Paying Attention #8: The Fable Endgame, the GPT-5.6 Sol-stice, and the Multi-Vendor Orchestra
Happy Sunday everyone! (Well, Monday. Sorry I’m a day late. ☺️)
The reason I’m a day late is because I went out on Lake Allatoona yesterday with my in-laws and 3-year old son. My 3 year old watched his two older cousins ride on a tube. After he saw that, he was not taking no for an answer and was going to ride that tube too. I was not prepared to ride a tube with him off the back of the boat but we both had so much fun.
Also, been trying to persuade my lawyer wife to let me download ChatGPT Work onto her laptop. It’s pretty great.
Last week was a huge one. We had important moves from OpenAI, Anthropic, Meta and SpaceXAI. Let’s discuss what I paid attention to.
1. The Fable Endgame
What happened
I recommend that Claude users max out their Fable usage on Monday because Fable was slated to leave the sub on July 7th. Turned out not to be the best advice, because a day later Fable access was extended for all subscriptions until July 12th. Then, on July 12th, Fable access was extended again until July 19th.
Why it held my attention
I am curious about Anthropic’s end-game with Fable. My perception is that people believe they have an underlying motive. I’m not so sure about that. My feeling is that they are legitimately trying to serve the best model they can, at the best price, with the compute they have available.
What I’m carrying forward
A take I had that went pretty viral on X is that they will remove Fable from the subscriptions once Opus 5 is ready. I still think that’s likely.
A lot can happen in a week, and the new end-date for Fable is July 19th so Opus 5 could be ready by next week. The timing is quite realistic.
2. The Summer GPT-5.6 Sol-stice
What happened
GPT-5.6 was a long time coming. We knew it was ready, but the US government was holding it back for security testing. This time OpenAI announced the release date a couple days ahead of time. The hype ran all week. That’s the fun part, isn’t it?
And on Thursday, July 8th, GPT-5.6 was released in three variants—Sol, Terra and Luna.
It tops the Artificial Analysis Agentic Coding index, and it cheaper than Claude Fable at $30 per 1mil output tokens compared to $50.
OpenAI was also released along with GPT-5.6 some pretty major updates to their desktop apps with ChatGPT Work. It’s similar to Codex but designed for knowledge workers.
Why it held my attention
I was keen on seeing if GPT-5.6 could stand up to Fable. It’s definitely nearing the ballpark, but after a few days of use Fable is still superior in terms of raw capability. That’s no surprise to me.
The area where GPT-5.6 Sol competes with Anthropic’s models is on cost and efficiency.
That, and OpenAI is living up to their name in making GPT-5.6 much more accessible for things like ML/AI development.
What I’m carrying forward
I don’t think GPT-5.6 is OpenAI’s answer to Fable. That will be GPT-6 and the rumors are already starting to leak about a new model with a brand new pretrain. The group chats I’m in say it will be ~10T param models and be the true response to Fable. View this in light of the fact that “The Information” reported OpenAI discovered a way to more than halve inference costs.
3. The Multi-Vendor Orchestra (Grok 4.5 + fable-advisor)
What happened
SpaceXAI released Grok 4.5 on Wednesday. It was trained in conjunction with the Cursor team using Cursor data. It’s surprisingly good, nearing Opus-class model benchmarks at a fraction of the price.
Why it held my attention
I created a Claude Code Plugin that uses Fable as the orchestrator model and combo of GPT-5.5, Opus 4.8 and Sonnet 5 as implementers. It’s not using GPT-5.6 and Grok 4.5 in place of GPT-5.5 and Sonnet 5, respectively.
What I’m carrying forward
It was a surprise that the SpaceXAI + Cursor tandem produced a model of this quality, at this cost, this soon. I used it and it’s quite good. SpaceXAI with the talent of the Cursor team is definitely back in the game.