Paying Attention #7: Sonnet 5 Ships, the Fable Advisor Pattern, and GPT-5.6 Closes In
Happy 4th of July weekend everyone! For my international readers—July 4th is America’s Independence Day, and this year was the 250th.
It was a real scorcher in Atlanta. That’s typical for us on the 4th, but this year it was extra humid. My wife and I were meant to run the Peachtree Road Race for the third year in a row. It’s the largest annual 10K in the world. We had to bail this year because the kids were sick. We’ll be back at it next year.
We had a few surprises this week in the world of AI. Here’s what I paid attention to.
1. Sonnet 5 Ships
https://x.com/daniel_mac8/status/2072036584635977819
What happened
Anthropic released Sonnet 5 on Tuesday. On the benchmarks it’s a step down from Opus 4.8 almost across the board, and much of the early reaction wrote it off for exactly that reason. The post above made the counter-case: pair it with Claude Code Dynamic Workflows—set /model to Sonnet 5, set /effort to Ultracode, and any complex task kicks off a dynamic workflow.
Why it held my attention
People were pretty down on the Sonnet 5 release. I get it. But a pattern I’ve been experimenting with a bunch—and what makes Sonnet 5 relevant for me—is this: how can you give a smaller, cheaper model to a larger, more expensive, more capable model as a tool?
What I’m carrying forward
The second item in today’s post gets deeper into this, but the best way to use a model like Sonnet 5 is as an implementer orchestrated by a larger model. This is a winning pattern, especially as mid-sized models keep getting more capable. I’m not sure it will always be necessary, but it’s a winner while it is.
2. The Fable Advisor Pattern
https://x.com/daniel_mac8/status/2073366129737683146
What happened
Fable 5 came back on Wednesday, as predicted in last week’s edition. Its return made a pattern I’d been sketching all the more relevant: Fable as the advisor doing the high-leverage judgment, Sonnet or Opus as the implementer doing the token-heavy work. By Friday I’d distilled it into the best way to use Fable right now, and on Saturday I shipped it as a working artifact: fable-advisor, a free, open-source plugin that runs exactly this orchestrator pattern (that’s the post above). Underneath it all, the cost tension is still live—Fable is expensive enough that I’ve been token-maxxing my Claude sub ahead of July 7th.
Why it held my attention
Fable is the most capable AI model ever created. It’s also the most expensive. You don’t need to blow your token budget by using it for every LLM call. Use it as an advisor and distill its expensive intelligence for smaller models, in real time.
What I’m carrying forward
I do wonder whether this will be an effective pattern forever. I suspect it won’t be, and that eventually there will be a single unified system that does it all for you, without any engineering. But for now I’ll keep looking for ways to optimize my token usage, and you should too.
3. GPT-5.6 Closes In
https://x.com/daniel_mac8/status/2072646722283622591
What happened
Details on GPT-5.6 Sol surfaced Thursday: 60% of Fable’s cost ($30 vs. $50 per 1M tokens), a reported win over Fable on TerminalBench 2.1, and availability through the ChatGPT subscription. The Information reported that OpenAI found a way to cut inference costs by more than half, which is what makes that pricing possible. Then the timing sharpened: Leo says OpenAI plans to release GPT-5.6 on Tuesday, July 7—the same day Fable access is set to leave Claude subscriptions—even as Anthropic signals that Fable is coming to Claude subs once compute capacity allows. There may also be a Sol Ultra tier aimed squarely at Fable.
Why it held my attention
It will be interesting to see what Anthropic does based on how strong GPT-5.6 is. On one hand, Fable is expensive and isn’t guaranteed to be available on consumer subscriptions after July 7th. On the other, many people and organizations simply want to use the best model. They are cost insensitive.
What I’m carrying forward
If I’m right about July 7th being the release date for GPT-5.6, I can’t wait to see what Anthropic does about Fable access that day. My hope is that they extend it to Claude subs, but I’m not that confident they will.