July 20, 2026: Anthropic split Claude Fable 5 access permanently along plan lines. Max and Team Premium subscribers keep it bundled, capped at 50% of standard weekly limits. Pro and Team Standard subscribers no longer get it included — instead, a one-time $100 usage credit becomes available. After that, Fable 5 is metered at $10 per million input tokens and $50 per million output tokens, Anthropic’s highest consumer rate.
How to actually get the $100
This credit is not automatically dropped into your account with no action needed. Log into claude.ai (or the Claude Console) on your Pro or Team Standard account, go to Settings > Usage, and look for the Fable 5 credit notice there. Claim/activate it from that tab before you start a Fable 5 session, then confirm it shows up in your usage balance. (Exact wording in the UI can shift as Anthropic iterates on the page, but Settings > Usage is where it lives.)
The six-week timeline that got us here
- June 9: Fable 5 launches, Anthropic’s first Mythos-class model, included free for Pro/Max/Team/select Enterprise during an initial window.
- June 12: The US Department of Commerce issues an export control directive under the Export Administration Regulations, requiring Anthropic to suspend Fable 5 and Mythos 5 for any foreign national anywhere, including its own foreign national employees. With no reliable way to verify nationality in real time platform-wide, Anthropic pulls both models for everyone.
- June 30: Commerce lifts the controls after Anthropic trains a new safety classifier targeting the bypass technique that triggered the suspension (an Amazon-researcher-found jailbreak that let the model identify vulnerabilities and, in one case, produce working exploit code). The Department’s Center for AI Standards and Innovation independently tested and approved the updated classifier first.
- July 1: Fable 5 returns. Anthropic says subscription-included access runs through July 7.
- July 7, 12, 19: Three separate extensions, each announced via @claudeai on X, each following user backlash over the deadline.
- July 20: The free window ends for good. The Max/Pro split described above becomes permanent policy, not another temporary extension.
Why $100 disappears fast
At $50/M output tokens, one intensive agentic coding session generating 2 million output tokens (a moderate scope for the multi-file autonomous work Fable 5 is built for) burns the entire credit in a single sitting.
Why Anthropic did this: Sol changed the math
The original plan was to move Fable 5 entirely to metered billing for everyone, no bundle at all. The reversal that kept Max subscribers covered wasn’t a capacity win, it was a competitive response. GPT-5.6 Sol reached general availability July 9 at $5/M input and $30/M output, half Fable 5’s input cost and 60% of its output rate. On the Artificial Analysis Intelligence Index, an independent composite benchmark, Fable 5 scores ~60 and Sol scores ~59, a statistical tie on general capability, with Sol completing equivalent tasks at roughly a third of Fable 5’s per-task cost because it emits fewer output tokens. A Max plan offering zero access to Anthropic’s flagship model, while a competitor matched it for less, was a retention risk Anthropic moved to close.
What actually justifies Fable 5’s premium
Fable 5 is the first Mythos-class model for general use, a tier above Opus that runs always-on adaptive reasoning (chain-of-thought engaged on every request, never returned raw, only the finished output is visible). That’s also why it carries a mandatory 30-day data retention requirement, Anthropic needs to monitor completed sessions for multi-request attack patterns, so zero data retention (available on Opus 4.8 and earlier) isn’t offered for Fable 5 or Mythos 5.
Specs: 1-million-token context window, 128,000-token max output per response, safety classifiers that reroute ~5% of sessions to Opus 4.8 for cybersecurity, bio/chem uplift, or suspected large-scale distillation requests (returns HTTP 200 with stop_reason: « refusal » so you can handle the fallback explicitly).
On benchmarks: Fable 5 leads clearly on SWE-bench Pro, 80% vs. Sol’s 64.6% (Anthropic’s own eval, not yet independently replicated on the same benchmark). Sol leads on agentic/terminal-based coding: 88.8% on Terminal-Bench 2.1 and stronger DeepSWE results. Long-horizon, large-context repo work favors Fable 5; routine agentic loops favor Sol on cost.
One migration gotcha: Fable 5’s new tokenizer can inflate prompts built for pre-Opus 4.7 models by up to 35% in token count, which compounds the per-token price gap on high-volume pipelines.
How to make the $100 last, once claimed
- Route routine work off Fable 5. Sonnet 5 is the new default, $2/M input and $10/M output through an introductory window to Aug 31, 2026. Reserve Fable 5 for jobs that genuinely need the 1M-token context or long-horizon coherence.
- Use the Batch API for anything non-realtime. $5/M input, $25/M output, half the live rate, matching Opus 4.8 standard pricing.
- Check your usage balance before big sessions, not after. The credit doesn’t refill.
- Do the Max math once usage is consistent. If you reliably burn past $100/month, compare that real token volume against Max’s $100-200/month before assuming metered stays cheaper.
The bigger picture
Anthropic hasn’t given a timeline to restore Fable 5 to Pro. Claude Code lead engineer Thariq has publicly said the goal is to bring it back once compute capacity allows, no date attached. The economics explain the hesitation: Fable 5’s adaptive reasoning costs more compute per token than Opus 4.8, and the 30-day retention requirement adds compliance overhead a $20/month plan doesn’t obviously cover. Anthropic also filed a confidential draft S-1 on June 1, 2026, targeting an October Nasdaq listing (last private valuation $965B). Proving Fable 5 earns durable per-token revenue, rather than being a subscription loss-leader, matters directly to how that IPO story gets told. Budget around the current split rather than a reversal.
Worth knowing about while you’re rerouting workloads: Moonshot AI’s Kimi K3 (2.8T total parameters, MoE with 16-of-896 experts active per token, ~50B active compute) goes open-weight on July 27, a genuine self-hostable alternative at this scale, though benchmarks are Moonshot’s own until the weights are public and independently tested. It’s also worth knowing Moonshot is subject to China’s National Intelligence Law, a fixed legal condition for anything processed through their infrastructure, separate from the cost question.
If your team is making these model-routing and cost-tradeoff calls every week and wants it wired into your actual tools instead of decided ad hoc, that’s the kind of automation work Kompose builds.
