August 24: ChatGPT Ads heads to Europe, AWS trims RAG context, and two more AI updates

The August 24 briefing covers the strongest verified AI launches and practical releases announced or newly reported from August 17–23, 2026. Marketers get a new ChatGPT ad surface, developers get a way to trim RAG context before expensive inference, and AI startups get two different models for specialized work: one optimized for high-volume agent steps and one trained to reproduce research.

What shipped: OpenAI announced on August 18 that ChatGPT Ads would expand the following week to 31 European countries, including Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands, and Austria. The expansion follows a six-month U.S. test and is OpenAI's largest market expansion for the ad product so far.1

Initial access goes through OpenAI's Ads Solutions team, agency partners, and technology partners. OpenAI says self-service access through Ads Manager will follow later this summer. Ads appear only for people on Free and Go plans; Plus, Pro, and Enterprise remain ad-free.1

OpenAI has added conversion optimization, geo-targeting, custom audiences, the OpenAI Pixel, the Conversions API, and third-party measurement integrations. The company says tens of thousands of marketers have advertised on ChatGPT, while ads remain labeled, separate from answers, and outside advertisers' access to conversations.1

Why it matters: ChatGPT is becoming an ad surface for people who describe a goal, compare options, and explain their constraints in one conversation. Marketers will need to test whether a product belongs in that decision context, rather than copy a keyword campaign into a new placement.

First test: Use one conversion goal and one European market. Compare qualified actions against a search or social campaign with the same offer, geography, and measurement window before expanding access.

What shipped: AWS published a query-aware compression pattern for Amazon Bedrock on August 21. After a retriever returns its top-k chunks, a smaller model such as Claude Haiku runs inside an AWS Lambda function and extracts the verbatim spans relevant to the user's question. A primary model such as Claude Sonnet then answers from the shorter context through the Bedrock Converse API.2

In AWS's benchmark, which used more than 500,000 documents and 500 questions, compression reduced cost to 67% of the baseline and sent 12% as many tokens to the primary model. The rerank-plus-compression version reduced cost to 64% and tokens to 10%. AWS measured a 19% latency increase for compression and a 12% increase for rerank plus compression; answer quality stayed at 97.5% and 97.6% of the baseline composite score.2

The trade-off is an extra model call. AWS's pattern needs a Lambda function, IAM permissions, and Bedrock access for both the smaller and primary models. AWS also says teams should test the pattern on their own queries, because the benchmark used one corpus, one domain, and one query distribution.2

Why it matters: Teams can keep a high-recall retriever while sending less irrelevant context to the costly answer model. The pattern fits documentation, support, compliance, and internal-operations assistants that retrieve long documents for narrow questions.

First test: Run the baseline and compressed paths against the same retrieved chunks behind a feature flag. Measure cost, end-to-end latency, correctness, completeness, citation accuracy, and faithfulness before changing the production path.

What shipped: AWS made NVIDIA Nemotron 3.5 Lightning available in Amazon SageMaker JumpStart for high-volume agentic workloads. The open model uses a hybrid mixture-of-experts design with 30 billion total parameters and 3 billion active parameters, a context window of up to 1 million tokens, text input and output, and DFlash speculative decoding. NVIDIA and AWS report up to four times higher throughput and up to 30% faster task completion for the workloads they evaluated.3

Developers can search for the model in SageMaker JumpStart, choose a supported GPU instance, and deploy an endpoint. An AWS account, scoped SageMaker permissions, and GPU service quota are prerequisites. SageMaker charges for the running endpoint, and AWS says the JumpStart model card does not expose customization even though the open model can be post-trained with NVIDIA NeMo.3

Why it matters: An agent does not need a frontier model for every step. Classification, field extraction, alert enrichment, policy checks, and record matching can move to a smaller specialized model while a larger model handles planning or difficult decisions.

First test: Route one repetitive, low-risk agent step to Nemotron 3.5 Lightning. Record accuracy, throughput, GPU cost, and failure cases beside the frontier-model baseline before widening the route.

What shipped: TechCrunch reported on August 22 that Inherent had released Faraday, an AI agent trained to reproduce findings from published scientific papers. Inherent's first-party research page describes Faraday as a 27-billion-parameter agent trained with long-horizon reinforcement learning and coding agents as tools.45

The training suite, called Replica, contains 310 tasks drawn from 100 machine-learning and AI-for-science papers. Each task asks an agent to reproduce a figure without seeing the original plot and within a limited time and compute budget. Inherent reports that Faraday outperformed Claude Opus 4.8 and GPT-5.5 baselines on its paper-replication evaluations.5

The public material is a research write-up and paper link, with the method and benchmark exposed more clearly than a user-facing product workflow. Faraday is therefore most useful to AI startups as a design reference: train a model to decide which experiments to run, then let a coding agent carry out the work under a measured budget.

Why it matters: The release tests a different division of labor from a chat assistant. A relatively small model supplies scientific judgment while a larger coding agent executes experiments, which gives builders a concrete pattern for training an agent to supervise tools instead of answering one prompt at a time.

First test: Take one reproducible internal analysis and remove the expected chart from the agent's input. Give the agent a fixed compute budget, require it to record its hypotheses and experiments, and score the final reproduction against a held-out reference.