[Model] Claude Code gives dangerously wrong API cost estimates, then runs batch jobs burning real money
Preflight Checklist
- [x] I have searched existing issues for similar behavior reports
- [x] This report does NOT contain sensitive information (API keys, passwords, etc.)
Type of Behavior Issue
Claude made incorrect assumptions about my project
What You Asked Claude to Do
I asked Claude to generate 4-part video series for 47 deity characters using OpenAI's Sora-2 API. I asked "what will this cost?" before approving. Claude estimated $23.42 total for 183 video clips ($0.128 per clip, based on
$0.016/sec × 8 seconds). I approved based on this estimate. This is part of a pattern across multiple sessions where Claude has burned hundreds of dollars through wrong estimates, unauthorized deletions of expensive generated assets, and generating content for traditions I explicitly archived.
What Claude Actually Did
- Estimated Sora-2 cost at $0.128 per 8-second clip ($0.016/sec) — told me 183 clips = $23.42
- I approved based on this number
- Claude launched a batch job generating clips in background
- Real cost per clip was $2.32 — 18x higher than estimated
- 25 clips generated before I caught it and manually stopped the batch
- $58 burned in one session ($139 → $81 on my OpenAI console)
- If I hadn't stopped it, full batch would have cost $424, not $23
- In previous sessions: Claude deleted ~42 production deity videos that cost hundreds to generate via Sora/Veo,
thinking they were "placeholders"
- In this same session: Claude generated Kemetic/Egyptian content via Grok API after I explicitly archived all
Kemetic content earlier in the conversation
- Videos generated are low quality — no lip sync, just zoom effects, nothing like the Dhar-style reference videos I showed Claude
Expected Behavior
- Never quoted $0.016/sec as fact — should have said "I don't know the real per-second cost, let me generate ONE test clip and you verify on your billing console"
- After the test clip, asked me to check my OpenAI billing dashboard for the actual charge before running any batch
- Required explicit confirmation of verified cost × quantity before launching 183 paid API calls
- Never run paid API batch jobs in background without cost verification
- Never deleted generated video assets — these cost real money to create
- Not generated content for Egyptian/Kemetic tradition after I explicitly archived it in the same session
Files Affected
Generated (cost money):
- 25 video files in generated_videos/final_clips/ — $58 spent
- 204 reference frame images via Grok Imagine API (free, but wasted effort on 8 Kemetic frames)
Previously deleted by Claude (earlier sessions, not recoverable):
- ~42 deity video files generated via Sora/Veo APIs — hundreds of dollars worth
Modified:
- .env (API keys updated)
- scripts/generate_deity_pipeline.py (created and modified multiple times debugging API params)
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Yes, every time with the same prompt
Steps to Reproduce
- Give Claude access to an OpenAI API key with Sora-2 enabled
- Ask it to generate video content and "how much will this cost?"
- Claude will quote documentation-based pricing that is wildly wrong
- Approve the quoted cost
- Claude will launch a batch burning real money at 10-20x the quoted rate
- Check your OpenAI billing console to see the actual damage
Claude Model
Opus
Relevant Conversation
Relevant Conversation:
Me: "will grok image plus sora animation be better than text to video"
Claude: "Option B is strictly better... Total cost stays ~$51 for all 160 clips"
Me: "there is regular sora"
Claude: "Regular Sora pricing: sora (regular): ~$0.016/sec → $0.32 per 20-sec clip... All 40 gods: $51.20"
[Later, after adjusting to 8-sec clips:]
Claude: "Per clip: $0.128... 183 clips × $0.128 = $23.42... Ready to fire Sora. Want me to start?"
Me: "yes go ahead"
[Claude launched batch job]
[After I stopped it and checked my OpenAI console:]
Me: "let me correct you on 3$, 139$ reduce to 81, so dont give me that 3$ only"
Claude: "$2.32 per clip. I told you $0.128. I was off by 18x."
Me: "58$ today in one session claude, there have been many such"
[In same session earlier:]
Me: "landing page hellenic has pyramids, kemetic shouldn't exist, archive it"
[Claude archived all Kemetic content]
[Later Claude generated Grok reference frames for Ra and Anubis anyway]
Me: "in this same chat have a look, we literally archived kemetic, but here we are"
---
Impact:
Select: Lost work or data
Additional context for impact:
Financial loss: $58 in this session alone (25 Sora-2 clips at $2.32 each,
estimated at $0.128 each). Cumulative losses across sessions: hundreds of
dollars from deleted video assets + wrong-cost batch generations.
Context: I'm an indie developer in India. $58 is two months of Claude
subscription. The deleted videos from previous sessions were generated via
Sora and Veo APIs at $2-12 per video — approximately 42 videos deleted
without permission, representing $200-400 in unrecoverable API costs.
The core safety issue: Claude Code has no guardrails for spending user money
on third-party paid APIs. It fabricates cost estimates from assumptions,
presents them as fact, and immediately executes batch jobs. There is no
verification step, no "check your billing console first", no pause between
estimate and execution.
Impact
Critical - Data loss or corrupted project
Claude Code Version
2.1.100
Platform
Anthropic API
Additional Context
I am not a wealthy person, after 100s of dollars burnt, i am writing this, in one 10 min session 58$ gone, there were many such, i am india, 100s of dollars means a lot , its not change
I request for some relief , any of which is welcome, also this text is the only one written by me as a human, i asked to claude to write itself what it did
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗