Token Cost: CFA Bench
How many tokens your AI reads for real Unreal Engine tasks with CodeFizz Editor Agent: the CFA Bench results, compact text against raw JSON, and the slimmer default replies in plugin 2.21.0.
Loading…
How many tokens your AI reads for real Unreal Engine tasks with CodeFizz Editor Agent: the CFA Bench results, compact text against raw JSON, and the slimmer default replies in plugin 2.21.0.
Every token your AI reads costs money and fills its context. We measure it with CFA Bench: a fixed set of real tasks on a real Unreal project, each done two ways. The compact way is what the bundled AI skill teaches. The raw way reads the same facts as full JSON.
Read an asset as compact text first. Export JSON only when you need exact values.
Compact text with every node and value:
cfa read_asset --asset /Game/Blueprints/BP_CharacterExact JSON, only when you need it:
cfa export_asset --asset /Game/Blueprints/BP_CharacterTokens the AI reads for each task. Lower is better.
| Task | Compact way | Raw JSON way | Smaller by |
|---|---|---|---|
| Understand a character Blueprint | 2,952 | 102,566 | 34.7x |
| Understand an Animation Blueprint and its state machine | 3,746 | 73,006 | 19.5x |
| Import an animation FBX onto an existing skeleton | 351 | 3,765 | 10.7x |
| Read an input mapping context | 452 | 3,205 | 7.1x |
| Read a locomotion blend space | 550 | 2,020 | 3.7x |
| Read a montage: slots, sections, notifies | 143 | 327 | 2.3x |
| Find the pistol assets in a project | 343 | 678 | 2.0x |
| All seven tasks | 8,537 | 185,567 | 21.7x |
Plugin 2.21.0 cuts the heaviest read commands. The same command on the same asset, before and after:
| Command | Before | After | What changed |
|---|---|---|---|
| export_asset | 102,501 | 1,861 | A large export now points at a short summary file first. |
| get_blueprint_variable_details | 4,142 | 735 | The list shows only values that differ from the defaults. --full shows all. |
| get_actors_in_level | 3,708 | 2,717 | New --filter and --max-results, a total, and no default rotation or scale. |
| get_blend_space | 2,445 | 2,020 | No default rate scale or mirror on each sample. |
| get_input_mapping_context | 1,912 | 1,812 | Triggers and modifiers show only changed properties. |
export_graph stays large on purpose. It is the input format for build_graph, diff_graph and apply_graph_patch. For a large graph, its reply now says that read_asset --graph is about ten times smaller for reading.
A token is four characters, the usual rule for English and JSON. We count each reply, plus any file the reply tells the AI to open. Both ways use the same rule, so the ratios are fair. The run used Unreal Engine 5.8.3 with plugin 2.21.0 and cfa 3.22.0. Each task took under a second.