Topic
The compute, latency and money required to serve each request to a large model, and the engineering and pricing decisions that follow from it.
The new Strands harness ships the tools, memory and context handling most teams assemble themselves, and AWS attributes its 28% saving to offloading bulky tool output to files and reusable caches across six unnamed benchmarks.
Reality
Sol lists at $2 and $10 per million tokens and Luna at $0.10 and $0.50. The per-task savings OpenAI published come mostly from the lower price, and the cheaper tier scores below its predecessor on computer use.
Reality
BDH-CQ updates a fixed-size internal memory instead of generating chain-of-thought tokens, and its authors put a single puzzle query at about $0.00070. Two outside researchers say the result does not yet credit the design.
Reality
A PyTorch case study credits Shopify's GraphQL agent with beating frontier-model quality at 96 percent lower serving cost. Most of what it documents is how the reward signal behind that gets defined and checked.
Reality
A Microsoft case study credits Docusign's swap to small task-specific models with 90% lower cost and eight times the throughput. Docusign's own accounting of the whole pipeline claims 50 times cheaper per document.
Reality
An engineer measured TypeSafe AI's Jev decision model on 10 public datasets and published the raw results. The two largest movements came from decomposing and staging the questions, with the model build held constant.
Reality
Jev returns typed answers with probability scores and no free-form text, and TypeSafe prices a decision at four hundredths of a cent. Its accuracy figures are scored against other models' probabilities.
Reality
The model writing each page's recommendations only ever saw the candidates the code handed it, and when similarity matching found nothing that list was the most-rated games in the catalog. A probability model now scores the candidates first.
Reality
Nvidia's Adel el Hallak told The New Stack that logs and outputs are not enough to diagnose an agent failure. The industry's shared answer so far is SAFE, a findings exchange backed by roughly 140 companies.
Reality
The test built three tokenizers that differed only in their language mix, and the halfway split cost English about 1% while a fully balanced five-language corpus made English 17% more expensive.
Reality
Researchers held a coding harness's execution loop fixed and varied planning, tools and context management across four models, and found that what each component is worth depends on how strong the model already is.
Reality
The number came out of one harness on OpenRouter at temperature 0.9 with a 64,000-token output cap, and it holds for a bounded translation task while the same model sits mid-table on open-ended terminal work.
Reality
Vals AI's 141-hour Minecraft run ended with OpenAI's newest model farming potatoes for hours after losing its end-game loot and its spawn point in one explosion, and the recovery policy it came away with was a note to self.
Reality
Retrieve-for-Train runs reinforcement learning once against a fixed corpus and distils the winning sub-query sets into a lightweight diffusion model. Adopting it puts a retrain schedule on whoever owns the catalogue.
Reality
The Kotlin Benchmark grades agents on 105 verified repository tasks. Its token column shows setups that solve within a few tasks of each other burning between 66,000 and 777,000 tokens per fix.
Reality
A retired production prompt measured 56,000 tokens on every call, ninety percent of it accumulated bans. Its author says the rule-per-failure loop looked like progress and produced committee prose.
Reality
Anthropic says preliminary second-quarter revenue passed $11.5B and adjusted operating income was $559M. That is a 4.9% margin, booked in the same quarter its models got 15 cents a dollar cheaper to run.
Reality
A proposed scaling law for distillation says the winner flips on two things: whether you already own a teacher, and how many students you intend to serve.
Reality
Neither the compute saved nor the share of reasoning that stopped being words has been published, which leaves customers policing agents with a control whose substrate OpenAI says it will describe later.
Reality
The model string keeps working after the cutover, so a pinned request comes back from V4.1-Flash. By the figures in the dev.to write-up, that model scores 90.6 on Terminal-Bench 2.1 and 42.3 on SimpleQA, against V4-Pro's 55.2.
Reality