Best Claude Model Now That Fable 5 Is Disabled (Mythos vs Opus vs GPT-5.5)

The Short Answer

With Fable 5 and Mythos 5 suspended, your best available model for most serious work is Claude Opus 4.8. It is the closest thing to Fable that you can actually call today, and it is fully unaffected by the directive.

But "best" depends on what you are doing. Here is the breakdown.

The Replacements, Ranked

Claude Opus 4.8 (claude-opus-4-8): the default pick. Closest in capability to Fable, available everywhere Fable was, and the lowest-friction swap. If you had pinned claude-fable-5, change it to claude-opus-4-8 and you are running again. The main thing you give up is Fable's long-horizon task persistence, the ability to grind on a multi-step goal without losing the thread, which was the clearest difference between the tiers.

OpenAI GPT-5.5: the cross-vendor hedge. If today taught you anything, it is that a single provider can vanish overnight. GPT-5.5 is a strong second source, and notably, Anthropic itself pointed to GPT-5.5 as having comparable capability on the specific task the government flagged. For agentic coding, Opus 4.8 still edges it in our testing, but having a tested fallback on a different vendor is no longer optional.

Claude Sonnet 4.6 (claude-sonnet-4-6): the cost-aware workhorse. For high-volume, latency-sensitive, or cost-sensitive workloads, Sonnet 4.6 is the right tool. You would not reach for it to replace Fable on your hardest reasoning tasks, but for most production traffic it is the sensible default that Fable was overkill for anyway.

Open-weight models (Qwen, Kimi, GLM, DeepSeek): the floor nobody can pull. The strongest argument for open weights was never benchmark parity. It is that no government letter can switch them off. They define your worst-case capability floor: the level your product degrades to but never below, because the weights are on your disk. After today, deciding your open-weight floor is a real architectural decision, not a hobby.

Pick by Task

• Agentic coding, hard reasoning, long multi-step jobs: Opus 4.8, with GPT-5.5 as your tested fallback.
• High-volume API traffic, summarization, classification: Sonnet 4.6.
• Cheap, fast, simple calls: Haiku 4.5 (claude-haiku-4-5-20251001).
• Anything that absolutely cannot go down: an open-weight model you control, as your floor.

Do Not Just Swap the String

The lesson of the Fable suspension is not "use Opus instead." It is that model availability is now operational risk. Hardcoding any single model ID, from any vendor, is a single point of failure.

The fix is a model abstraction layer: one config switch for provider and model, a fallback chain you actually test against your eval suite, and a canary that detects silent model substitution. We wrote the full playbook in Your Stack Has a Single Point of Failure.

Background

If you are still catching up on why your model disappeared: what happened, why it happened, and the difference between Mythos and Fable. To check whether Fable is back yet, see Is Claude Fable 5 Down?.