Anthropic is making Claude Fable 5 access permanent in Max and Team Premium plans from July 20, but capped at 50% of regular limits. The limits themselves drop by a third that same day. Pro and Team Standard users lose access: they receive a one-time $100 credit, then pay API rates.

As we reported on July 14, Anthropic had extended Fable 5 availability in Max plans until July 19, already the third extension. Now the decision is final, but with tight constraints: the model stays, but at half capacity on already-reduced limits.

Why it matters to you. If you use Fable 5 in daily work, your consumption changes today. Those on top plans get stable access, but with a lower ceiling than you had during the bonus phase. Those on base plans need to do the math: $100 in API credit doesn’t last long if you run coding agents or long sessions. Now’s the time to compare Fable 5 via API against GPT-5.6 Sol, which costs a third for similar performance, and to measure actual consumption instead of relying on impressions.

Anthropic admitted that demand for Fable has been hard to manage. The original plan was to remove it from subscriptions entirely. The reversal comes from competition: OpenAI launched GPT-5.6 Sol at a much lower price point, and Chinese models are pressing on everything below the frontier.

In detail

What came before.

Claude Fable 5 had a rocky path through subscription plans. Blocked for weeks in June by US export restrictions, reactivated July 1, then extended three times in Max plans (until July 19). Each extension was presented as temporary. Anthropic’s original plan was to pull Fable 5 from subscriptions and push everyone to API.

July 20 shifts the paradigm: Fable stays in subscriptions, but under much tighter terms.

What actually changes.

Three things happen at once:

• Regular limits drop by a third. If you had a certain number of messages or tokens per day, you now get roughly 67% of that.
• Fable 5 is available at 50% of those already-reduced limits. In practice, a Max user who could use Fable 5 freely before now has access to roughly a third of the original consumption.
• Pro and Team Standard lose access. They get $100 in API credit once. After that, they pay per use.

The $100 credit needs to be weighed against your actual usage. A coding agent session consuming 50,000 input tokens and 10,000 output tokens costs, at Fable 5 API prices, on the order of a few dollars. But long sessions with accumulated context, repeated tool calls, or pipelines running across dozens of files can burn through that $100 in a week of heavy work.

Why the reversal.

Anthropic acknowledged that managing demand has been frustrating for users. But the decisive factor is competition. GPT-5.6 Sol from OpenAI delivers performance comparable to Fable 5 at a third of the cost, as we saw in the practical comparison from July 13. Below the frontier, Chinese models like Kimi K3 are pressing on pricing.

If Anthropic had pulled Fable 5 from subscriptions, many users would have migrated to GPT-5.6 Sol without looking back. Keeping it, even at reduced capacity, is a way to retain users invested in the Claude ecosystem (Claude Code, MCP, tools running on Anthropic) and to not cede ground in market perception.

How reliable are these numbers.

Data comes from Anthropic’s statement on X, reported by The Decoder on July 18, 2026. The math of 50% on limits reduced by a third is straightforward, but effective limits per user depend on your specific plan and real-time system load. The $100 credit is confirmed, but how long it lasts depends entirely on your usage pattern: no one can tell you without measuring your actual consumption.

What this means for those choosing models.

The move makes the question concrete that anyone using AI for work needs to ask: stay within a vendor’s ecosystem or optimize for cost. Fable 5 in Max plans offers convenience (no API setup, Claude Code integration) but limited capacity. GPT-5.6 Sol via API offers more scaling room at lower cost, but requires managing keys, costs, and tooling. To decide on solid ground instead of impressions, the playbook for comparing two models in fifteen minutes is the starting point.