MifBridge against plain Python, run by run.

The short version

• Right results

  MifBridge got 46 of 50 runs right and plain Python 45 of 50. MifBridge got more runs right on A3 and A8 (A3: 4 of 5 against 3 of 5; A8: 5 of 5 against 4 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 stopped in 1 MifBridge run (X2 1 of 5, Blender + Unreal). A crashed run left nothing to grade, so it counts against its arm.

• The average

  Averaged task by task, MifBridge used 1.237x plain Python’s tokens: 121,348 against 98,111. Plain Python’s costliest run was on X1 (507,954 tokens); leave X1 out and the average is MifBridge 72,686 against 80,646, so it tips the other way.

• The typical run

  On 7 of the 10 tasks (A1, A2, A3, A5, A6, A8 and X1) MifBridge’s median run used more tokens than plain Python’s. It used fewer on A4, A7 and X2.

• MifBridge, Python off

  On the 4 tasks it ran (A3, A5, A8 and X1), it used 3.125x plain Python’s tokens, averaged task by task, and got 18 of 20 runs right. It called a typed asset tool in 18 of 20 runs, where MifBridge as shipped did in 6 of 50. The headline figures compare plain Python with MifBridge as shipped only.

• Time

  All 50 MifBridge runs took 4,827 s of model time and all 50 plain Python runs 4,239 s. Task by task, MifBridge’s median run took longer on 6 of the 10 (A1, A3, A5, A7, A8 and X1).

• The worst case

  MifBridge’s costliest run used 876,512 tokens (X1); plain Python’s used 507,954 (X1).

• Dollars

  Priced at API list prices, all 50 MifBridge runs came to $15.33 and all 50 plain Python runs to $13.29. On a Claude plan these runs use the plan’s allowance; nobody is billed these amounts.

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.

Its median run used more tokens than MifBridge as shipped on 3 of 4 tasks, and more than plain Python on 4 of 4; above 1 means Python off used more.

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_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 44 of its 50 runs the model wrote Python and nothing else. With Python off, 18 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 71 times the cheapest. Each task names the application it drives.

Task by task: what each arm built

One run from each arm, side by side, each labeled with the application it shows: a verified run whose tokens sit nearest its cell’s median, or, where an arm had no verified run, its most common outcome, labeled as such. Blender results are rendered from the scene the run saved, every arm from the same camera and light. X1 and X2 work in both applications, so each arm has two pictures there: the Blender scene the run made, and what it delivered to Unreal, put back into the editor from the run’s own saved content and photographed in the run’s own level, lit for the photograph only. Nothing is drawn that a run did not produce. Select a picture to open it full size.

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 full waves. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 86,577, Python 69,544, 1.245x.

Blender

Blender

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 full waves. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 38,518, Python 36,662, 1.051x.

Blender

Blender

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 full waves. Right results: MifBridge 4/5, Python 3/5. Median run: MifBridge 98,732, Python 95,290, 1.036x. Python off: 5/5 right, median 287,287, 2.910x of as shipped.

Blender

Blender

Blender

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 full waves. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 103,023, Python 132,040, 0.780x.

Blender

Blender

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 full waves. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 57,116, Python 36,731, 1.555x. Python off: 5/5 right, median 280,954, 4.919x of as shipped.

Blender

Blender

Blender

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 full waves. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 53,886, Python 50,859, 1.060x.

Blender

Blender

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 full waves. Right results: MifBridge 5/5, Python 5/5. Median run: MifBridge 19,490, Python 21,070, 0.925x.

Blender

Blender

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 full waves. Right results: MifBridge 5/5, Python 4/5. Median run: MifBridge 85,907, Python 81,005, 1.061x. Python off: 5/5 right, median 266,867, 3.106x of as shipped.

Blender

Blender

Blender

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 full waves. Right results: MifBridge 3/5, Python 3/5. Median run: MifBridge 664,950, Python 201,973, 3.292x. Python off: 3/5 right, median 453,426, 0.682x of as shipped.

In Blender: the scene each arm made

Blender

Blender

Blender

In Unreal: what each arm delivered

Unreal

Unreal

Unreal

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 full waves. Right results: MifBridge 4/5, Python 5/5. Median run: MifBridge 111,614, Python 180,118, 0.620x.

In Blender: the scene each arm made

Blender

Blender

In Unreal: what each arm delivered

Unreal

Unreal

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.

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, time and dollars

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.

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's development version at commit 2a026a58 (2026-10-01), 55 commits after the 1.1.0 release and before 1.2. It is not the build on sale: that is 1.1.0 (4684a30f), which round nine measured. Every MifBridge run in this round was checked to be on 2a026a58.

Model and client

claude-opus-5-5 through Claude Code 2.1.280, at the client’s default effort (medium), recorded for every run that finished (a run stopped at the client’s time limit records neither), 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 (A3 and A8).

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

2026-10-01, 120 runs, in two waves; the table below lists them.

The pictures

24 renders of saved Blender scenes (EEVEE, one camera and light per task for every arm, with any staging stated under the picture) and 5 photographs of what a run delivered to Unreal, put back into the editor from the run’s own saved content. Every task has its pictures from each arm.

The waves

The tasks

• A1Blender

  Make a messy imported crate game-ready: welded, fixed normals, applied transforms, base pivot, UVs, a padded lightmap and one tight hull.

• A2Blender

  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.

• A3Blender

  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.

• A4Blender

  Make three different boulders of one family sharing one material, each with its own pivot and hull.

• A5Blender

  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.

• A6Blender

  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.

• A7Blender

  Write byte-exact masks into a color attribute: a value per loose part, an up-facing mask and a height ramp.

• A8Blender

  Model a 5 m young broadleaf tree with 400+ real leaves, bark and leaf slots, a wind weight and a trunk-only hull.

• X1Blender + 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.

• X2Blender + 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.

Download the data (JSON): every number on this page, as MifBench’s exporter wrote it from the run records on 2026-10-01.

What this round cannot show

• One model (claude-opus-5-5) through one client (Claude Code), at its default effort, on one machine.

• UE 5.8.2 and Blender 5.2 only: 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.

• Three to 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 proxies only rule out clean primitives (a jittered smooth blob passes the rock proxies); the rendered pictures are for a person to judge, and nothing here scores them.

• 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.