Intelligence Is Starting to Be About Wealth

The day after Fable 5 dropped, I threw it at a research project that had been stuck for two weeks.

Inference engine performance optimization. A 78-layer model running on hardware, baseline 10 token/s, target 100+ tps. GPT 5.5 had already sunk a lot of hours into it without much luck.

It ran for 15 hours. API bill: $420. Performance went from 10 to 13 token/s, a 30% boost. Still far from 100, but it moved.

Two Approaches

I had GPT 5.5 working the same project in another worktree, instructions basically identical. The two sides tackled it very differently.

GPT 5.5 started by scaffolding with just two layers. Out of 78 total, it first got the pipeline running on two, then dug in, running hundreds of rounds on a single point, squeezing every drop of local speed before scaling up. Steady. Slow. Following the roadmap one step at a time.

Fable 5 did the exact opposite. It spent an hour or two feeling out the single-layer ceiling, then stitched all 78 layers together and went straight for an end-to-end result.

Its engineering was on another level. It built a runnable inference engine from scratch in just over ten hours, a level of completeness previous models couldn't match. But 30% is nowhere near the target, and burning a few hundred dollars a pop isn't sustainable. I ended up switching back to GPT 5.5.

But that's not the point of this post.

An Order-of-Magnitude Gap Between Subscription and API

At launch, Fable 5 was put into Pro and Max subscriptions, counted against quota at twice the Opus 4.8 rate, for a two-week window. After June 22 it was pulled from subscriptions. To keep using it you had to buy usage credits at API rates.

API pricing: $10 input,$50 output, per million tokens.

Max 20x runs $200 a month. Before, one Max 20x seat was enough for Opus 4.8 on complex tasks. Fable 5 at 2x burns through quota faster, but you could still pick which tasks to spend it on. On API, the math changes entirely. A three-hour subtask costs$150; the full project ran $420. That's two months of Max fees.

From Opus 4.5 to 4.6 to 4.8, every upgrade just tweaked the subscription quota ratios. The best model was always included. $200 a month got you access. Now it's not a ratio tweak. It's a straight removal.

Most people can't afford to burn hundreds on a single task.

Anti-Distillation

Anti-distillation is where things get really twisted.

I get not wanting to be distilled. Fable 5 has a built-in two-layer classifier: requests flagged as distillation, cybersecurity, or biochemistry automatically fall back to Opus 4.8. Anthropic says the trigger rate is under 5%.

In reality? People on Reddit and Twitter are saying a routine bloodwork interpretation got flagged as biochemistry, or normal questions got silently downgraded. Distillation is just intensive model use. Once the model is publicly available, you can't fully stop it. Using an opaque classifier to guess intent? Getting it wrong is only a matter of time.

The worst part is that the downgrade is silent. Answer quality tanks out of nowhere, and you think that's just the model, but you're actually getting Opus 4.8. You're paying Fable 5 prices for Opus 4.8 output, and they don't even tell you. I find this more disgusting than the pricing issue.

The Other Side

OpenAI has been doing something different lately. In March they launched Codex for Open Source: six months of free Pro for maintainers of GitHub projects with 1000+ stars. It goes to individuals writing code in the community, not enterprises.

The scale isn't massive. But the direction is toward letting more people use the good stuff, not locking it away.

The two are going head-to-head; who's right or wrong is anyone's guess. But one pulls the best model from subscriptions and silently downgrades users, while the other hands out free accounts to open-source contributors. You tell me. Do those two signals look the same to you?

The World Is at a Crossroads

Anthropic has elitism in its bones. AI constitution, moral charter—the intentions aren't bad, but who gets to define "responsible"? It's always that tiny group, deciding from the top down. The business strategy follows the same path: those with budgets get the best, those without get second tier, with a tenfold gap between.

You can charge a premium for the strongest model. But keeping it in subscriptions lets individuals and small teams allocate their own quotas. Pulling it out and putting it on API changes the signal completely.

A model company's pricing is its stance on how intelligence should be distributed. This fracture is bigger than Anthropic realizes.

References

1. Anthropic, "Introducing Claude Fable 5 and Mythos 5", 2026-06-09
2. Claude official pricing: Fable 5 input $10/M token, output$50/M token; GPT-5.5 input $5/M token, output$30/M token
3. Fable 5 subscription availability: June 9 to 22, after which usage credits at API rates are required
4. Fable 5 anti-distillation mechanism: two-stage classifier detection, falls back to Opus 4.8 when triggered, official trigger rate claimed below 5%
5. Fortune, "Anthropic accused of secretly limiting Claude Fable 5 capabilities", 2026-06-10
6. OpenAI, "Codex for Open Source" program, 2026-03-07, providing 6 months of free Pro accounts to maintainers of 1000+ stars projects
7. SWE-bench Pro benchmark: Fable 5 scored 80%, GPT-5.5 scored 58.6%