[mirror] fix(block_cost_config): audit + correct stale LLM/block rates + migrate generic ReplicateModelBlock to COST_USD

[yashwant86]

Mirror of upstream Significant-Gravitas#12912 for benchmark. Do not merge.

---

Summary by MergeMonkey

• Docs Updates:
  • Updated docstrings and comments for Replicate block cost tracking and LLM/block rate audits.
• Fresh Additions:
  • Replicate block now emits provider_cost via predict_time metrics for accurate per-second billing instead of flat RUN charge.
• Fixes & Patches:
  • Corrected stale LLM token rates (Grok, DeepSeek, Mistral, Kimi, Perplexity) to match current pricing.
  • Fixed Unreal Speech block to bill per-character USD instead of flat 5 credits.
  • Adjusted FAL video generator rate from 3 to 15 credits/second to match actual pricing.
  • Normalized COST_USD block margins (Jina, ZeroBounce, Unreal) to consistent 150 cr/$ baseline.
• Maintenance:
  • Updated .secrets.baseline line numbers and timestamp.

[bot-mergemonkey]

⚡ Risk Assessment — CRITICAL · ~45 min review

Focus areas: Replicate block status/output validation order · LLM rate accuracy (DeepSeek, Kimi, Mistral, Grok) · COST_USD margin normalization (150 cr/$ baseline justification) · FAL video 5× rate increase verification

Assessment: Refactors billing logic for Replicate block and audits/corrects LLM+block rates across 10+ models.

Walkthrough

User calls ReplicateModelBlock.run_model() with a model reference and inputs. The block now parses the reference to determine if it's version-pinned (contains ':') and calls predictions.async_create with either version= or model= keyword. After awaiting async_wait(), it checks prediction.status for 'failed' or 'canceled' and raises if found. If metrics.predict_time exists and is non-zero, it emits provider_cost = predict_time * $0.0014/sec as cost_usd via merge_stats. Finally, it extracts and returns the output. The COST_USD resolver then bills ceil(provider_cost * 150) credits.

Changes

Files	Summary
Replicate Block Cost Tracking Refactor autogpt_platform/backend/backend/blocks/replicate/replicate_block.py, replicate_block_cost_test.py	Migrated ReplicateModelBlock from flat RUN billing to dynamic COST_USD via predictions.async_create + predict_time metrics. Emits provider_cost = predict_time * $0.0014/sec; handles version-pinned vs unpinned refs, status validation, and gracefully skips billing when metrics unavailable.
Unreal Speech Block Cost Migration autogpt_platform/backend/backend/blocks/text_to_speech_block.py	Changed Unreal Speech billing from flat 5 credits to per-character USD ($0.000016/char). Block now emits provider_cost = len(text) * 0.000016 with cost_usd type for proportional billing.
Block Cost Configuration Audit autogpt_platform/backend/backend/data/block_cost_config.py	Audited and corrected stale LLM rates (Grok, DeepSeek, Mistral, Kimi, Perplexity). Normalized COST_USD block margins to 150 cr/$ baseline (Jina 100→150, ZeroBounce 250→150, Unreal 5→150). Updated FAL video rate from 3 to 15 credits/second. Added detailed pricing comments for transparency.
Test Updates for Cost Changes autogpt_platform/backend/backend/blocks/block_cost_tracking_test.py autogpt_platform/backend/backend/data/block_cost_config_test.py autogpt_platform/backend/backend/executor/block_usage_cost_test.py autogpt_platform/backend/backend/copilot/tools/helpers_test.py	Updated test assertions to reflect new cost models: Unreal Speech character-based USD billing, FAL video 15 cr/s rate, ZeroBounce 150 cr/$ margin, and Replicate provider_cost emissions.
Secrets Baseline Maintenance .secrets.baseline	Updated line number reference and timestamp due to code additions in replicate_block.py.

Sequence Diagram

sequenceDiagram
    participant User
    participant ReplicateBlock
    participant ReplicateClient
    participant Prediction
    participant StatsResolver
    User->>ReplicateBlock: run_model(model_ref, inputs, api_key)
    ReplicateBlock->>ReplicateBlock: Parse model_ref for ':'
    alt version-pinned
        ReplicateBlock->>ReplicateClient: predictions.async_create(version=...)
    else unpinned
        ReplicateBlock->>ReplicateClient: predictions.async_create(model=...)
    end
    ReplicateClient-->>Prediction: return Prediction object
    ReplicateBlock->>Prediction: async_wait()
    Prediction-->>ReplicateBlock: metrics populated
    ReplicateBlock->>ReplicateBlock: Check status
    alt status == 'failed' or 'canceled'
        ReplicateBlock-->>User: raise RuntimeError
    else status == 'succeeded'
        alt metrics.predict_time exists and > 0
            ReplicateBlock->>ReplicateBlock: merge_stats(provider_cost=predict_time*0.0014, cost_usd)
            ReplicateBlock->>StatsResolver: emit NodeExecutionStats
        end
        ReplicateBlock->>ReplicateBlock: extract_result(prediction.output)
        ReplicateBlock-->>User: return result
    end

Loading

Dig Deeper With Commands

• /review <file-path> <function-optional>
• /chat <file-path> "<question>"
• /roast <file-path>

Runs only when explicitly triggered.