Cheap models change the budget, while Opus owns the hard work
GPT 6.1 Soul changes the economics of agent loops, yet Opus still wins on complex autonomous work.
Sonnet 5.5 is a sub-agent for Opus, and its cache-read pricing is the trap.
The cost math flips. Both OpenAI and Anthropic moved the cost math, in opposite directions.
6.1 Soul's 5-cent cache reads make agent loops cheap, but it still fails on complex, unattended work.
Sonnet 5.5 can end up costing more than Opus for real agent tasks because cache reads aren't discounted proportionally.
Governance is the bottleneck. OpenAI's Dots demos are impressive but gloss over agent governance.
Missing: audit trails, agent-to-agent auth, rollback procedures, cost guardrails.
Treat the Dots demo as a vendor pitch; the Lean exploit is the evidence that the gap is real.
The reliability ceiling. Both 6.1 Soul and Sonnet 5.5 hit the same wall: they're unreliable for autonomous, multi-day tasks.
6.1 Soul burned tokens for four days on a Rust port, then Opus finished it in under a day.
Sonnet 5.5 degrades with bad reasoning settings and is best used as a sub-agent called by Opus.
My take
I think the smart budget play is to run both model families side by side instead of picking a winner.
6.1 Soul is the cheapest code reviewer you've ever hired, but I wouldn't let it run unattended on anything complex.
The Sonnet cache-read pricing is a trap your finance team will find on the first monthly bill.
The Dots governance gap matters, but the Lean exploit is the real warning: verification pipelines are now a security boundary.
Buy the savings, but audit everything twice.
The videos behind this brief
We watched these so you don't have to.
Video and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.