gpt-image-1-mini incorrect price estimation in UI

[xinpehr]

Description

The model selector displays 3 credits as the estimated price per image for gpt-image-1-mini, but the actual cost deducted is ~10 credits. This causes confusion and a poor user experience.

Root Cause

The estimation in ImagineView.php only calculates output token cost:

// src/Presentation/RequestHandlers/App/ImagineView.php (line 111-114)
$model['estimation'] = $this->calc->estimate(
    $model['key'],
    CostCalculator::OUTPUT  // only output tokens — misses image_input tokens!
)->value;

Estimation formula:

multiplier (1875) × output_rate (0.0016) = 3 credits  ← shown in UI

However, OpenAI charges for 3 token types, which are all summed in ImageService.php:

// src/Ai/Infrastructure/Services/OpenAi/ImageService.php (line 154-157)
$tc = $this->calc->calculate($resp->usage->input_tokens_details->text_tokens,  $model, CostCalculator::INPUT);
$ic = $this->calc->calculate($resp->usage->input_tokens_details->image_tokens,  $model, CostCalculator::IMAGE);
$oc = $this->calc->calculate($resp->usage->output_tokens,                       $model, CostCalculator::OUTPUT);
$cost = new CreditCount($tc->value + $ic->value + $oc->value);

Actual cost breakdown for a real generation (1536×1024):

Token type	Tokens	Rate	Cost
Text input (prompt)	~500	0.0004	~0.20 credits
Image input	~13,480	0.0005	~6.74 credits
Output	~1,875	0.0016	~3.00 credits
Total			~9.94 credits

The image_input_tokens (~13,480) returned by the OpenAI API (usage.input_tokens_details.image_tokens) represent internal generation compute and are never accounted for in the estimation.

Fix Applied

Updated multiplier for gpt-image-1-mini from 1875 → 6250 in:

• config/registry.json
• config/registry.base.json

- "multiplier": 1875,
+ "multiplier": 6250,

New estimation: 6250 × 0.0016 = 10 credits ✓

Affected Models

Model	Status	Was affected
gpt-image-1-mini	✅ Active	✅ Fixed
gpt-image-1.5	✅ Active	No (already had multiplier 6250)
gpt-image-1	❌ Disabled	—
dall-e-3, fal.ai, Stable Diffusion, etc.	—	No (flat-rate pricing)

Long-term Recommendation

The estimation logic should be refactored to separately estimate all three token types (INPUT + IMAGE + OUTPUT) using dedicated multipliers per token type in the model registry, rather than relying on a single multiplier value applied only to the output rate.