Developments that matter in production
What changed, how it works and what it means for the people running reliable systems.
46 stories · 25–36
How we reportThe last answer may be the cheapest part of a repair. Sonnet 5.5’s cache prices and API changes matter most when you follow the attempts all the way to a usable patch.
Background reading: AI in productionA replacement agent can send fresh measurements that an old alert never sees. Google’s maintenance deadline makes this a good time to follow one disk name all the way into the query.
Background reading: ObservabilityAn access ticket cannot reach the access team if the router offers only database, network and application. CLM-8B makes supplied choices comparable, not complete.
Background reading: AI in productionStorage Intelligence can spot unusual reads and preview a bulk change. The missing detail may be simple: a new job needs that old dataset every morning.
Background reading: Platform engineeringA suspected flaw gives a developer a question. Google's internal PageBreak project tries to hand over a reproducible failure, so the next step can be a repair.
Background reading: AI in productionAn assistant looking for affected jobs should not make customers wait longer. Google's preview separates its query compute, but the first question after idle time still deserves a trial.
Background reading: Platform engineeringAn order lookup can fail in one account while its explanation sits in another. Omni brings agent evaluations and telemetry together, but the investigation still needs the right records or a clear handoff.
Background reading: ObservabilityA passing test inside one rack says little about a job spanning two. NVIDIA's controller organizes workload checks so operators can connect the result to the nodes, paths and targets that matter.
Background reading: Platform engineeringA host-update package can be correct and still reach too many machines. Follow one proposed kernel-setting change from the selected worker to the check that tells the team what happened.
Background reading: Platform engineeringOpus 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.
Background reading: AI in productionNVIDIA 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.
Background reading: Platform engineeringSWE-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.
Background reading: AI in productionDevelopments since our original reporting, with dated updates and supporting sources.
New model documentation makes retained instructions and thinking-block compatibility part of the compaction migration review.
Original report: Claude compaction: keep the reason the worker is waiting