OpenRouter Batch API for code review: limits, 402 errors and reasoning-token budgets

OpenRouter Batch API

• :batch model variants cost 50% of the synchronous price. completion_window accepts only 24h; in practice batches finished in minutes.
• There is no cancel endpoint. A running batch can only be deleted after it finishes, so it is billed.
• Reasoning-mandatory models (Claude Fable 5.1, Gemini 3.8 Flash) spend max_tokens on reasoning: at 4000 they returned no content (finish_reason: length). 32000 with reasoning.effort: medium was enough for a 4-file review; Gemini used about 25k tokens and Fable 3-6k. 65k gave Gemini nothing extra.
• Before accepting a batch, OpenRouter estimates cost from max_tokens and returns HTTP 402 if the estimate exceeds the account credit balance (the message says "available balance"). This is not the per-key daily cap, so raising the key limit does not help.
• Batch inputs are retained for 30 days, which matters when deciding what code may be sent.
• Trial result: GPT-6 Astra ($0.10) found a real backslash-URL SSRF that the others missed; Fable ($0.28) invented one finding; Gemini ($0.10) was solid. Two cheap models outperformed one expensive one.

Claude Haiku review subagents printed secrets

Four of nine Haiku review subagents printed .env secret values into the transcript (via cat, cut -d= -f1 and grep) despite a prompt rule not to. A follow-up warning only works if it arrives before the first command. Permission deny rules for reads of .env, *.pem, *key*.json and credential files are more reliable than prompt text. They should be scoped to the reviewed project paths: a Bash(grep *.env*) deny in a project's local settings also blocked an unrelated read of a user-level .env file the same session needed, and the rules had to be removed. Adding such rules only for the duration of a review run avoids that side effect.