Through July 7, Claude Fable 5 is still available inside Claude subscriptions. On July 8, Anthropic pulls it from the Pro, Max, and Team plans: the company's most capable model will only be available pay-per-token, through separate usage credits billed at API rates. Those were the official terms of the model's return on July 1; PCWorld and BleepingComputer have the details.
At Gless we build client systems on Anthropic models every week, so for us this isn't a line in a news feed, it's a line in a budget. Below: the timeline of a turbulent month, how billing works from July 8, why Anthropic is doing this, and the practical question that actually matters — when Claude Fable 5 is worth its $50 per million output tokens, and when you're better off staying on Opus 4.8.
The timeline: launch, ban, comeback, cutoff
Fable 5 spent less than a month inside subscriptions, and most of that month was turbulent. Here's the sequence:
- June 9 — Anthropic releases Claude Fable 5 and Claude Mythos 5, the first models of the Claude 5 family. Subscribers are promised 14 days of access, through June 23.
- June 12 — the U.S. government imposes export restrictions following an Amazon security report. Anthropic takes both models offline for everyone.
- July 1 — Fable 5 returns to subscriptions on reduced terms: up to 50% of a plan's weekly limits, and only through July 7.
- July 8 — the model leaves subscriptions. Access continues only through usage credits at API rates.
Subscribers feel shortchanged. One put it bluntly: "We got to use it for like 3 days out of the 14 we were told, and now we get it for just 7 days at half usage?" Anthropic hasn't budged.
Mythos 5, for the record, never appeared in subscriptions at all: it remains limited to select U.S. organizations under Project Glasswing.
Why people are so attached to Fable 5
Claude Fable 5 is the first model of the new Mythos class, which Anthropic positions above the Opus line. It ships with a 1M-token context window by default and up to 128K output tokens. Reasoning is always on: the model decides for itself how long to think about a task.
In practice the gap with Opus 4.8 shows up not in chat, but in long autonomous work. Fable 5 will grind on a single request for 10–15 minutes: gathering context, building, checking its own output. It reliably runs parallel sub-agents and carries multi-hour runs to completion where previous models needed a human hovering nearby. That's exactly what the now-angry Max subscribers loved it for.
A word on the split in the lineup, because the confusion around it is constant. Fable 5 and Mythos 5 are the same model under the hood. Fable 5 carries additional safety measures for dual-use capabilities — primarily biology and offensive cybersecurity — which is why it's sold to everyone. Mythos 5 runs without those measures, and Anthropic grants it only to vetted organizations.
The price matches the tier. On the API, Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens — exactly double Opus 4.8. Cached input is 90% cheaper: $1 per million on cache hits.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 |
| Claude Opus 4.8 | $5.00 | $25.00 |
| Claude Sonnet 5 | $3.00 | $15.00 |
The output column is the one that matters: on agentic work the model writes and reasons a lot, and reasoning tokens are billed as output. That's why moving Fable 5 to pay-per-token hits hardest for exactly the people who run long autonomous sessions.
How billing works from July 8
From July 8, every Fable 5 token for subscribers is metered through usage credits at standard API rates — the same $10/$50 per million. Credits sit as a separate layer on top of your subscription: they're only spent when you pick Fable 5, while Opus 4.8 and Sonnet keep working within your plan's normal limits.
Anthropic did build in guardrails: a $2,000 daily top-up limit, a configurable monthly spending cap, alerts as you approach it, and optional auto-reload. What the announcement doesn't include: discounted credit bundles, monthly allotments baked into plans, or rollover rules for unused credits. Reading between the lines, none of those exist yet.
One thing worth stating plainly: nothing changes on the API. If your production system calls claude-fable-5 directly over the API, you were paying $10/$50 and you still are. This news only affects access through Claude subscriptions — which mostly means developers running Claude Code on Pro and Max plans.
Why Anthropic is doing this
The official reason is compute scarcity. Anthropic describes demand for Fable 5 as "very high, and difficult to predict," and says it chose to restore subscription access after the export ban "more conservatively, in stages." A Claude Code engineer added: "we aim to restore Fable as a standard part of our subscriptions as soon as capacity allows."
So formally, the measure is temporary. But there's no return date, and "as soon as capacity allows" comes with no guarantees. A month ago these same subscribers were promised 14 days and got three. The sane way to plan a budget is around what exists today: from July 8, Fable 5 means pay-per-token, full stop.
What this means for your business
Price the solved task, not the token. Let's run the numbers on the actual rates. A typical long agent session: 3M input tokens, 80% of them served from cache at $1, plus 300K output tokens. Output: 0.3 × $50 = $15. Input: 0.6 × $10 + 2.4 × $1 ≈ $8.40. Call it $23 for a session that closes several hours' worth of engineering work unattended.
Whether that's expensive depends on the alternative. If Fable 5 closes a task in one run where Opus 4.8 needs two or three iterations with an engineer in the loop, the "expensive" model comes out cheaper: the specialist who reruns and reviews those iterations costs more than the tokens. You can see how we run this math on real projects in the Gless case studies.
Where Opus 4.8 remains the rational choice: interactive development, chatbots, knowledge-base answers, short tasks with a human in the loop. There's no reason to pay double for what already works. Fable 5 earns its keep at the other pole: large code migrations, multi-hour research, overnight autonomous builds — anywhere an extra review cycle costs more than the tokens.
And a third takeaway, the second one this month. In June, access to a frontier model was cut off for political reasons — we covered that episode in our GLM-5.2 post. In July, one is being pulled from a pricing plan over compute scarcity. Different causes, same conclusion: your access to any specific model can change at any moment, and you have no say in it.
Here's what we bake into client architectures so news like this doesn't turn into a fire drill:
- an abstraction layer over the model — switching providers is a config change, not a pipeline rewrite;
- an eval set built on your own tasks — so you can verify a replacement in a day instead of guessing from someone else's benchmarks;
- spending limits and alerts — so a model moving to a new billing scheme shows up in your budget immediately, not in the invoice at month's end.
This is a standard part of our AI implementation services, and June–July 2026 made a vivid case for why it's there.
What to do before July 7
The window is still open, and it's worth spending deliberately:
- Run your real tasks on Fable 5, not someone else's benchmarks. Half of a Pro or Max weekly limit is enough for several serious sessions.
- Record the numbers against Opus 4.8: iterations to a finished result, engineer touch-ups needed, wall-clock time from request to acceptance.
- Use those numbers to decide whether to enable usage credits on July 8, and what monthly cap to set. If your tasks show no difference — congratulations, you just saved money by doing nothing.
- If you work through the API — exhale. None of this applies to you.
Fable 5 remains the strongest model money can buy. From July 8, the money is just counted differently, and it's better to count it before the deadline than after. If you'd like to scope something like this for your business, get in touch.