Good morning, and welcome back to Techzip AI Newsletter.

Today’s Techzip covers Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Sol and Luna as frontier AI gets cheaper, Qualcomm’s push to put agentic AI on devices, Epoch AI’s falling cost of thought, and DeepSeek’s bet on Huawei chips for future model training.

Anthropic has launched Claude Opus 5.5, which the company says reaches Claude Fable 5.1-level performance on most work while costing 40% less to run than Opus 5. Anthropic says the model is available across major clouds, includes preserved-thinking and compliance safeguards, and is built for longer sessions and coding, research and computer-use tasks that require multiple rounds of work. Anthropic’s ClaudeDevs account also says Opus 5.5 is about 30% faster per task and raises the five-hour Claude Code session limit by 20%.

In Claude’s playbook, Anthropic focuses on how to use the model for long-running work: hand over the full task and define “done,” set clear stopping rules, track progress in a file, split large audits across subagents and check the evidence before accepting the output. Opus 5.5 also brings clearer status reporting and lower cache-read pricing.

OpenAI has launched GPT-6 Sol and GPT-6 Luna, two faster, cheaper models built from GPT-6 Astra advances. Sol is priced at $2 per million input tokens and $10 per million output tokens; Luna is priced at $0.10 and $0.50. OpenAI positions Sol for tasks that need stronger reasoning, while Luna is aimed at high-volume, low-latency workloads.

In results published by OpenAI, GPT-6 Sol at reasoning xhigh scores higher than Claude Opus 5 on AutomationBench at 91% lower cost per task. GPT-6 Luna at high improves 5.4 percentage points over its predecessor while cutting cost per task by 58%. OpenAI also highlights improvements in factuality, coding, computer use, collaboration style and prompt caching, which allows reasoning or tool controls to change without reprocessing the entire cached context.

Sol and Luna turn frontier-derived capability into a cheaper routing layer for repetitive tasks and agent workloads. At scale, the cost of each model call becomes an architecture decision, not just a line item on a pricing page.

Here's what's happening with the Groundfloor Preferred Note: while the market moves on headlines, this rate doesn't move once you commit. Lock in a 7% fixed annualized return over a defined 6-month term, paid out monthly — income that isn't tied to earnings surprises, rate cuts, or the next correction. Groundfloor has a 100% repayment track record on Notes since launching them in 2018, and has paid out over $38M in Notes interest alone. New investors get a $100 bonus when they get started.

Invest Now

Anthropic’s official Opus 5.5 playbook makes one useful change to how you prompt the model: stop telling it to “think carefully.” Give it a complete task, define what finished looks like and tell it exactly when to come back to you.

Anthropic’s example is a migration: move every payment endpoint to a new client, delete the old client and pass the test suite; ask only if a test fails for a reason the model cannot explain.

Stanford finds that AI is already widely used across organizations, while agent deployment remains early. The report puts the gap between adoption and autonomous deployment in one data-backed frame.

This paper proposes measuring how often agents mention a brand inside shopping workflows through a concept similar to share of search. It expands the SEO question from ranking on results pages to appearing in an agent’s answer.

Simon Willison puts Opus 5.5, Sol and Luna against tasks he actually uses, rather than looking only at benchmarks. That lens puts model price next to the ability to complete a real workflow.

The paper proposes linked Markdown wikis as a long-term memory layer that agents can retrieve from and update over time. It replaces chat history with a document structure that people can read, edit and inspect.