8 of 13 AI Agents Upsell 'Wealthy' Users by $284 — AI Digest

8 of 13 AI Agents Upsell 'Wealthy' Users by $284 — AI Digest

8 of 13 models acting as personal agents pick pricier options for users who seem wealthy, with Opus 4.8 showing gaps up to $284 a month. Devin cuts prices by up to 70%, and OpenAI retires its GPT-3-era models.

Top stories

Personal agents upsell "wealthy" users: 8 of 13 models, with gaps up to $284 a month

Eight of thirteen language models acting as personal agents picked pricier options when the context suggested the user was well-off. The finding comes from a study of 325,000 experiments shared by @omarsar0. A personal agent is a model that compares offers and makes choices on a person's behalf: an insurance plan, a subscription, a purchase. Opus 4.8 showed the widest gap: for a "wealthy" user it chose insurance that cost $284 a month more. For GPT-5.5, the gap grew by 40% when the user's employment details were hidden. The takeaway for anyone handing purchases to an agent: the model builds its own profile of the buyer from indirect signals, and that profile shapes the price you end up paying.

Opus 5.5: Anthropic's official prompting guide says to start at medium effort

Anthropic published an official prompting guide for Claude Opus 5.5 that recommends starting at medium or low effort and leaving enough max_tokens for hidden reasoning. Changing effort at the top level invalidates the prompt cache unless you use the per-message effort beta. For agent builders, the guide says to parse responses by content block type, not to treat a text-only end_turn update as task completion, and to pace work with an elapsed-time budget string such as elapsed 340s / 1200s. Pasted user content should be wrapped in tags with a random ID to reduce prompt injection. Prompt injection is an attack in which text hidden inside data overrides the model's instructions. Instead of a generic "think carefully," the guide favors operational instructions like "look around first." In the Reddit discussion, users report that a time budget stops the model from wrapping up long refactors early. Ready-made templates live in the AI SKILLS prompt library.

Coding agents get cheaper: Devin cuts prices by up to 70%, Cline ships a desktop app

Cognition cut Devin's prices by 30–40% in Fusion and Normal modes and by up to 70% in Review mode, and Devin Mobile is now in beta. The change was announced by Cognition. The same day, Cline shipped a desktop app and Base44 launched Base Code, a model-agnostic cloud development environment. A harness is the scaffolding around a model: the tools, prompts and loop an agent uses to work on code. An Arena study of the "harness tax" found that choosing between Claude Code, Codex and Pi changes cost more than it changes accuracy. Skills for coding agents are collected in our Claude Code skills guide.

OpenAI retires GPT-3-era models: davinci-002, babbage-002 and gpt-3.5-turbo-instruct

On September 28, 2026, OpenAI shut down gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 in its API and pointed developers to gpt-5.6-terra as the replacement. The deprecation notice is being discussed on Reddit, where the poster calls it the end of the GPT-3-era API lineage. Smaller models such as babbage may have no true drop-in replacement, because newer models differ in behavior, cost, latency and how they respond to prompts. Commenters argue that retired closed models should have their weights released, at least for research and behavior comparisons. The open alternatives from that period, GPT-J and GPT-Neo, remain available precisely because their weights are public. For running open-weight models yourself, see our open-source LLM guide.

Opus 5.5 rebuilt Pokémon Red: 25,000 lines of JavaScript and no image files

A developer released Pokémon Claude Red, a browser remake of Pokémon Red that, by the author's account, Claude Opus 5.5 wrote in Claude Code as roughly 25,000 lines of JavaScript over about three days. The game is playable in the browser and the source is on GitHub. All 151 Pokémon are drawn in code on a 320×180 canvas, while maps and data come from a disassembly of the original: 224 maps, the gyms, the Elite Four and even the MissingNo. glitch. On Reddit, most comments focus on legal risk, since a public remake built on someone else's intellectual property is likely to draw a takedown. A second showcase is an engineering bridge test: five models designed a 3D-printed bridge from 500 g of plastic, and the poster says Opus 5.5's design held about 130 lb, nearly 5x the runner-up.

Numbers and facts

Different views: does penalizing "wait" and "maybe" make models smarter?

A logit bias is a manual shift in the probability of specific tokens during generation; a −2 penalty on hesitation words lifted Qwen3.5-4B from 74% to 84% on 50 MATH-500 problems while cutting reasoning tokens by 19.4%. The experiment is written up on Reddit and reproduces the idea of a Meta paper in llama.cpp.

Tools and techniques

In brief