On June 23, 2026, Claude Fable 5 usage credits become mandatory on every Anthropic paid plan. From June 9 through June 22, 2026, Fable 5 was included at no extra cost on Pro, Max, Team, and seat-based Enterprise. Starting June 23 it is removed from those plans, and any Fable 5 request after a user exhausts the included plan allowance must be funded by usage credits at API rates of $10 per million input tokens and $50 per million output tokens. This post explains exactly what changes, how to model the cost, and how builders are routing around the shock.
Quick Answer
- June 22, 2026 is the last day Claude Fable 5 is included free on Pro, Max, Team, and seat-based Enterprise plans.
- On June 23, 2026, Fable 5 is removed from those plans. Continued access requires usage credits billed at $10 per million input tokens and $50 per million output tokens.
- Inside the plan window, Fable 5 already counted roughly 2x toward usage limits versus Opus 4.8, so the practical cost was never zero.
- Fable 5 was restored on approximately June 18, 2026, six days after the June 12 suspension, with nationality-based access controls and increased fallback to Opus 4.8 on cybersecurity, chemistry, and biology prompts.
- Builders are responding with three patterns: pay the credits, fall back to Opus 4.8 at $5 input and $25 output per million, or route through model-agnostic stacks (Vercel AI Gateway with GLM 5.2, Kimi K2.7 Code, or builders like Totalum that do not pin a single foundation model).
What changes on June 23, 2026
Two changes land at the same time.
First, Fable 5 leaves the included-models list. From June 9 through June 22, the model was bundled on Pro, Max, Team, and seat-based Enterprise plans. On June 23 it is removed. Existing seats still see Fable 5 in the Claude app and via the API, but every request that would previously have drawn from plan usage now requires usage credits.
Second, the cost mechanic switches from "counts double against your plan limits" to straight API-rate billing. While Fable 5 was bundled, sessions counted roughly twice as much against the plan allowance compared to Opus 4.8. That was the implicit price. From June 23 onward, the explicit price applies: $10 per million input tokens and $50 per million output tokens, billed against prepaid usage credits. Refusals before output are not billed, and the fallback-credit mechanic refunds the prompt-cache cost when a refused Fable 5 request retries on Opus 4.8.
Third, the audience matters. Claude Mythos 5 stays in limited release through Project Glasswing. Most builders do not have Mythos 5 access. For the rest of the market, Fable 5 is the highest-capability widely available Claude model, and as of June 23 it is the model that is no longer included in any subscription plan.
If your application uses the claude-fable-5 model id from the Messages API, this is the pricing reality you should plan against starting tomorrow. If your application calls Claude through the AWS Claude Platform, Amazon Bedrock, Vertex AI, or Microsoft Foundry, the same per-token API pricing has always applied, and nothing structural changes on those channels.
Claude Fable 5 usage credits: how they work
Usage credits are Anthropic's prepaid overage system for paid plans. They are available on Pro, Max 5x, Max 20x, Team, and seat-based Enterprise. The mechanic is:
- You use the model normally inside your plan's included usage.
- When you hit the plan's limit, requests stop unless usage credits are enabled and funded.
- With credits enabled, requests continue, and every Fable 5 call burns against credits at the API rate.
The two practical implications for builders:
- Capacity planning becomes line-item, not flat. The old "this is in my plan" assumption breaks. Every Fable 5 message after the plan threshold is a real dollar cost.
- The Opus 4.8 fallback gets sharper teeth. Fable 5 at $10 input and $50 output is exactly 2x Opus 4.8 at $5 input and $25 output. For most product surfaces, the right move is to test whether Opus 4.8 hits the quality bar and reserve Fable 5 for the long-horizon agentic work it was actually built for.
See our earlier coverage of the Claude Agent SDK credits pause for the parallel mechanic on Agent SDK usage. The Agent SDK credit changes are paused as of June 15, 2026, but the Fable 5 plan removal on June 23 is moving forward on schedule.
Cost math: Fable 5 vs Opus 4.8 after June 22
The simplest way to internalize the change is a single-task cost example. Assume an agentic coding task with 80,000 input tokens (context, repo files, prior turns) and 8,000 output tokens (the model's plan plus diffs).
| Model | Input cost | Output cost | Total per task |
|---|---|---|---|
| Claude Opus 4.8 ($5 / $25 per M) | $0.40 | $0.20 | $0.60 |
| Claude Fable 5 ($10 / $50 per M) | $0.80 | $0.40 | $1.20 |
| GLM 5.2 on Vercel AI Gateway ($0.60 / $2.20 per M) | $0.048 | $0.0176 | ~$0.066 |
| Kimi K2.7 Code ($0.95 / $4.00 per M) | $0.076 | $0.032 | ~$0.108 |
For a single user running 10 such tasks per day, Fable 5 lands at $12.00 per user per day, $360 per user per month. Opus 4.8 cuts that in half. Routing the same workload through GLM 5.2 on the Vercel AI Gateway drops it to roughly $2 per user per month, and Kimi K2.7 Code sits at roughly $3 per user per month. The headline benchmarks are not identical, but the order-of-magnitude cost gap is real, which is why so many shipping teams are rebuilding their model routing right now. The GLM 5.2 benchmarks cover the head-to-head numbers in detail.
This math does not mean Fable 5 is the wrong default. For long-horizon agentic work that runs autonomously for an hour and writes tens of thousands of output tokens, the capability difference still favors Fable 5. The math just makes the previous habit of "always pick the flagship" expensive in a way the plan inclusion was hiding.
Three ways builders are responding
There are three patterns we are seeing teams adopt this week as the June 23 transition lands.
Pattern 1: Pay the credits, narrow the surface
Teams that want Fable 5 capability for production workloads are enabling usage credits and narrowing the model's surface area. Instead of Fable 5 on every internal turn, they route the user-facing turn through Fable 5 and the planning and tool-call turns through Opus 4.8. This pattern keeps the headline-quality reasoning on Fable 5 while cutting credit burn by 50 to 70 percent on average. Anthropic's server-side fallback parameter makes this routing easier than a year ago, particularly with the fallback-credit mechanic that refunds the prompt-cache cost on retries.
Pattern 2: Default to Opus 4.8, escalate to Fable 5
Many teams ran the AB test in the last 48 hours and concluded that Opus 4.8 is "good enough" for 80 percent of their requests. They have flipped the default model to Opus 4.8 and reserved Fable 5 for explicit escalation. We covered the Claude 4 deprecation and migration playbook on June 15, and that migration guide is the right reference for teams switching Opus 4.7 or older code paths to 4.8 as the new default.
Pattern 3: Decouple from the Anthropic plan entirely
The most aggressive response is to stop treating the Anthropic plan as the source of truth for model capacity. Teams in this camp are routing through model-agnostic gateways (Vercel AI Gateway, OpenRouter, custom proxies) and treating Claude as one of several backends. This pattern matters because the second Anthropic pricing or availability event in two weeks (the Agent SDK credits pause on June 15 and the Fable 5 plan removal on June 23) is enough signal that single-vendor capacity planning is now a real business risk.
The same logic applies to the Fable 5 suspension. We covered the Fable 5 access suspension on June 13 and the restoration timeline on June 19. Teams that had hard-coded claude-fable-5 saw six days of broken production. Teams that routed through model-agnostic abstractions kept shipping.
The deeper lesson: vendor-pricing volatility
The Fable 5 plan removal is not an isolated event. In the last two weeks alone, Anthropic has:
- Released Fable 5 and Mythos 5 on June 9.
- Suspended access on June 12 in response to a US government directive.
- Paused the Claude Agent SDK credit changes on June 15.
- Restored Fable 5 around June 18 with nationality-based access controls and a more aggressive fallback profile.
- Removed Fable 5 from Pro, Max, Team, and seat-based Enterprise plans on June 23.
Five material events affecting the cost, availability, and behavior of the same model in 14 days. Any product roadmap that assumes one foundation model's pricing or availability is stable enough to plan a quarter against is now operating on faulty assumptions. The teams that are calmest right now are the ones that already separated their product layer from their model layer, so that any individual model change is a config edit, not a rewrite.
The same volatility shows up at the open-weight end. GLM 5.2 on Claude Code shipped a credible 62.1 SWE-bench Pro score on June 17, with full MIT-licensed weights on Hugging Face. Kimi K2.7 Code from Moonshot landed June 12 at a 256K context window with Modified MIT weights. The composition of credible model choices is shifting weekly. Treating any single one as fixed is the same mistake in a different direction.
How Totalum sidesteps single-model risk
Totalum is an AI app builder that creates real, production-grade Next.js applications, owned by the customer, with backend, database, auth, payments, hosting, and custom domain included. Because the app Totalum builds is a normal Next.js project and the source code is fully downloadable, the underlying foundation model that powers the build is an implementation detail. The customer's product does not change when the model behind Totalum changes.
For builders responding to the Fable 5 plan removal, two of Totalum's usage modes are directly relevant:
- Web UI: describe your idea on totalum.app, get a complete production app. Model routing is handled inside the platform; the app you ship is yours.
- AI agent mode (MCP): drive Totalum from Claude, OpenClaw, Cursor, Codex, or any agent via MCP. The agent orchestrates, Totalum builds and maintains the project. If you change which agent you use next month, your app does not change.
The point is not that the foundation model does not matter. It is that the foundation model should not be in the critical path of your product's pricing and availability assumptions.
FAQ
Q: Is Claude Fable 5 still available after June 22, 2026?
Yes. Fable 5 remains generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. The change on June 23 is that it is no longer included in the Pro, Max, Team, or seat-based Enterprise subscription plans. Continued access from inside a plan requires usage credits at API rates.
Q: How much do Claude Fable 5 usage credits cost?
Usage credits bill at the standard Fable 5 API rate of $10 per million input tokens and $50 per million output tokens. This is exactly double Claude Opus 4.8 at $5 per million input and $25 per million output. Refusals before output are not billed, and the fallback-credit mechanic refunds the prompt-cache cost when a refused Fable 5 request retries on another Claude model.
Q: What happened with the Fable 5 suspension and restoration?
Anthropic suspended access to Fable 5 and Mythos 5 on June 12, 2026, in response to a US government export-control directive. Access was restored on approximately June 18 with nationality-based access controls and enhanced compliance screening during API onboarding. Developers report a more aggressive fallback profile to Opus 4.8 on cybersecurity, chemistry, and biological research prompts. Anthropic has not published updated fallback-rate data.
Q: Should I switch to Opus 4.8 by default?
For most product surfaces, yes. Opus 4.8 is half the price of Fable 5 and clears the quality bar for the majority of agentic coding, writing, and analysis tasks. Reserve Fable 5 for explicit escalation on long-horizon agentic work where the additional capability is measurable in your evaluation set. The fallback parameter on the Messages API makes the routing straightforward.
Q: What are the open-weight alternatives if I want to leave Anthropic plans?
The two credible open-weight alternatives shipping this month are GLM 5.2 from Z.ai (62.1 on SWE-bench Pro, MIT-licensed weights on Hugging Face, available on the Vercel AI Gateway at $0.60 input and $2.20 output per million tokens) and Kimi K2.7 Code from Moonshot (256K context window, Modified MIT weights, $0.95 input and $4.00 output per million tokens). Both can be routed through a model-agnostic gateway alongside Claude.
Q: How does Totalum handle the Fable 5 plan change?
Totalum is a foundation-model-agnostic AI app builder. The application Totalum builds is a real Next.js project that the customer owns and can deploy anywhere. The model that powers the build is internal to Totalum. When pricing or availability changes on any foundation model, the customer's product and code do not change. This is the structural difference between an app builder and a coding agent that calls a model directly.
Ready to build something the next Anthropic announcement does not affect?
If the Fable 5 plan removal made you rethink how much of your stack is pinned to a single model, Totalum is built for exactly that scenario. Describe your idea on totalum.app, get a real production Next.js application with backend, database, auth, payments, and hosting included. Start free, scale as your traffic grows, and keep your code portable.
External references for this post: Claude Fable 5 docs on platform.claude.com, GLM 5.2 on the Vercel AI Gateway, Anthropic statement on the June 12 directive.