The same model got the same 10 tasks twice: once with only plain Python inside the application, and once with MifBridge as shipped. A separate process checked every result. 120 runs in four waves, recorded on to . A1 to A8 run in Blender; X1 and X2 build in Blender and deliver into Unreal Engine. On A3, A5, A8 and X1 a third arm ran as well: MifBridge, Python off.
The build measured. MifBridge 1.3.0, the tagged release at commit 6cd630ca9 (2026-10-29). The Fab listing's recorded build when this was read (2026-11-01) was 1.2.0 (c2536bb8). Every MifBridge run in this round was checked to be on 6cd630ca9.
48/50right results, MifBridgePlain Python: 47/50.
0.94xtokens, average taskMifBridge 106,015 against 113,226, the mean of each task’s mean
69,007tokens, median run, MifBridgePlain Python: 84,130, the middle run of all 50
3/10tasks where MifBridge used moreits median run above plain Python’s
0/0crashes, MifBridge / PythonNeither arm crashed the application
The short version
What this measures
Tokens and verified output. A separate process checks each saved result against the task's stated spec. Nobody scored how the assets look, and this round captured no pictures, so a verified run is a correct one by the spec and says nothing about its looks.
Right results
MifBridge got 48 of 50 runs right and plain Python 47 of 50. MifBridge got more runs right on A3 (A3: 5 of 5 against 3 of 5). Plain Python got more runs right on X2 (X2: 5 of 5 against 4 of 5).
Crashes
The application never stopped answering in a plain Python run. It never stopped in a MifBridge run. A crashed run left nothing to grade, so it counts against its arm.
The average
Averaged task by task, MifBridge used 0.94x plain Python’s tokens: 106,015 against 113,226. Plain Python’s costliest run was on X2 (517,728 tokens); leave X2 out and the average is MifBridge 71,351 against 100,223, so it still favors MifBridge.
The typical run
On 3 of the 10 tasks (A2, A5 and X2) MifBridge’s median run used more tokens than plain Python’s. It used fewer on A1, A3, A4, A6, A7, A8 and X1.
The costly task
X2 is the one large loss, at 1.48x by median. Two of its five runs hit the same failed collision read-back in the hand-off to Unreal. What happened, and the fix.
MifBridge, Python off
On the 4 tasks it ran (A3, A5, A8 and X1), it used 3.22x plain Python’s tokens, averaged task by task, and got 19 of 20 runs right. It called a typed asset tool in 20 of 20 runs, where MifBridge as shipped did in 15 of 50. The headline figures compare plain Python with MifBridge as shipped only.
Time
Not reported this round. Other work shared the machine during waves 2 to 4, so times are not comparable. Tokens are this round's measure.
The worst case
MifBridge’s costliest run used 676,791 tokens (X2); plain Python’s used 517,728 (X2).
Dollars
Priced at API list prices, all 50 MifBridge runs came to $12.92 and all 50 plain Python runs to $14.53. On a Claude plan these runs use the plan’s allowance; nobody is billed these amounts.
What moved since round ten
The same tasks, prompts, checks and arms on three builds. Round ten B re-ran round ten on a later development build and has no page of its own. Its figures here come from its export, as this round’s do.
Measure
Round ten
Round ten B
Round eleven
Build measured
development build 2a026a58
development build bdba507e
1.3.0, a tagged release (6cd630ca)
Unreal Engine on X1 and X2
5.8.2
5.8.2
5.8.3
MifBridge as shipped against plain Python, tokens, mean of task means
1.24x
0.76x
0.94x
The same, geometric mean of the task medians
1.13x
0.97x
0.83x
Tasks at or below plain Python’s median
3 of 10
4 of 10
7 of 10
Verified, MifBridge as shipped
46 of 50
43 of 50
48 of 50
Verified, plain Python
45 of 50
42 of 50
47 of 50
MifBridge, Python off against plain Python, on its 4 tasks
3.13x
2.09x
3.22x
Verified, MifBridge, Python off
18 of 20
20 of 20
19 of 20
Plain Python’s own cost swings between rounds, so each ratio belongs to its own round. Round ten B’s mean leans on one plain Python X1 run of 1,511,421 tokens, which is why the geometric mean sits beside it.
The build
Round eleven measured the tagged 1.3.0 release. Rounds ten and ten B measured development builds. Tasks, prompts, checks, thresholds and arms are the same in all three.
The engine
X1 and X2 ran on UE 5.8.3, because no 5.8.2 is left on the machine. Rounds ten and ten B ran them on 5.8.2. The round changes the engine and the build together, so neither can take the credit or the blame alone for what moved on those two tasks.
The login
The pilot wave ran on the machine's default Claude login, as rounds ten and ten B did. Waves 2 to 4 ran on the benchmark owner's own login. The first request moved by 5 tokens: 3,971 to 3,976 for MifBridge and 3,135 to 3,140 for plain Python. Claude Code stayed at 2.1.280 throughout. Whether the login moved anything past the first request is not established.
The project
The bench project was deleted by mistake in a disk cleanup and rebuilt for the later waves from the same MifBridge checkout, with the same plugin list and fixture. The checkers were tested again before the counted runs: 86 cases on A1 to A8 and 18 of 18 on X1 and X2.
The change cap
1.3's shipped cap of 25 distinct targets a call is on, as a buyer gets it. Whether any run met its preview was not checked.
Runs that did not count
Six runs ended as harness errors with no model run, because what one X2 run left behind blocked the reset. One X1 run with Python off, 242,537 tokens in, ended when the editor was stopped at the machine's free memory floor. All seven were run again; none is in a table here, and those tokens are outside the round's total.
Where MifBridge used more tokens (3)
Task
Median run, Python
Median run, MifBridge
Median
Mean
X2, cross-tool: kit to UnrealBlender + Unreal
188,092
278,562
1.481x
1.815x
A2, modular kitBlender
35,646
40,584
1.139x
1.184x
A5, organic: plantsBlender
41,396
47,032
1.136x
1.133x
Where MifBridge used fewer (7)
Task
Median run, Python
Median run, MifBridge
Median
Mean
A4, organic: rock variantsBlender
177,089
86,049
0.486x
0.753x
X1, cross-tool: prop to UnrealBlender + Unreal
223,625
108,841
0.487x
0.428x
A3, organic: rockBlender
104,833
64,213
0.613x
0.736x
A6, LODs and exportBlender
60,376
47,642
0.789x
0.591x
A8, organic: treeBlender
84,230
70,371
0.835x
0.777x
A7, vertex-color masksBlender
20,867
19,081
0.914x
0.956x
A1, game-ready passBlender
96,315
92,980
0.965x
0.981x
Median and Mean are MifBridge’s tokens divided by plain Python’s on the same task, so above 1 means MifBridge used more. A task is listed as a loss when its median is above 1, whichever arm got it right; the table below and the task rows say which did. MifBridge, Python off is read in its own section.
X2, the costly task
The numbers
MifBridge's five X2 runs read 676,791, 278,562, 217,842, 672,423 and 244,337 tokens. Three cost 218k to 279k and the other two more than twice as much, so the median and the mean both hide the shape.
What happened
In the two expensive runs, the first send to Unreal with kit: true failed its collision read-back: "the asset's collision is not its hulls". Each took 28 tool calls against 15 to 17, mostly Python in Unreal inspecting and repairing collision on a mesh that was right. One recovered by sending the pieces one at a time. The other ended wrong: its assets fell out of the registry, and the check read two of them as missing.
The cause
Two faults, found after the round. A send over a deleted asset that a Python name still held in the editor imported into it, which saved a file that never opens and has no registry row. And from the third send to one path on 5.8, the asset's stored import data blocked the collision setting, so the collision came back empty. The failed reply gave no sign the mesh itself was right, and its advice was to send under a new name.
How often
Read-back failures on this task: 0 of 5 in round ten, 1 of 5 in round ten B and 2 of 5 here. Five runs a round cannot separate those rates.
Fixed, and not measured again
The bug is fixed on MifBridge's master branch, after 1.3.0 and in no release yet. Nobody has measured it again, so every number on this page is 1.3.0's with the bug in it, and the fix says nothing about them. A 1.3.0 install still behaves this way, and what the fix saves is unknown until a later round runs.
A third arm: MifBridge, Python off
The same install with the Blender add-on's own 'Allow run_python' preference unticked, so Blender work goes through the typed tools (round ten's diagnostic arm; Unreal's Python stays reachable, MifBridge has no switch for it). It ran on A3, A5, A8 and X1, 20 runs, and is read here against MifBridge as shipped and against plain Python on the same tasks. The headline numbers above compare the other two arms only.
Task
Right, Python off
Right, MifBridge
Right, Python
Median run, Python off
Against MifBridge, median / mean
Against Python, median / mean
A3, organic: rockBlender
5/5
5/5
3/5
181,464
2.826x / 2.554x
1.731x / 1.880x
A5, organic: plantsBlender
5/5
5/5
5/5
162,782
3.461x / 3.765x
3.932x / 4.267x
A8, organic: treeBlender
5/5
4/5
4/5
320,289
4.551x / 4.552x
3.803x / 3.537x
X1, cross-tool: prop to UnrealBlender + Unreal
4/5
5/5
5/5
374,200
3.438x / 8.216x
1.673x / 3.514x
Its median run used more tokens than MifBridge as shipped on 4 of 4 tasks, and more than plain Python on 4 of 4; above 1 means Python off used more. X1 with Python off is very wide, 154,888 to 1,870,167 tokens: two of its runs took 43 and 47 requests with repeated delete and save work, and their transcripts were not read closely enough to say why.
Which tools MifBridge called
Counted from each run’s transcript: the runs that called at least one of MifBridge’s typed asset tools (bl_build_kit_piece, bl_build_lod_chain, bl_build_plant, bl_build_rock, bl_build_tree, bl_create_collision_hull, bl_mesh_quality and mif_send_to_unreal are the ones called this round), and the runs that only wrote Python. MifBridge as shipped keeps its Python tool, and in 35 of its 50 runs the model wrote Python and nothing else. With Python off, 20 of 20 runs called them.
A tool reached through MifBridge’s mif_call is counted under its own name. A run counted in neither column called only other tools.
Every task’s spread
Each row covers one arm’s runs of one task: the thin line runs from the cheapest run to the costliest, the bar is the middle half of the runs, and the tick is the median. The scale is logarithmic, because the costliest run is about 125 times the cheapest. Each task names the application it drives.
Plain PythonMifBridge as shippedMifBridge, Python off (A3, A5, A8 and X1)tokens per run
20k50k100k200k500k1M2M
A1Blender: game-ready pass
Plain Python (Blender), A1: fewest 78,084, middle half 81,149 to 98,081, median 96,315, mean 96,031, most 126,524 tokens over 5 runsPython
MifBridge as shipped (Blender), A1: fewest 75,146, middle half 87,674 to 103,916, median 92,980, mean 94,254, most 111,552 tokens over 5 runsMifBridge
A2Blender: modular kit
Plain Python (Blender), A2: fewest 28,132, middle half 35,126 to 36,145, median 35,646, mean 34,671, most 38,305 tokens over 5 runsPython
MifBridge as shipped (Blender), A2: fewest 39,287, middle half 40,228 to 41,423, median 40,584, mean 41,063, most 43,795 tokens over 5 runsMifBridge
A3Blender: organic: rock
Plain Python (Blender), A3: fewest 82,614, middle half 104,788 to 105,486, median 104,833, mean 106,780, most 136,178 tokens over 5 runsPython
MifBridge as shipped (Blender), A3: fewest 49,182, middle half 56,901 to 67,643, median 64,213, mean 78,601, most 155,067 tokens over 5 runsMifBridge
MifBridge, Python off (Blender), A3: fewest 164,045, middle half 169,035 to 182,183, median 181,464, mean 200,725, most 306,896 tokens over 5 runsPython off
A4Blender: organic: rock variants
Plain Python (Blender), A4: fewest 104,628, middle half 170,755 to 192,418, median 177,089, mean 178,348, most 246,851 tokens over 5 runsPython
MifBridge as shipped (Blender), A4: fewest 72,574, middle half 77,223 to 101,229, median 86,049, mean 134,330, most 334,573 tokens over 5 runsMifBridge
A5Blender: organic: plants
Plain Python (Blender), A5: fewest 33,133, middle half 35,746 to 43,574, median 41,396, mean 39,642, most 44,359 tokens over 5 runsPython
MifBridge as shipped (Blender), A5: fewest 27,898, middle half 35,887 to 50,148, median 47,032, mean 44,927, most 63,671 tokens over 5 runsMifBridge
MifBridge, Python off (Blender), A5: fewest 124,389, middle half 162,283 to 187,920, median 162,782, mean 169,148, most 208,365 tokens over 5 runsPython off
A6Blender: LODs and export
Plain Python (Blender), A6: fewest 58,551, middle half 60,115 to 81,504, median 60,376, mean 90,157, most 190,240 tokens over 5 runsPython
MifBridge as shipped (Blender), A6: fewest 37,866, middle half 39,737 to 58,854, median 47,642, mean 53,313, most 82,468 tokens over 5 runsMifBridge
A7Blender: vertex-color masks
Plain Python (Blender), A7: fewest 14,964, middle half 20,687 to 27,639, median 20,867, mean 22,416, most 27,924 tokens over 5 runsPython
MifBridge as shipped (Blender), A7: fewest 18,135, middle half 18,551 to 25,361, median 19,081, mean 21,439, most 26,069 tokens over 5 runsMifBridge
A8Blender: organic: tree
Plain Python (Blender), A8: fewest 73,249, middle half 76,792 to 102,755, median 84,230, mean 89,845, most 112,197 tokens over 5 runsPython
MifBridge as shipped (Blender), A8: fewest 48,888, middle half 54,613 to 86,981, median 70,371, mean 69,803, most 88,160 tokens over 5 runsMifBridge
MifBridge, Python off (Blender), A8: fewest 304,989, middle half 314,624 to 323,939, median 320,289, mean 317,766, most 324,988 tokens over 5 runsPython off
X1Blender + Unreal: cross-tool: prop to Unreal
Plain Python (Blender + Unreal), X1: fewest 126,625, middle half 215,558 to 252,580, median 223,625, mean 244,118, most 402,201 tokens over 5 runsPython
MifBridge as shipped (Blender + Unreal), X1: fewest 88,875, middle half 89,124 to 110,673, median 108,841, mean 104,427, most 124,624 tokens over 5 runsMifBridge
MifBridge, Python off (Blender + Unreal), X1: fewest 154,888, middle half 283,145 to 1,607,327, median 374,200, mean 857,945, most 1,870,167 tokens over 5 runsPython off
X2Blender + Unreal: cross-tool: kit to Unreal
Plain Python (Blender + Unreal), X2: fewest 84,029, middle half 141,682 to 219,713, median 188,092, mean 230,249, most 517,728 tokens over 5 runsPython
MifBridge as shipped (Blender + Unreal), X2: fewest 217,842, middle half 244,337 to 672,423, median 278,562, mean 417,991, most 676,791 tokens over 5 runsMifBridge
Hover a row for its numbers; the table below has all of them.
Task by task
This round captured no pictures. The look checks only rule out clean primitives, so the page makes no claim about how any asset looks; round ten's page shows pictures of these same tasks.
A1 game-ready pass Blender
Make a messy imported crate game-ready: welded, fixed normals, applied transforms, base pivot, UVs, a padded lightmap and one tight hull.
Blender, pilot and wave 2. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 92,980, Python 96,315, 0.965x.
A2 modular kit Blender
Build four modular kit pieces on a 100 cm grid: corner pivots, tiling ends, world-scale UVs, lightmaps and hulls that keep the doorway open.
Blender, pilot and wave 2. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 40,584, Python 35,646, 1.139x.
A3 organic: rock Blender
Make one 180 x 120 x 90 cm boulder of 3k-6k triangles with a seated pivot, two UV maps and one tight convex hull.
Blender, pilot and wave 2. Right results: MifBridge 5/5, Python 3/5. Median run: MifBridge 64,213, Python 104,833, 0.613x. Python off: 5/5 right, median 181,464, 2.826x of as shipped.
A4 organic: rock variants Blender
Make three different boulders of one family sharing one material, each with its own pivot and hull.
Blender, pilot and wave 2. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 86,049, Python 177,089, 0.486x.
A5 organic: plants Blender
Model a grass tuft and a fern in real geometry (no alpha cards), with per-blade UVs and a wind weight in a color attribute.
Blender, pilot and wave 3. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 47,032, Python 41,396, 1.136x. Python off: 5/5 right, median 162,782, 3.461x of as shipped.
A6 LODs and export Blender
Make three LODs of a statue at 50, 25 and 12.5% of its triangles, grouped for FBX, with a hull, exported as one file.
Blender, pilot and wave 3. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 47,642, Python 60,376, 0.789x.
A7 vertex-color masks Blender
Write byte-exact masks into a color attribute: a value per loose part, an up-facing mask and a height ramp.
Blender, pilot and wave 3. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 19,081, Python 20,867, 0.914x.
A8 organic: tree Blender
Model a 5 m young broadleaf tree with 400+ real leaves, bark and leaf slots, a wind weight and a trunk-only hull.
Blender, pilot and wave 3. Right results: MifBridge 4/5, Python 4/5. Median run: MifBridge 70,371, Python 84,230, 0.835x. Python off: 5/5 right, median 320,289, 4.551x of as shipped.
X1 cross-tool: prop to Unreal Blender + Unreal
Deliver a boulder into Unreal as a saved static mesh: size and pivot, Nanite or LODs, hugging hulls, the material slot and lightmap index, placed.
Blender + Unreal, pilot and wave 4. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 108,841, Python 223,625, 0.487x. Python off: 4/5 right, median 374,200, 3.438x of as shipped.
X2 cross-tool: kit to Unreal Blender + Unreal
Deliver three kit pieces into Unreal as saved static meshes with corner pivots, simple collision that keeps the doorway open, slots and lightmap index.
Blender + Unreal, pilot and wave 4. Right results: MifBridge 4/5, Python 5/5. Median run: MifBridge 278,562, Python 188,092, 1.481x.
Every task, every outcome
Five runs an arm on each task, by wave. The four outcomes are counted apart, with runs that could not be graded in a column of their own. Tokens are per run.
Task
Arm
Outcomes
Median
Mean
Tokens per run, range
Against Python, median / mean
A1 game-ready passBlender
Plain Python (Blender)
5 verified
96,315
96,031
middle half81,149 to 98,081all runs78,084 to 126,524
MifBridge as shipped (Blender)
5 verified
92,980
94,254
middle half87,674 to 103,916all runs75,146 to 111,552
0.965x / 0.981x
A2 modular kitBlender
Plain Python (Blender)
5 verified
35,646
34,671
middle half35,126 to 36,145all runs28,132 to 38,305
MifBridge as shipped (Blender)
5 verified
40,584
41,063
middle half40,228 to 41,423all runs39,287 to 43,795
1.139x / 1.184x
A3 organic: rockBlender
Plain Python (Blender)
3 verified2 wrong
104,833
106,780
middle half104,788 to 105,486all runs82,614 to 136,178
MifBridge as shipped (Blender)
5 verified
64,213
78,601
middle half56,901 to 67,643all runs49,182 to 155,067
0.613x / 0.736x
MifBridge, Python off (Blender)
5 verified
181,464
200,725
middle half169,035 to 182,183all runs164,045 to 306,896
1.731x / 1.880x
A4 organic: rock variantsBlender
Plain Python (Blender)
5 verified
177,089
178,348
middle half170,755 to 192,418all runs104,628 to 246,851
MifBridge as shipped (Blender)
5 verified
86,049
134,330
middle half77,223 to 101,229all runs72,574 to 334,573
0.486x / 0.753x
A5 organic: plantsBlender
Plain Python (Blender)
5 verified
41,396
39,642
middle half35,746 to 43,574all runs33,133 to 44,359
MifBridge as shipped (Blender)
5 verified
47,032
44,927
middle half35,887 to 50,148all runs27,898 to 63,671
1.136x / 1.133x
MifBridge, Python off (Blender)
5 verified
162,782
169,148
middle half162,283 to 187,920all runs124,389 to 208,365
3.932x / 4.267x
A6 LODs and exportBlender
Plain Python (Blender)
5 verified
60,376
90,157
middle half60,115 to 81,504all runs58,551 to 190,240
MifBridge as shipped (Blender)
5 verified
47,642
53,313
middle half39,737 to 58,854all runs37,866 to 82,468
0.789x / 0.591x
A7 vertex-color masksBlender
Plain Python (Blender)
5 verified
20,867
22,416
middle half20,687 to 27,639all runs14,964 to 27,924
MifBridge as shipped (Blender)
5 verified
19,081
21,439
middle half18,551 to 25,361all runs18,135 to 26,069
0.914x / 0.956x
A8 organic: treeBlender
Plain Python (Blender)
4 verified1 wrong
84,230
89,845
middle half76,792 to 102,755all runs73,249 to 112,197
MifBridge as shipped (Blender)
4 verified1 wrong
70,371
69,803
middle half54,613 to 86,981all runs48,888 to 88,160
0.835x / 0.777x
MifBridge, Python off (Blender)
5 verified
320,289
317,766
middle half314,624 to 323,939all runs304,989 to 324,988
3.803x / 3.537x
X1 cross-tool: prop to UnrealBlender + Unreal
Plain Python (Blender + Unreal)
5 verified
223,625
244,118
middle half215,558 to 252,580all runs126,625 to 402,201
MifBridge as shipped (Blender + Unreal)
5 verified
108,841
104,427
middle half89,124 to 110,673all runs88,875 to 124,624
0.487x / 0.428x
MifBridge, Python off (Blender + Unreal)
4 verified1 wrong
374,200
857,945
middle half283,145 to 1,607,327all runs154,888 to 1,870,167
1.673x / 3.514x
X2 cross-tool: kit to UnrealBlender + Unreal
Plain Python (Blender + Unreal)
5 verified
188,092
230,249
middle half141,682 to 219,713all runs84,029 to 517,728
MifBridge as shipped (Blender + Unreal)
4 verified1 wrong
278,562
417,991
middle half244,337 to 672,423all runs217,842 to 676,791
1.481x / 1.815x
Verified
the scene is correct.
Wrong
it ran and the scene is not.
Refused
the tool declined (not a failure and not a success).
Crashed
the application stopped answering.
No measurement
the harness could not grade the run.
Who grades
A third process grades every run and talks to neither arm.
Tokens and dollars
Arm
Runs
Median run
Mean run
Tokens per run, range
Dollars, at API list prices
Plain Python
50
84,130
113,226
middle half41,941 to 140,306all runs14,964 to 517,728
$14.53
MifBridge as shipped
50
69,007
106,015
middle half42,016 to 99,167all runs18,135 to 676,791
$12.92
MifBridge, Python off A3, A5, A8 and X1 only
20
245,755
386,396
middle half167,788 to 321,202all runs124,389 to 1,870,167
$7.84
Tokens
Tokens are input + output + cache read + cache write, all counted at full weight, as the API reports them per run: the context processed, largely cache reads, not a bill.
Range
Middle half: where the typical runs fall, leaving out the cheapest quarter and the costliest quarter. All runs: the cheapest run to the costliest.
Dollars
Dollars are the round's tokens priced at the API's list prices for claude-opus-5-5 (input $4, output $20, cache read $0.20 and cache write $8 a million tokens): what an API-key user would pay. On a claude.ai plan the same runs draw on the plan's usage allowance instead of being billed.
What one request costs before any work
The tokens in the very first request of a session, before the model has done anything: the client’s own prompt plus whatever the tools add. Measured on and , before the waves of each day, on the builds named in each row (the round itself ran on 6cd630ca99837c4e64bb791dba319b1a8f7f63f5). With tool search on, the client loads a tool’s description only when the model asks for it.
Setup
Tool search
Tools offered
Client environment
First request, tokens
MifBridge (6cd630ca)
on
3
clean
3,971
Plain Python for Blender
off
1
clean
3,135
MifBridge (6cd630ca)
on
3
clean
3,976
Plain Python for Blender
off
1
clean
3,140
How this was measured
The method
The same model got the same task text and the same starting scene twice, once through each arm. Only the tools changed. Each run is one fresh session of Claude Code that ends when the model says it is done; then a separate process opens what the run saved and checks it.
The applications
A1 to A8 run in Blender; X1 and X2 build in Blender and deliver into Unreal Engine. Every arm drives the same application on a given task, and each label on this page names it: Plain Python (Blender) on A1 to A8 and Plain Python (Blender + Unreal) on X1 and X2, and the same for every other arm.
The arms
Plain Python: one tool that runs the model's Python inside the editor, exactly as written. It ran A1 to A8 in Blender; X1 and X2 in Blender + Unreal.MifBridge as shipped: MifBridge as a buyer installs it: its default tools, Claude Code's tool search, and its own Python tool left on as shipped. It ran A1 to A8 in Blender; X1 and X2 in Blender + Unreal.MifBridge, Python off: the same install with the Blender add-on's own 'Allow run_python' preference unticked, so Blender work goes through the typed tools (round ten's diagnostic arm; Unreal's Python stays reachable, MifBridge has no switch for it). It ran A3, A5 and A8 in Blender; X1 in Blender + Unreal.The build: MifBridge 1.3.0, the tagged release at commit 6cd630ca9 (2026-10-29). The Fab listing's recorded build when this was read (2026-11-01) was 1.2.0 (c2536bb8). Every MifBridge run in this round was checked to be on 6cd630ca9.
Model and client
claude-opus-5-5 through Claude Code 2.1.280, at the client’s default effort (medium), recorded for every run, started from an empty folder so no project files or settings reach the model.
Grading
A third process grades every run and talks to neither arm. For Blender it is a fresh headless Blender that opens the saved scene; for Unreal Engine, a separate headless editor that loads the saved level and assets; for X1 and X2, both: the Unreal editor grades the saved asset and exports it, and a fresh Blender grades that export. It runs the same checks for every arm, and every checker was first tested against a correct build and against deliberately broken ones. A tool reporting success is never taken as the evidence.
What counts as a loss
A task where MifBridge’s median run used more tokens than plain Python’s median run. It is listed as a loss whether or not either arm got the task right, so a task MifBridge got right and plain Python got wrong can still be a loss on tokens (none this round).
Tokens and dollars
Tokens are input + output + cache read + cache write, all counted at full weight, as the API reports them per run: the context processed, largely cache reads, not a bill. Dollars are those tokens priced at API list prices, what an API-key user would pay; a Claude plan user is not billed them.
When
to , 120 runs, in four waves; the table below lists them.
The pictures
This round captured no pictures. The look checks only rule out clean primitives, so the page makes no claim about how any asset looks; round ten's page shows pictures of these same tasks.
Deliver three kit pieces into Unreal as saved static meshes with corner pivots, simple collision that keeps the doorway open, slots and lightmap index.
Download the data (JSON): every number on this page, as MifBench’s exporter wrote it from the run records on 2026-11-02.
What this round cannot show
One model (claude-opus-5-5) through one client (Claude Code), at its default effort, on one machine.
UE 5.8.3 on X1 and X2, and Blender 5.2: a claim about Python on UE 5.3-5.7 needs a bench project on that engine, and 5.8 added Blueprint graph APIs the earlier engines lack.
Every task is a small, fully specified job: none asks the agent to find its way around a large existing scene it did not make.
Five runs a cell: cells are wide, so a per-task ratio is a rough guide.
Ratios to plain Python are read within one round only; plain Python's own cost swings between rounds.
Every round ran at the CLI's default effort (medium for claude-opus-5-5 on Claude Code 2.1.280; round nine records it per run). Rounds one to eight passed a desktop session's variables to the CLI, which adds ~208 tokens of client prompt to every request on every arm alike: measured against rounds six to eight's own probes; round nine removes them.
Whether the assets look good: the look checks only rule out clean primitives (a jittered smooth blob passes the rock checks), this round captured no pictures, and nothing here scores how anything looks.
Knowledge: every prompt states the conventions it grades (UCX_ names, fbx_type, the byte rule, padding), so the round measures carrying out a spec, not knowing the conventions unprompted.
Textures, baking and Unreal-side use (lighting, streaming, LOD switching, Nanite's look) are not measured.
The Python-off arm turns off Blender's Python only: MifBridge has no switch for Unreal's (run_python app='unreal' goes through exec_console), so on X1 that arm could still script Unreal.
Whether UE 5.8.3 or the build moved X1 and X2 against rounds ten and ten B: the round changes both.
Whether the change of login moved anything past the first request: the pilot wave is one run a cell.