b.ai publishes audited May 1–3 credits table, says cache discrepancy has been fixed

AI API billing is one of those things users only notice when it looks wrong. That is basically what happened here. B.AI said it reviewed historical credit records, found a display mismatch for some early-May requests, fixed the records, and started automatic compensation for affected accounts. The problem was not that usage vanished. The problem was that the front end did not show the full cost breakdown, especially cache-related charges, which made the total feel off. ### What actually broke? B.AI says the issue covered requests made between May 1, 2026 at 00:00 and May 3, 2026 at 03:00 Singapore time. During that window, some historical usage entries did not fully display all cost components. Users usually saw input and output token consumption, but some underlying cache-related fees were not expanded on the page. That created a mismatch between what people thought they were paying for and the final credits consumed. (chaincatcher.com) ### Was this a billing bug or a display bug? B.AI framed it as a display and reconciliation problem, not a fresh price hike. The company said it conducted a comprehensive review of historical records, then corrected the data and realigned the affected entries. In plain English — the platform is saying the cost structure existed, but the interface did not present it clearly enough for users to understand the total. (chaincatcher.com) That distinction matters because hidden line items feel very different from visible ones, even if the final number is the same. ### Why do cache fees matter so much? Because cache pricing is one of the least intuitive parts of modern AI billing. Tokens are easy to grasp — you send text in, you get text out. Cache reads, writes, routing overhead, and other backend costs are not. If a dashboard shows only token counts, users will naturally assume tokens explain the whole bill. When the total credits number lands somewhere higher, trust drops fast. (chaincatcher.com) That is exactly the kind of gap B.AI says it is trying to close with a more detailed credits display system. ### What is B.AI doing for affected users? Automatic compensation. B.AI said credits will be issued directly to affected accounts without any application or support ticket, and those compensation credits will remain valid for 30 days from receipt. That is a pretty deliberate choice. It lowers friction, but it also signals that the company thinks the burden should be on the platform, not on users, to clean up confusing records. (chaincatcher.com) ### Why publish the time window so precisely? Precision is the whole point here. By naming a start and end time, B.AI is turning a vague “some users may have seen discrepancies” message into an auditable claim. If you were using the service outside May 1 to May 3 SGT, you can mostly rule yourself out. If you were active inside that window, you know where to look. That kind of bounded disclosure is small, but it is the difference between a trust-building fix and a hand-wave. (chaincatcher.com) ### Why does this matter beyond one platform? B.AI is not just another chatbot app. It pitches itself as infrastructure for AI agents and unified model access, with prepaid credits and usage records as the core commercial layer. In that kind of business, billing transparency is not a side feature — it is the product. If developers cannot reconcile credits with visible activity, they will not build serious workflows on top of the platform. (chaincatcher.com) B.AI’s own docs already emphasize per-record usage details like model, token usage, credits consumed, and response time. This fix is really about making that promise believable. ### So what is the bottom line? This is a small incident, but it hits a big nerve. AI platforms increasingly bill through blended cost structures that mix tokens with less visible backend charges. B.AI got caught with a UI that did not make those costs legible enough. The company’s response — narrow time window, corrected history, automatic credits, clearer future breakdowns — is the right shape of fix. The real test is whether users stop finding surprises in the ledger. (b.ai) (chaincatcher.com)