// Intelligence Brief · Week 39
7 items that need action
2026-09-21 → 2026-09-27 UTC · 7 items published
ACTION NEEDED
VERIFY_ONLY
WHAT CHANGEDOpenRouter's DeepSeek V4 Flash (deepseek/deepseek-v4-flash) reduced max_completion_tokens from 384,000 to 131,072 tokens. Completion pricing dropped 65.9% ($384,000/M to $131,072/M tokens) and prompt pricing dropped 0.3% ($0.047/M to $0.0469/M tokens).
WHY IT MATTERSApplications generating outputs longer than 131,072 tokens will now fail or truncate. The pricing drop is substantial but does not offset the hard limit reduction if your workload depends on the previous 384k completion capacity.
WHAT TO DOGrep your codebase for 'deepseek-v4-flash' or 'deepseek/deepseek-v4-flash' and check max_tokens or equivalent parameters in calls. If any exceed 131,072, either chunk the generation into multiple requests or switch to a model with higher completion limits.
Full item and evidence ↗ · Primary source ↗
ACTION NEEDED
BUDGET
WHAT CHANGEDOpenRouter changed pricing for deepseek/deepseek-v4-pro-0813: prompt tokens decreased from $0.000000264 to $0.00000024502 per token (7% drop), completion tokens increased from $0.000000792 to $0.0000035 per token (342% increase). Maximum completion tokens increased from 384,000 to 943,718.
WHY IT MATTERSCompletion tokens cost 4.4x more than before. Applications generating long outputs will see material cost increases: a 100k token completion that cost $0.08 now costs $0.35. The higher output limit enables longer generations but at the new rate.
WHAT TO DOGrep your codebase for 'deepseek-v4-pro-0813' or 'deepseek/deepseek-v4-pro-0813'. For each usage, calculate current monthly completion token volume and multiply by 4.4 to estimate new cost. Update budget forecasts and alert whoever approves AI spend.
Full item and evidence ↗ · Primary source ↗
ACTION NEEDED
VERIFY_ONLY
WHAT CHANGEDOpenRouter reduced max_completion_tokens for deepseek/deepseek-v4-flash-vision-exp from 943,718 to 262,144 tokens (72% reduction). Prompt pricing decreased from $0.00000022 to $0.0000002156 per token; completion pricing decreased from $0.00000066 to $0.0000006468 per token.
WHY IT MATTERSApplications requesting or expecting output longer than 262K tokens will now be truncated or rejected. The pricing changes are negligible (2% decrease), but the token limit reduction breaks any workflow that relied on the previous 943K ceiling for long-form generation.
WHAT TO DOGrep your codebase for 'deepseek-v4-flash-vision-exp' or 'deepseek/deepseek-v4-flash-vision-exp' and check max_tokens parameters in requests. If any exceed 262,144, either chunk the output into multiple calls or switch to a model with a higher completion limit.
Full item and evidence ↗ · Primary source ↗
ACTION NEEDED
BUDGET
WHAT CHANGEDOpenRouter pricing for deepseek/deepseek-v4.1-flash changed from $0.000000099 to $0.0000003 per prompt token (3x increase) and $0.0000006 to $0.0000012 per completion token (2x increase). No effective date stated.
WHY IT MATTERSIf you are calling this model through OpenRouter, your inference costs will increase by 2-3x on the next invoice. A workload processing 1B tokens monthly goes from ~$660 to ~$1500.
WHAT TO DOCalculate your current monthly token volume for this model, multiply prompt tokens by 3x and completion tokens by 2x to get the new cost, and update your budget forecast. If the increase is material, evaluate switching to a different model or provider before the next billing cycle.
Full item and evidence ↗ · Primary source ↗
ACTION NEEDED
MIGRATE
WHAT CHANGEDOpenRouter's deepseek/deepseek-flash-latest model pricing dropped from $0.000000075 to $0.000000035 per prompt token (53% reduction) and $0.0000003 to $0.00000029 per completion token (3% reduction). The max_completion_tokens limit decreased from 943,718 to 384,000 tokens.
WHY IT MATTERSExisting requests generating more than 384k completion tokens will now fail or truncate. Cost projections for this model are no longer accurate—workloads will cost roughly half as much for prompts but may require splitting long completions across multiple requests.
WHAT TO DOAudit any code calling this model: check if you request or rely on completions exceeding 384k tokens, and if so, implement chunking or switch models. Update cost forecasts to reflect the new prompt pricing.
Full item and evidence ↗ · Primary source ↗
ACTION NEEDED
BUDGET
WHAT CHANGEDOpenRouter changed pricing for deepseek/deepseek-pro-latest: prompt tokens decreased from $0.0000002528 to $0.00000025 per token (1% drop), completion tokens increased from $0.0000019584 to $0.0000035 per token (79% increase).
WHY IT MATTERSCompletion tokens are the bulk of cost in most workloads. A 79% increase on completions means total cost per request will rise substantially for any usage pattern that generates more than trivial output.
WHAT TO DOPull last 30 days of usage for this model, calculate new cost at the updated completion rate, and update your budget forecast. If cost increase is material, evaluate whether to pin to a specific non-latest DeepSeek model or switch providers.
Full item and evidence ↗ · Primary source ↗
ACTION NEEDED
BUDGET
WHAT CHANGEDEffective now: prompt pricing decreased from $0.0000003498 to $0.00000026 per token (26% reduction). Completion pricing increased from $0.0000010494 to $0.0000035 per token (233% increase). Maximum completion tokens increased from 393,216 to 943,718 tokens (140% increase).
WHY IT MATTERSApplications generating long outputs will see significantly higher costs despite cheaper prompts. A 100K token completion now costs $0.35 instead of $0.10. The higher output limit enables longer generations but at the new rate.
WHAT TO DOCalculate actual cost impact using your prompt-to-completion ratio from logs. If completions dominate your token usage, reforecast monthly spend and update budgets. Consider capping max_tokens in requests if the new limit exceeds your use case.
Full item and evidence ↗ · Primary source ↗
What the week cost
84runs
$1.3163model cost
27sources watched
7corrections
Printed because they are checkable. 7 items were published this week and later
found wrong — every correction, with its cause and how long it stood.
RADAR watches 27 sources every six hours and writes up only
what changed materially. Get it weekly or daily by email —
or take the feed, which is free and has
terms that say so.