AI News Briefing — Google launches Gemini Decision 1 to score questions in bulk

Model releases

• [2026-11-11] Google launches the Gemini Decisions API with gemini-decision-1, a closed model that answers instead of writing: it scores up to 1,000 yes/no, multiple-choice or rubric questions against one shared input in a single call, each with a confidence between 0 and 1. Inputs can be text, images, audio, video or PDFs up to 256K tokens. Input costs $0.10 per million tokens, billed once per request however many questions ride on it, and output costs $0. Google quotes 150–250 ms end to end for inputs up to 16K tokens, nearly flat from 1 to 100 questions. The post doesn’t say whether the model is in preview or generally available. (official, source)

  Classification, routing and eval-grading jobs that now loop one generation call per label can collapse into one call, and the bill in bulk depends only on the size of the shared input. Simon Willison’s quick test found it returns softer probabilities than rival decision models, which matters if your code thresholds at 0.5.

  For ML / Data Engineers: An eval pass that grades each transcript against five rubric criteria as five generation calls fits in one score request per transcript, and the 0-to-1 confidence gives a cut-off for sending borderline grades to a human reviewer.

Self-hosting & compute

• [2026-11-10] NVIDIA open-sourced Snapshot, which restores a checkpointed inference replica in seconds instead of reloading and warming the model. Crusoe reports GLM-5.3 going from about eight minutes to 31 seconds before it can serve, and Saturn Cloud cut vLLM startup from 2–8.5 minutes to 6–34 seconds. Both figures come from customers quoted by NVIDIA, not independent testing. (official)

  Long cold starts are why teams keep spare replicas warm. Seconds instead of minutes makes scaling on demand realistic.

  For Platform / DevOps Engineers: Autoscaling thresholds set around a multi-minute model load can be revisited once a replica restores in seconds. A single vLLM deployment on Kubernetes is the safe first trial, since vLLM is the only engine NVIDIA’s post names.

Coding agents

• [2026-11-11] AWS documented how to run OpenAI Codex against Amazon Bedrock: Codex stays on the developer’s machine, and its model calls go to Bedrock under short-lived IAM Identity Center credentials instead of an OpenAI key. The examples use GPT-5.6 Sol in us-east-2, set through model_provider in ~/.codex/config.toml, and spend shows up in the AWS Cost and Usage Report. AWS recommends log_user_prompt = false in the telemetry settings. (official)

  For a company that already buys through AWS, this puts Codex usage under existing IAM and billing rather than a separate OpenAI contract. Model availability still varies by Region, so check yours before a pilot.

  For Security Engineers: Prompts stay out of Codex telemetry only when log_user_prompt = false is set, as AWS recommends; ship it in the same ~/.codex/config.toml that sets the Bedrock model_provider, so it isn’t left to each developer.

MCP

• [2026-11-11] The MCP specification moves to a fixed six-month release schedule, normally the last Tuesday of March and September; the next version is 2027-03-30. Each cycle has a soft freeze 2.5 months before release, Tier 1 SDK betas 1.5 months before, and a hard freeze a month out. A proposal that misses the freeze moves to the next release rather than delaying the spec. Existing clients and servers need no changes. (official)

  Server authors now get dated SDK betas to test against roughly six weeks before each spec version lands.

Agent frameworks & platforms

• [2026-11-10] NVIDIA donated OpenShell, its runtime for enforcing policy on what agents do with tools, inference services and workloads, to the CNCF as a Sandbox project. It also joined Google’s Agent Substrate, which suspends idle agent environments on Kubernetes, and is integrating the two. (official)

  Under the CNCF, OpenShell’s roadmap is no longer NVIDIA’s alone to set, though Sandbox is the foundation’s earliest stage.

• [2026-11-11] Bedrock AgentCore added a Web Search API, InvokeWebSearch, served from an index Amazon runs itself, so queries aren’t sent to a third-party search engine. It takes domain and date filters and returns 1–25 results; CloudTrail logs the call but not the search text. Calls are signed with ordinary AWS credentials, with no separate API key. The post gives no price. (official)

  Agents already on AgentCore get search without another vendor contract or key to store. Cost planning waits on a price.

• [2026-11-11] Notion says Fireworks AI storage reached in the October breach held Notion AI chat transcripts from classification jobs behind its Insights feature, affecting fewer than 1% of workspaces, and tells customers to treat that data as exposed. Fireworks’ own update, dated November 9, says its zero-retention inference was untouched; Fireworks’ retention setting is Enterprise-only and off by default, and neither company has said which feature stored the transcripts. (source, official)

  A provider’s “zero retention by default” can cover inference only. Batch, dataset and training features may still store what you send, so check which ones your pipeline actually uses.

Practice & craft

• [2026-11-11] AWS open-sourced UAEF under Apache 2.0, an agent-evaluation framework that turns LangGraph, LangChain, Strands and AgentCore traces into one format and scores them on 36 metrics. Its LLM judges treat the agent’s output as untrusted: the rubric is sent separately, content is escaped, and a score is rejected unless the judge quotes a verbatim span of what it graded. (official)

  The verbatim-quote rule is the part worth copying: a judge that has to point at evidence is harder for injected text to steer.

• [2026-11-11] Thoughtworks’ Matteo Vaccari argues for writing customer-level tests as data, in tables or a small DSL, kept apart from test code, read-only to agents and checked in CI. The DSL makes assertions mandatory so an agent can’t write a test that checks nothing. The case rests on interviews with 26 developers who didn’t trust tests an agent had refactored along with the code, not on measurements. (source)

  An agent that edits the code and its test in one change can make both wrong together and still pass CI; tests it can’t touch remove that path.

• [2026-11-11] Arctic Wolf linked AntiFox, an open-source multi-agent pentesting framework on GitHub, to intrusions that started from a compromised IIS web app or a valid AWS credential of unknown origin. With the credential, most of dozens of service calls were denied, but one read a production secret from Secrets Manager. (source)

  The agent tried dozens of services and got in where the key’s permissions allowed. A credential scoped to its job would have turned that one success into another denial.

Watch list

• Google AI Plus: Gemini Pro models leave the plan from November 13, with dates set per account.
• Gemini 4 Carbon introductory rate: $2 input and $10 output ends November 30.