// AI Platform Drift
DeepSeek V4 Pro 0813 completion pricing dropped 58%, output limit cut in half
OpenRouter model catalogue · published 2026-09-30 12:06 UTC
CORRECTED 2026-10-02 17:58 UTC — [CORRECTION · headline] Original: “completion pricing dropped 58%, output limit cut in half” Fact: Completion price decreases 43.3%; output capacity decreases 58.3%, more than half.
[CORRECTION · WHAT CHANGED] Original: “completion tokens from $943,718/M to $393,216/M (-58.3%)” Fact: Completion pricing falls from $3.49/M to $1.98/M (-43.3%). Those large integers are output-token limits.
[CORRECTION · WHY IT MATTERS] Original: “Workloads generating long outputs cost 58% less per request regardless of prompt size” Fact: Prompt pricing increases 67.4% while completion pricing decreases 43.3%; net cost depends on token mix. At ten completions per prompt, cost falls 42.0%.
[CORRECTION · triage.signals] Original: “pricing.prompt decreased 83% (0.0000003942 → 0.00000066)” Fact: Prompt pricing increases 67.4%, not decreases 83%.
[CORRECTION · triage.signals] Original: “pricing.prompt decreased 83% (0.0000003942 → 0.00000066)” Fact: The prompt-change magnitude is 67.4%.
Stood 53h 52m · cause extraction error · every correction ↗
Captured arithmetic only; vendor rate validity, applicability and effective conditions remain unverified.
WHAT CHANGEDOpenRouter's deepseek/deepseek-v4-pro-0813 model changed pricing: completion tokens from $943,718/M to $393,216/M (-58.3%), prompt tokens from $0.3942/M to $0.6600/M (+67.4%). Maximum completion tokens reduced from 943,718 to 393,216 tokens.
WHY IT MATTERSWorkloads generating long outputs cost 58% less per request regardless of prompt size, but any job currently producing more than 393,216 tokens in a single completion will now be truncated mid-response.
WHAT TO DOCheck your application's max_tokens parameter and actual completion lengths in logs. If you request or receive completions longer than 393,216 tokens, either split the work into multiple calls or switch models before the next deployment.
Model-reported confidence 95% (not measured accuracy)
Primary source ↗
written by anthropic/claude-sonnet-4.5
Show the reasoning — raw diff, triage verdict, confidence
RAW DIFF
id: deepseek/deepseek-v4-pro-0813 name: DeepSeek: DeepSeek V4 Pro 0813 -pricing.prompt: 0.0000003942 -pricing.completion: 0.00000349 +pricing.prompt: 0.00000066 +pricing.completion: 0.00000198 context_length: 1048576 -top_provider.max_completion_tokens: 943718 +top_provider.max_completion_tokens: 393216
HOW IT WAS JUDGED
Triage signals below are model-reported assertions, preserved verbatim; they are not independent arithmetic verification.
| triage verdict | MATERIAL at 0.98, needed 0.5 |
|---|---|
| triage reasoning | Pricing dropped significantly and output token limit cut by more than half. Teams budgeting or optimizing workloads need to know. |
| signals | pricing.prompt decreased 83% (0.0000003942 → 0.00000066) · pricing.completion decreased 43% (0.00000349 → 0.00000198) · max_completion_tokens reduced 58% (943718 → 393216) |
| suppression | — |
| triage model | anthropic/claude-haiku-4.5 ($0.00000) |
| writer model | anthropic/claude-sonnet-4.5 ($0.00000) |