Benchmark · November 12, 2026
Saymock vs Figma (MCP): the same agent on three screens
Claude Opus 5.5 reproduced three Wikipedia screens as editable designs, once with Saymock and once with Figma’s official MCP server. Same prompt, same assets, same budget, three runs per screen and per tool, scored blind. Here are the protocol and every number, including the ones that don’t flatter us.
Results
Median of each screen’s three runs, then the mean of the three screens. Tokens and dollars are what Claude Code measures for the agent doing the design, at API prices; export and scoring are excluded.
| Per screen | Saymock | Figma (MCP) | Difference |
|---|---|---|---|
| Tokens processed | 1.88M | 2.74M | −31% |
| of which output tokens | 30k | 61k | −50% |
| Cost | $1.70 | $2.69 | −37% |
| Time | 5.4 min | 9.8 min | 1.8× faster |
| Agent turns | 34.7 | 38.7 | −10% |
| Tool errors | 1.0 | 4.0 | 4× fewer |
| Score /100 | 79.0 | 82.2 | +3.2 for Figma: a tie |
| Visual fidelity /60 | 47.0 | 49.7 | +2.7 for Figma |
| Editability /40 | 31.3 | 32.8 | +1.5 for Figma |
| Successful runs | 9/9 | 9/9 |
Takeaways
- Quality: a tie. Figma leads by 3.2 points on average, under the 5-point threshold our protocol sets. It is ahead on each of the three screens, by 1.5 (iPhone) to 6 points (Android).
- Cost: Saymock uses less. 31% fewer tokens, half the output tokens, 37% fewer dollars per screen.
- Time: 1.8× faster. 5.4 minutes per screen against 9.8.
- Not on every screen. On the iPhone, Saymock’s median run processed more tokens than Figma’s (3.07M against 1.91M), while still costing less. On Android and the web, Saymock processes 2.3 to 2.6 times fewer.
Why the dollar gap is wider than the token gap: a token’s price depends on whether it is re-read from the cache ($0.20 per million), read for the first time ($4 to $8) or written by the model ($20). With Saymock, the model writes half as many tokens, the most expensive ones.
By screen
Median of each tool’s three runs.
| Screen | Score /100 | Tokens | Cost | Time | ||||
|---|---|---|---|---|---|---|---|---|
| Saymock | Figma | Saymock | Figma | Saymock | Figma | Saymock | Figma | |
| iPhone | 78.5 | 80.0 | 3.07M | 1.91M | $2.29 | $2.58 | 6.9 min | 9.7 min |
| Android | 78.0 | 84.0 | 1.20M | 3.14M | $1.44 | $2.88 | 4.6 min | 10.5 min |
| Web | 80.5 | 82.5 | 1.38M | 3.18M | $1.38 | $2.62 | 4.8 min | 9.2 min |
Protocol
The screens
Three Wikipedia screens of different kinds, captured from the real apps. Each agent gets the screenshot and the app’s real files: photos, icons, logo, and the fonts used for each text role.
| Screen | Content | Size |
|---|---|---|
| iPhone | iOS app home: feed of cards, photos, segmented control, translucent tab bar | 430×932, @3x |
| Android | Android app home: banner, image card, Material 3 navigation | 411×923, @2.625x |
| Web | Wikipedia’s Main Page: columns, colored boxes, photos, side panel | 1440×900, @2x |
The same conditions for both tools
- Same agent: Claude Code in non-interactive mode, model Claude Opus 5.5. Each run starts from an empty folder, with no context or previous session.
- Same prompt: “reproduce this screen as an editable design”. Each tool only adds how to create its file.
- One tool loaded: Saymock’s MCP server (the production version) or Figma’s official MCP server, never both, and no other connectors.
- Same budget: $15 and 45 minutes per run, one run at a time.
- Hands off: nobody fixes, nudges or resumes an agent. A failed run would count as one; all 18 succeeded.
Scoring
- Each result is scored out of 100: visual fidelity out of 60 (layout, typography, colors, icons and imagery, content, pixel metrics) and editability out of 40 (layer hierarchy and names, flowing layout, reuse, cleanliness).
- Double blind: two independent judges (Claude Opus 5.5, a fresh session per result) score each result without knowing which tool made it. They were never more than 5 points apart.
- For each screen and tool we keep the median of the three runs, then average the three screens. Each criterion is its own median, so visual plus editability does not add up exactly to the score.
- A gap under 5 points is a tie, and is reported as one.
All 18 runs
The raw numbers of every run. Saymock’s first iPhone run (85 turns, 6.83M tokens) is the one that processed the most tokens; the median leaves it out, this table shows it.
| Screen | Tool | Run | Score | Visual | Edit. | Tokens | Output | Cost | Time | Turns | Errors |
|---|---|---|---|---|---|---|---|---|---|---|---|
| iPhone | Saymock | 1 | 80.0 | 46.5 | 33.5 | 6.83M | 50k | $3.59 | 17.1 min | 85 | 2 |
| iPhone | Saymock | 2 | 78.5 | 45.5 | 33.0 | 3.07M | 36k | $2.29 | 6.9 min | 49 | 1 |
| iPhone | Saymock | 3 | 78.5 | 45.5 | 33.0 | 1.63M | 23k | $1.54 | 4.2 min | 28 | 0 |
| iPhone | Figma | 1 | 81.5 | 48.0 | 33.5 | 5.99M | 101k | $4.52 | 16.7 min | 66 | 4 |
| iPhone | Figma | 2 | 74.5 | 45.0 | 29.5 | 1.68M | 45k | $1.95 | 7.6 min | 29 | 3 |
| iPhone | Figma | 3 | 80.0 | 46.5 | 33.5 | 1.91M | 65k | $2.58 | 9.7 min | 29 | 4 |
| Android | Saymock | 1 | 78.0 | 46.0 | 32.0 | 1.15M | 28k | $1.40 | 4.6 min | 25 | 1 |
| Android | Saymock | 2 | 79.0 | 46.0 | 33.0 | 1.20M | 27k | $1.44 | 4.3 min | 21 | 1 |
| Android | Saymock | 3 | 77.0 | 45.5 | 31.5 | 1.30M | 30k | $1.56 | 4.9 min | 32 | 1 |
| Android | Figma | 1 | 85.5 | 53.0 | 32.5 | 2.56M | 52k | $2.30 | 9.4 min | 33 | 3 |
| Android | Figma | 2 | 84.0 | 50.0 | 34.0 | 3.50M | 65k | $2.92 | 10.8 min | 46 | 3 |
| Android | Figma | 3 | 82.5 | 52.5 | 30.0 | 3.14M | 61k | $2.88 | 10.5 min | 41 | 2 |
| Web | Saymock | 1 | 78.0 | 49.5 | 28.5 | 1.28M | 26k | $1.38 | 4.8 min | 30 | 2 |
| Web | Saymock | 2 | 80.5 | 49.5 | 31.0 | 1.38M | 24k | $1.38 | 4.3 min | 28 | 1 |
| Web | Saymock | 3 | 80.5 | 51.5 | 29.0 | 1.74M | 32k | $1.76 | 5.8 min | 35 | 1 |
| Web | Figma | 1 | 83.5 | 51.0 | 32.5 | 3.18M | 55k | $2.62 | 9.2 min | 46 | 5 |
| Web | Figma | 2 | 82.0 | 48.5 | 33.5 | 1.72M | 46k | $2.03 | 8.1 min | 31 | 3 |
| Web | Figma | 3 | 82.5 | 50.0 | 32.5 | 3.61M | 67k | $3.06 | 10.5 min | 50 | 6 |
Where each tool loses points
Mean points lost per criterion (maximum in parentheses). Higher means weaker.
| Criterion | Saymock | Figma |
|---|---|---|
| Layout (15) | 2.5 | 2.2 |
| Typography (12) | 2.8 | 2.2 |
| Colors and surfaces (8) | 1.5 | 1.3 |
| Icons and imagery (10) | 2.0 | 1.2 |
| Content (10) | 0.7 | 0.3 |
| Pixel metrics (5) | 3.3 | 3.3 |
| Hierarchy and naming (10) | 1.3 | 1.0 |
| Flowing layout (12) | 2.2 | 1.8 |
| Reuse (10) | 3.7 | 3.2 |
| Cleanliness (8) | 1.0 | 1.0 |
- Saymock loses most on text (letter spacing of a substitute font, which shifts the line breaks), on a few icon glyphs, and on values repeated by hand instead of shared tokens.
- Figma loses most on fonts (Inter instead of SF Pro on the iPhone), on corner radii that are too large, and on the lack of text and color styles.
Limits
- Three screens from a single app: a solid trend, not a general law.
- One model (Claude Opus 5.5). Another model or agent may change the gaps.
- The judges are models, not human designers.
- Costs are API costs. On a Claude subscription you pay in quota, not dollars; tokens remain the comparable measure.
- We build Saymock: this benchmark is not independent. That is why we publish the protocol and every run.