=128)
| Model | Score | Task completion | Content quality | Visual design | Process quality |
|---|---|---|---|---|---|
| Gen-1 Slides | 0.932 | 0.874 | 0.804 | 0.898 | 0.768 |
| Claude Opus 5 | 0.883 | 0.878 | 0.800 | 0.876 | 0.615 |
| Kimi K3 | 0.874 | 0.888 | 0.833 | 0.765 | 0.792 |
| GPT-5.6 Sol | 0.798 | 0.860 | 0.800 | 0.671 | 0.765 |
| MiniMax M3 (off-the-shelf) | 0.777 | 0.852 | 0.766 | 0.722 | 0.664 |
UniPPTBench (n=126)
| Model | Score | Task completion | Content quality | Visual design | Process quality |
|---|---|---|---|---|---|
| Gen-1 Slides | 0.867 | 0.846 | 0.816 | 0.814 | 0.722 |
| Claude Opus 5 | 0.859 | 0.884 | 0.812 | 0.866 | 0.518 |
| Kimi K3 | 0.855 | 0.884 | 0.845 | 0.751 | 0.768 |
| GPT-5.6 Sol | 0.780 | 0.836 | 0.798 | 0.662 | 0.726 |
| MiniMax M3 (off-the-shelf) | 0.579 | 0.817 | 0.713 | 0.513 | 0.616 |
The score is a weighted blend of four dimensions (task completion, content quality, visual design, process quality), adjusted by penalties for layout defects, fabrication, misrepresentation and ignored instructions, plus a bonus for a clean deck, and capped at 1.
C2. PPTEval scores (1–5 scale)
Real tasks
| Model | PPTEval average | Content | Design | Coherence |
|---|---|---|---|---|
| Gen-1 Slides | 3.783 | 3.423 | 3.915 | 4.010 |
| Kimi K3 | 3.714 | 3.429 | 4.039 | 3.674 |
| Claude Opus 5 | 3.655 | 3.412 | 3.808 | 3.746 |
| MiniMax M3 (off-the-shelf) | 3.655 | 3.274 | 3.681 | 4.010 |
| GPT-5.6 Sol | 3.416 | 3.283 | 3.547 | 3.420 |
UltraPresent
| Model | PPTEval average | Content | Design | Coherence |
|---|---|---|---|---|
| Gen-1 Slides | 3.812 | 3.483 | 3.952 | 4.000 |
| Kimi K3 | 3.770 | 3.524 | 4.005 | 3.782 |
| Claude Opus 5 | 3.701 | 3.460 | 3.843 | 3.800 |
| MiniMax M3 (off-the-shelf) | 3.677 | 3.305 | 3.652 | 4.073 |
| GPT-5.6 Sol | 3.440 | 3.359 | 3.487 | 3.473 |
UniPPTBench
| Model | PPTEval average | Content | Design | Coherence |
|---|---|---|---|---|
| Gen-1 Slides | 3.840 | 3.441 | 3.860 | 4.218 |
| Kimi K3 | 3.796 | 3.525 | 3.903 | 3.960 |
| MiniMax M3 (off-the-shelf) | 3.687 | 3.285 | 3.518 | 4.258 |
| Claude Opus 5 | 3.610 | 3.389 | 3.506 | 3.936 |
| GPT-5.6 Sol | 3.459 | 3.358 | 3.455 | 3.564 |
C3. UniPPTEval scores (0–10 scale)
Real tasks — scenario dimension N/A (no scenario spec in this dataset)
| Model | Overall score | Base average | Visual integrity |
|---|---|---|---|
| Gen-1 Slides | 8.858 | 8.834 | 9.099 |
| Claude Opus 5 | 8.709 | 8.699 | 8.804 |
| Kimi K3 | 8.684 | 8.623 | 9.287 |
| MiniMax M3 (off-the-shelf) | 8.401 | 8.428 | 8.128 |
| GPT-5.6 Sol | 7.697 | 7.707 | 7.598 |
UltraPresent — scenario dimension N/A
| Model | Overall score | Base average | Visual integrity |
|---|---|---|---|
| Gen-1 Slides | 8.954 | 8.993 | 8.571 |
| Claude Opus 5 | 8.915 | 8.936 | 8.718 |
| Kimi K3 | 8.821 | 8.890 | 8.152 |
| MiniMax M3 (off-the-shelf) | 8.452 | 8.491 | 8.060 |
| GPT-5.6 Sol | 7.825 | 7.829 | 7.817 |
UniPPTBench
| Model | Overall score | Base average | Visual integrity | Scenario average |
|---|---|---|---|---|
| Kimi K3 | 8.995 | 9.026 | 9.288 | 6.759 |
| Gen-1 Slides | 8.834 | 8.964 | 8.822 | 6.363 |
| Claude Opus 5 | 8.805 | 8.817 | 8.397 | 6.916 |
| MiniMax M3 (off-the-shelf) | 8.499 | 8.587 | 8.418 | 6.326 |
| GPT-5.6 Sol | 8.069 | 8.028 | 7.987 | 6.259 |
C3b. The 10 base sub-dimensions of UniPPTEval (0–10)
The base average column in C3 is the mean of these 10 dimensions; we break them out for per-dimension audit. Columns: instruction fulfillment, content accuracy, information density, clarity & legibility, narrative coherence, engagement, visual consistency, layout quality, hierarchy & emphasis, visual element quality.
Real tasks
| Model | Instr. | Accuracy | Density | Clarity | Narrative | Engage. | Consist. | Layout | Hierarchy | Vis. elem. |
|---|---|---|---|---|---|---|---|---|---|---|
| Gen-1 Slides | 9.741 | 9.135 | 8.870 | 8.710 | 9.565 | 8.378 | 9.503 | 9.078 | 8.990 | 6.373 |
| Claude Opus 5 | 9.736 | 9.145 | 8.829 | 8.580 | 9.497 | 7.917 | 9.135 | 8.772 | 8.700 | 6.679 |
| Kimi K3 | 9.596 | 9.072 | 8.731 | 8.845 | 9.342 | 7.617 | 9.300 | 8.964 | 8.793 | 5.974 |
| GPT-5.6 Sol | 9.119 | 8.912 | 7.668 | 7.197 | 8.896 | 7.135 | 8.026 | 7.249 | 7.492 | 5.378 |
| MiniMax M3 (off-the-shelf) | 9.420 | 8.482 | 8.596 | 8.197 | 9.378 | 7.824 | 9.124 | 8.461 | 8.539 | 6.259 |
UltraPresent
| Model | Instr. | Accuracy | Density | Clarity | Narrative | Engage. | Consist. | Layout | Hierarchy | Vis. elem. |
|---|---|---|---|---|---|---|---|---|---|---|
| Gen-1 Slides | 9.907 | 8.667 | 9.263 | 8.927 | 9.621 | 8.724 | 9.569 | 9.345 | 9.351 | 6.638 |
| Claude Opus 5 | 10.000 | 9.182 | 9.276 | 8.793 | 9.627 | 8.373 | 9.379 | 8.862 | 9.000 | 6.911 |
| Kimi K3 | 9.980 | 9.231 | 9.017 | 9.086 | 9.561 | 8.121 | 9.492 | 9.345 | 9.069 | 6.158 |
| GPT-5.6 Sol | 9.700 | 8.836 | 7.982 | 7.085 | 9.017 | 7.825 | 7.862 | 6.949 | 7.536 | 5.786 |
| MiniMax M3 (off-the-shelf) | 9.782 | 8.000 | 8.982 | 8.362 | 9.559 | 8.271 | 9.368 | 8.321 | 9.017 | 5.288 |
UniPPTBench
| Model | Instr. | Accuracy | Density | Clarity | Narrative | Engage. | Consist. | Layout | Hierarchy | Vis. elem. |
|---|---|---|---|---|---|---|---|---|---|---|
| Gen-1 Slides | 9.871 | 8.766 | 9.121 | 8.581 | 9.718 | 8.839 | 9.444 | 8.968 | 8.952 | 7.379 |
| Claude Opus 5 | 9.702 | 8.903 | 9.024 | 8.210 | 9.710 | 8.637 | 9.097 | 8.452 | 8.573 | 7.863 |
| Kimi K3 | 9.839 | 9.008 | 9.169 | 8.831 | 9.702 | 8.508 | 9.548 | 9.306 | 9.089 | 7.258 |
| GPT-5.6 Sol | 9.419 | 9.089 | 8.000 | 6.814 | 9.419 | 8.065 | 8.113 | 7.073 | 7.484 | 6.806 |
| MiniMax M3 (off-the-shelf) | 9.710 | 8.516 | 8.726 | 8.008 | 9.734 | 8.565 | 9.145 | 8.065 | 8.653 | 6.740 |
C4. Per-column rank and average rank
Ranks are taken within each column (scores sorted high-to-low), then averaged across the nine columns. Lower is better.
| Grader × dataset | Gen-1 Slides | Claude Opus 5 | Kimi K3 | GPT-5.6 Sol | MiniMax M3 (off-the-shelf) |
|---|---|---|---|---|---|
| Internal v3.2 · Real tasks | 1 | 2 | 3 | 4 | 5 |
| Internal v3.2 · UltraPresent | 1 | 2 | 3 | 4 | 5 |
| Internal v3.2 · UniPPTBench | 1 | 2 | 3 | 4 | 5 |
| PPTEval · Real tasks | 1 | 3 | 2 | 5 | 4 |
| PPTEval · UltraPresent | 1 | 3 | 2 | 5 | 4 |
| PPTEval · UniPPTBench | 1 | 4 | 2 | 5 | 3 |
| UniPPTEval · Real tasks | 1 | 2 | 3 | 5 | 4 |
| UniPPTEval · UltraPresent | 1 | 2 | 3 | 5 | 4 |
| UniPPTEval · UniPPTBench | 2 | 3 | 1 | 5 | 4 |
| Average rank | 1.11 | 2.56 | 2.44 | 4.67 | 4.22 |
D. Notes on Figure 1
Vertical axis: aggregate internal grader (v3.2) score, with each point the mean over the real tasks completed by all eleven models. The body text and tables use a separate five-model run (n=200; Gen-1 Slides 0.821, Claude Opus 5 0.810), so the figures there differ from the values plotted here by a different batch; the gap sits within generation variance. Axis truncated to 0.45–0.9. Horizontal axis: list input-token price (USD per million tokens, log scale). Claude Opus 5 is the primary baseline because it is the model Genspark AI Slides currently runs at the largest scale, so we have real, production-scale data on its slide quality and cost. Claude Fable 5 / 5.1 have limited data on real tasks and are shown for reference only.
E. Human evaluation: method and agreement audit
How we annotate slides: layout
Reviewers inspect the rendered slide for concrete layout defects, checking the grader's flags and adding issues it missed. The example below shows four problems on the same page: three flagged by the AI grader and one added by human review. These are checks on readability and arrangement, separate from whether the underlying claims are supported by the source.
Annotated example: Q3 2026 at a Glance. Numbers 1-4 identify the regions discussed below; they are not severity scores. The AI grader flagged 1, 3, and 4. A human reviewer added 2. Explanations follow the supplied annotation notes.
1. Overlap and overflow: flagged by AI. The lower-right "Sales Mix by Channel" table is taller than its card. Its final row, "Catering 3% 3%," spills outside the card and overlaps the takeaway text below, leaving two layers of text in the same space and making them difficult to read.
2. Text too small: missed by AI, added by a human. The value labels, region names, and vertical-axis tick labels in the left-hand bar chart are only 10-11 px, compared with 17-19 px for the page's body text, according to the annotation notes. They are difficult to read when projected. The grader did not flag this; the reviewer added it.
3. Excessive empty space: flagged by AI. The upper-right line chart is compressed into the left side of its card, leaving roughly the rightmost third unused. The annotation identifies an imbalance within this chart card, not a rule that all whitespace is a defect.
4. Alignment and grid inconsistency: flagged by AI. The top KPI cards do not form a consistent grid: the second card is taller, so its bottom edge extends below the others, and the gaps between cards are uneven.
What human review adds. A page can have a coherent color palette and still fail basic readability checks. Here, human review adds a small-text defect that the grader missed, alongside its overlap, whitespace, and alignment flags. The example illustrates how the labels are assigned; aggregate layout results are reported separately below.
How we annotate slides: source faithfulness
Reviewers check individual claims against the supplied source material, distinguishing directly supported facts, supported inferences, and fabrications (claims neither stated nor supported by the source). The example below shows seven checks and two cases where human review changed the automated judgment.
Annotated example: Strategic Options Matrix. Green (1-2): directly supported facts. Yellow (3): supported inference. Red (4-5): fabrications. Purple (6-7): human-grader disagreements, not a fourth label. Explanations below summarize the supplied annotation notes.
1. Fleet size: directly supported. The source states 22 vehicles, with 18 in daily service and four in reserve. Humans and grader agree.
2. Load factors: directly supported. Both 82% in 2024 and 54% in H1 2026 come from the source. The annotation treats the slide's "FY 2024" wording as equivalent to "2024." Humans and grader agree.
3. A 28-percentage-point decline: supported inference. The source does not state 28, but it follows from 82 minus 54. Humans and grader agree.
4. Cargo-pivot capex of NGN 85-120 million: fabrication. The source gives no investment amount and does not ask the model to estimate one. Humans and grader agree.
5. Downsizing to 12-14 vehicles: fabrication. The source gives the current fleet size, not a target fleet size or enough evidence to derive one. Humans and grader agree.
6. "Strong" B2B charter margins: human correction to fabrication. The grader initially accepted this as a supported inference because the task requested a comparison matrix with a margin column. The reviewer rejected that rationale: requesting a column is not evidence for any particular rating. Without supporting facts, the cell should be blank or marked "To be assessed." After this rule was clarified, the grader also labeled it a fabrication.
7. "Declining" passenger operating margins: mathematically supported, not fabricated. Monthly passenger revenue fell by 35.7%, while user-provided, like-for-like data show that operating costs for the same business declined by only 11%. This mathematically establishes a lower operating margin, although no margin figure was directly reported.
The boundary matters. In the same column, human review rejected one unsupported rating and accepted another as evidence-based. Reviewers check the support for each claim rather than accepting or rejecting an entire column at once.
