AI price cuts and the trade-offs of going local

Max Woolf traces OpenAI and Anthropic price reductions over recent months, with cuts reaching 80% on specific offerings. His analysis examines how competition from open-weight models puts pressure on hosted pricing.

IBM’s CEO makes the case for on-prem open weights

The Register

The Register reports that IBM CEO Arvind Krishna sees open-weight models as good enough and cheaper for many enterprise tasks.

One game, two coding workflows: cheaper took longer

Reddit · GapNew4766

A builder reports completing the same game with Opus alone for $2.52 in 16 minutes. Using Opus for planning and local Qwen for coding cost $0.62 and took 42 minutes in this single reported comparison.

A Qwen setup for 8 GB of video memory

Reddit · BleynSpecnaz

A builder shares a Qwen 3.8 Flash Next configuration using Strata and memory-mapped experts. The reported local run uses 8 GB of video memory and 32 GB of system memory.

Gemini Managed Agents puts a credentials proxy outside the sandbox

AI Engineer · Jack Witherspoon

Jack Witherspoon demonstrates cloud-hosted agents with stateful sandboxes and a credentials proxy. The sandbox uses placeholder tokens; the proxy substitutes actual credentials for requests to matching external endpoints.

Builders discuss letting one session handle orchestration

Reddit · gcaussade

A Reddit discussion examines relying on the model to manage orchestration within a session, including reported changes in guidance about hand-designed subagents.

Greptile: review with a different model than the one that wrote the code

AI Engineer

The speaker recommends targeting bugs from your own outage history, retaining failing pull requests for evaluation, and retrieving context about architecture and authentication state. He claims that using a different model for review improves incident detection by about 10%.

Farol maps pull-request risk for the approver

GitHub · winnin

Farol presents a pull request’s risk map for the person approving the changes.

Task completion and code quality can diverge

Reddit · aj_marshall

A builder’s research project, inspired by SlopCodeBench, examines Opus 5.5 completing tasks while producing low-quality code.

Larger models did not absorb MCP faults better in this study

arXiv

The authors inject tool faults across three agent stacks and two model sizes. Under their permanent-heavy fault profile, the larger gpt-oss model retained no greater share of its fault-free performance; the result is specific to the tested models and conditions.

Identifying an agent’s model from its repair behavior

arXiv

The authors probe tool interfaces and model the sequence of repairs that follows. They report 96% model-attribution accuracy under adversarial identity spoofing on the evaluated model set.

Saving lessons from mistakes in small adapters

arXiv

Self-Direction distills lessons from failed tasks into lightweight adapters while keeping base-model weights frozen. The authors evaluate a persistent library of these lessons across research, tool-use, and terminal tasks.