ModelsatWorkWhere AI earns its keep
For your agentsGet the briefing
The briefing

October 14, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Orca Security says it cut an AI query agent's cost 90% by sending the model about 140,000 input tokens per workload instead of 2 million. Correction: our Gemini 4 Argon report wrongly said Google's launch post omitted the later price; its footnote lists $4/$20, and the pricing page dates it from 1 December 2026. We also separated Gemini app subscriptions from API access.

By the numbers

  • -90% Orca says input fell from 2 million to 140,000 tokens per workload and latency from 60 to 15 seconds; AWS published it.
  • 812 of 1,221 ShipStation says its Slack help desk agent resolved these in chat; 409 became tickets. Reply is not resolution.
  • 15 of 18 AWS's 72-run Grok 4.7 test: low effort used 1,184 output tokens a run; xhigh solved 18 of 18 on 9,147.
  • 30,000+ Anthropic says merges rose from about 2,500 a day in May 2026, and source control became its bottleneck.
Send the model less

Orca Security says a nine-stage pipeline cut its Bedrock agent's cost 90% and latency from 60 to 15 seconds

AWS published a customer post on 13 October 2026, co-written with Orca Security, about Discovery AI, an agent that turns plain-English questions into queries over Orca's cloud-security data. Orca says its first version loaded the whole schema into one prompt on the largest model, at about 2 million input tokens and 60 seconds per workload. The production version uses about 140,000 tokens and 15 seconds, and Orca says cost fell 90%.

Why it matters: If your CFO asks what an answer costs, Orca's savings came mostly from sending less, not from one cheaper model. Per the post: compact summaries of 900+ resource types replaced the full schema, five reranked examples replaced a large dump, and regex rules answer obvious patterns such as IP addresses and CVE IDs with no model call. Small models, Amazon Nova 2 Lite and Claude Haiku 4.5, handle screening and checking.

Zoom out: Orca calls example retrieval its biggest accuracy lever. It compared plain top-K search, which returned near-duplicates, with diversity-weighted search, which lost relevance when the top hit was right. Retrieving 20 examples and reranking to five won.

The catch: This is a vendor-published, single-company account. It gives no pass rate or size for the evaluation set, and the grader is itself a language model. The 90% is measured against Orca's own deliberately heavy prototype, with no dollars per query. It also gives no count of failed or blocked requests.

What to do: Pull a week of your busiest agent's logs and sort by input tokens per request. Take the biggest prompt, replace the pasted reference material with retrieval of the five best matches, and run both versions on 100 real questions with known answers. Compare pass rate and cost per request before you pay for a larger model.

Our coverage →AWS Machine Learning Blog: Natural language security queries at Orca Security using Amazon Bedrock (13 October 2026, vendor-published, co-written with Orca) ↗

A bot that stays quiet

AWS says ShipStation's Slack help desk agent kept 812 of 1,221 chats out of Jira

AWS published a customer post on 13 October 2026, co-written with ShipStation Global, on Deckhand, an IT help desk agent in Slack. ShipStation says that in its first three months from April 2026 Deckhand handled 1,935 messages and started 1,221 conversations. Only 409 became Jira tickets.

Why it matters: ShipStation says positive ratings rose from 82% in April to 96% in June after it revised the system prompt and knowledge base in response to complaints. The design is the lesson: ticket creation needs the employee to press a button, and the agent replies only to the original poster and stays silent once a human has answered.

The catch: The post gives no running cost, so the roughly 406 saved hours, ShipStation's own 30-minute estimate per ticket, cannot be set against spend. The 96% is a thumbs-up rate, not an audit of correct answers, and a reply is not a resolution.

What to do: Before launching a support bot, write its silence rules: which messages it ignores and when it steps back for a human. Each month, have a technician grade 50 chats that ended without a ticket for correctness.

Our coverage →AWS Machine Learning Blog: How ShipStation Global built an AI IT help desk on Amazon Bedrock AgentCore (13 October 2026, vendor-published, co-written with ShipStation) ↗

Low effort, a fraction of the tokens

AWS says Grok 4.7 at low effort solved 15 of 18 tasks on about one-eighth of xhigh's output tokens

AWS said on 13 October 2026 that it ran an agent on SpaceXAI's Grok 4.7 through Amazon Bedrock, three search-heavy tasks at four effort settings, six times each: 72 runs. Low effort got 15 of 18 right, medium and high 17, xhigh 18. Mean output was 1,184 tokens at low and 9,147 at xhigh.

Why it matters: Reasoning effort is a cost dial you set per request. AWS says low suits work where a slightly lower completion rate is acceptable or answers are reviewed downstream, and suggests running planning and verification steps high and mechanical steps low.

The catch: AWS says six runs per cell show a direction, not a confidence interval; one miss moves a figure by 5.6 points. The tasks are synthetic, the post reports tokens rather than dollars, and it tests one model.

What to do: Take your 20 hardest real tasks and run each five times at low and at high effort. Route a step to low only if its pass rate holds within what a reviewer would accept.

Our coverage →AWS Machine Learning Blog: Build long-horizon agents with Grok 4.7 on Amazon Bedrock (13 October 2026, vendor-published; AWS ran the evaluation) ↗

What else

  1. 01AWS and Anchor Browser say agents can pay HTTP 402 paywalls once one session flag is set; the flag defaults to off, and the post publishes no spending limits. Our coverage →
  2. 02Google says adults in select markets, including the US, can build private Android apps in AI Studio without a developer account; neither Google nor TechCrunch says whether work accounts or admins can control it. Our coverage →
  3. 03AWS says its Well-Architected Agent is in preview and reviews workloads across more than 65 services; it requires an AWS support plan, so check yours before assigning anyone to try it. Our coverage →

Stat of the day

Under 0.1%

Anthropic says its new classifier against distillation, the unauthorised training of a smaller model on Claude's outputs, blocks benign requests in fewer than 0.1% of cases. It also says that since 24 September it charges again for requests blocked before Claude responds. A false-positive rate that low still bills real traffic at volume. Anthropic measured this on its own platform and does not say how to appeal a block. Ask your account contact whether blocked requests appear separately on your invoice. Anthropic: Strengthening our defenses against distillation attacks (12 October 2026) ↗

Field note

Anthropic's engineers report in the Scion post that agents sometimes broke core system constraints: several systems that needed to know when a change had merged read the commit log linearly instead of asking the code review tool. They also report that adversarial review could be myopic, with Claude adding excessive validation after one over-defensive review. Engineers still review every change request, now for intent rather than line by line. Write your system's constraints where agents and reviewers can both read them.

One question

Which stage of your busiest AI agent could a cheaper model, or no model at all, take over?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedInXBlueskyHNEmail

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →