@jerryjliu0 Same, I told everyone tokenmaxxing as much as you can (to explore and live on the edge) before the employer changes its mind. we need small, local and vertical models to be ready.

@nickbaumann_ Would be incredible if the Fable release pushed our team to codex 🙂

The model subsidies will eventually end and this workflow of “creating loops that will prompt your agents” will result in massive amounts of code that’s not well understood that you will have to pay lots of money to maintain in the future.

@jerryjliu0 Were you able to switch over to your codex plans when they ran out?

napkin math: a $350K engineer is roughly $175/hr, so 10 hours of their work equates to ~$1.75k. If Fable shipped >2x what that eng would've in those 10 hours, the spend argument can still be justified. The harder q is whether you can attribute $ to output, which is difficult to scale org wide. Token allocation will eventually follow what people ship

@jerryjliu0 Yah it’s powerful for small specific tasks but if you let it spawn a dozen agents it will incinerate tokens

@jerryjliu0 inevitable beginning of the end

Of subsidies

Codex crazy plans ended last week Claude -p moves to API in 2 days Models like fable not being on API after the 22nd

It's been fun but was bound to end eventually

@jerryjliu0 time to also improve the llama router 😅

Anthropic has already been preparing for that approach where Opus can be the “verifier” Model at the API layer. We can probably use sonnet with Fable as the verifier and get amazing results.

I’ve been having Fable orchestrate the workflow of sonnet and opus agents and have it verify the results. Way less “fable tokens”, the initial vibe is of great results. No science behind it yet.

@jerryjliu0 Federate your creds bro. Don't use the underwear account to test new stuff - set up a new account with different billing.

@jerryjliu0 you guys work on parsing pdfs so there's def AI/ML involved. you guys know that Fable is nerfed for AI/ML, right?

@jerryjliu0 yeah fair point - per-run attribution breaks under tokenmaxxing. Perhaps the math just moves up a level where team spend vs what the team shipped that quarter is what matters. Ceiling at that point might be how much agent output the team can actually absorb and act on vs the bill

@jerryjliu0 I used 86% of my weekly limits in 16 hours on 200$ license. No more tokenmaxxing

@jerryjliu0 imagine burning $150/hr on claude and still having to beg for more tokens. fable changed the math, guess api calls finally cheaper than peace of mind

@jerryjliu0 good idea to use fable as a planning agent->delegate to cheaper coding models

@jerryjliu0 This is at least partly inconsistent with the VC thesis on Anthropic investment.

@ahmedrsyed @jerryjliu0 Senior SWEs do more than code.

They set the technical direction, evaluate considerations that go beyond the codebase, and act a a reservoir of institutional knowledge.

AGI will be able to do all of that but we don't have AGI yet.

Dollar equivalence is a category error.

@jerryjliu0 The rate limit's the real constraint now, not cost.

Cheaper tokens mean you can push harder but hit the ceiling sooner.

For systems optimized for throughput, you've flipped which problem matters.

@jerryjliu0 If fable goes for api based billing and Gpt5.6 doesnt, do you think there will be a shift in dev market share?

@ahmedrsyed i think that's super hard to say, because tokenmaxxing == you're fine wasting tokens massively as long as SOME of it produces meaningful output relative to your bandwidth towards processing