AI in production: tools, operating patterns, costs, and failure modes.
Opus 5.5 lowers the listed token prices. A small restart decision shows what to compare before moving the conversation to it: the exception, the client’s handling of earlier reasoning and the cost of an answer someone can use.
Plan an inference-service node drain with a scoped disruption budget, replacement-capacity checks and a practical maintenance worksheet.
NVIDIA shortened a Cosmos 3 Nano generation from about 156.6 seconds to 34.2 by using eight GPUs. That helps one request; a busy service must also account for the others.
SWE-Serve tests whether an AI-written patch works through a running model server. Its different verdicts on the same patches show what an output count can leave unanswered.
Antigravity's SDK can run local models, including Gemma through LiteRT. A repository audit shows why local computation, tool permission and a cloud planner need separate decisions.
What prompt-cache hits save once retries and failed work are included in task cost, with a local calculator and ledger.
NVIDIA's confidential-inference results show a modest serving cost on eight B200 GPUs. A replacement worker has another job first: show that it meets the service's trusted-state requirements.
NVIDIA Topograph gives schedulers a map of hardware connections. The useful result is a job that finishes sooner, including the time spent waiting for a better group.
Saving the unchanged beginning of a prompt can cut the bill for later updates. A summary that drops an important dependency can still leave the reviewer with more work.
Build repeatable cases and check recovery independently, with a runnable Python grader.
A power reduction can leave an AI evaluation intact but too late for the release decision. Follow the job through the pause, the restart and the agreement its owners need to make.
“Restart after the queue drains” is not the same as “restart approved.” Follow that small difference through Claude’s conversation summaries and Opus 5.5 replay checks.
