- Claude Fable 5 is Anthropic’s first publicly available Mythos-tier model (GA June 9, 2026), sharing the same underlying architecture as the restricted Mythos 5
- Compared to Opus 4.8, the big shift isn’t “smarter chat”—it’s async autonomy over hours or days: plan, execute, test, and iterate without you babysitting every turn
- For developers: treat Fable 5 in Claude Code like a senior contractor who can work overnight; keep Opus 4.8 as the default for daily pairing and Dynamic Workflows; both need a Cloud Mac that stays online to close the verification loop
On June 9, 2026, Anthropic shipped Claude Fable 5 and Claude Mythos 5 together. Fable is the first time Mythos-tier capability—previously Mythos Preview in April, Glasswing partners only—lands on the public API, Claude Pro/Max/Team, and Claude Code. Mythos 5 stays partner-only, with some safety classifiers removed.
The dev-world reflex is usually: “New model—how many benchmark points?” That’s the wrong question. The line between Fable 5 and Opus 4.8 is about work shape: from “you watch it edit code” to “you clock out and it’s still editing.”
On June 12, Anthropic temporarily suspended global access to Claude Fable 5 and Mythos 5, citing U.S. government export-control directives, and said it’s working to restore service. The capability and integration differences below still apply once access returns; check the official status page for current availability. Day-to-day development can continue on Opus 4.8 and other Claude models.
Difference 1: Model tier—from Opus flagship to Mythos class
In Anthropic’s product ladder, Opus has long been the “strongest public release” bar; Mythos is the higher tier introduced in 2026, first as Mythos Preview for Project Glasswing (cybersecurity and critical-infrastructure partners).
Claude Fable 5 matters because it puts Mythos-tier capability on the normal path—API, Claude Pro/Max/Team, Claude Code—for the first time. It shares architecture and benchmarks with Claude Mythos 5; the split is mostly the safety shell: Fable ships with classifiers; Mythos is partner-only with some limits removed.
For integrators, this isn’t just swapping a model string. Default task duration, failure modes, and invoice shape all change.
Difference 2: The upgrade is long-horizon autonomy, not chat
Anthropic’s Fable 5 messaging keeps circling three words: thorough, proactive, tests its own work. In plain terms:
Older models felt like a sharp intern who answered when asked. Fable 5 feels like a senior engineer you hand a goal to—it breaks work down, writes code, runs tests, sees failures, and fixes them.
Public case studies: Stripe said Fable 5 compressed months of migration into days, completing a global migration across a 50-million-line Ruby monorepo in a single day; Anthropic’s own teams ran week-plus autonomous genomics research. On Opus 4.8, that usually meant hand-rolling Dynamic Workflows, sub-agents, or a human hitting “continue” every few minutes.
Like Siri becoming a system agent after WWDC 2026, the change is in the execution layer: the model doesn’t wait for the next human prompt—it keeps pushing inside a task budget.
Difference 3: Agentic benchmarks show a step change
On “chat IQ,” Opus 4.8 → Fable 5 is incremental. On whether an agent can finish software engineering work alone, the curve is steep. Anthropic’s published comparison (Fable 5 and Mythos 5 differ by ~1–3 points on most rows; * marks items where Fable’s safety classifiers pull scores closer to Opus):
| Benchmark | Fable 5 / Mythos 5 | Opus 4.8 | How to read it |
|---|---|---|---|
| SWE-Bench Pro | 80.3% | 69.2% | Real-repo issue fixes—~11pt gap |
| FrontierCode (Diamond) | 29.3% (xhigh) | 13.4% (xhigh) | Maintainable agent-written code—roughly 2× gap |
| Terminal-Bench 2.1 | 88.0%* | 82.7% | Multi-step terminal work—Claude Code’s home turf |
| Humanity's Last Exam (tools) | 64.5%* | 57.9% | Cross-domain long-chain reasoning |
| OSWorld-Verified | 85.0% | 83.4% | Computer use—both close |
For iOS / Swift teams, SWE-Bench Pro + Terminal-Bench matter more than HLE: one tests “can this diff merge,” the other tests “will it get lost in shell”—an abstract of the claude CLI ↔ xcodebuild loop.
Opus 4.8’s launch story emphasized honesty, Effort, and Dynamic Workflows; Fable 5 bites off bigger chunks inside a single agent session, so you write fewer orchestration scripts by hand.
Difference 4: 1M context and 128k output
Fable 5 / Mythos 5 ship with a 1 million token default context window and up to 128k output tokens per turn. For long-running agents that means:
- You can load whole-repo READMEs, architecture docs, and CI log excerpts into working memory at once
- With the memory tool and compaction, intermediate state doesn’t have to crowd the chat window
- Anthropic claims file-backed memory makes Fable 5’s long-context task gains roughly 3× Opus 4.8
Opus 4.8 already supports long context and adaptive thinking, but Fable 5 treats “long” as the default agent assumption, not a paid add-on. Migration note: thinking: {"type": "disabled"} is not supported on Fable 5—adaptive thinking stays on; tune depth with effort.
Difference 5: Safety classifiers and Opus 4.8 fallback
This is where Fable 5 and Mythos 5 are same capability, different shell.
Fable 5 runs separate safety classifiers on high-risk queries—cybersecurity, biology, chemistry, distillation. On trigger, the API returns HTTP 200 with stop_reason: "refusal" and which classifier fired—not a classic 4xx error. Integrators need three things (official docs):
- Refusal handling—parse refusals; don’t treat them as successful output
- Fallback—server-side
fallbacks, SDK middleware, or manual retry toclaude-opus-4-8 - Billing—refusals with no output are not charged; fallback has prompt-cache credits so you don’t pay double cache fees
That’s why on ExploitBench, BioMysteryBench, and other * rows, Fable 5 scores hug Opus 4.8—not weaker weights, but the public build refuses to answer. Mythos 5 in trusted environments shows fuller capability (e.g. ExploitBench Cap% ~78% vs Opus 4.8’s 40%).
One ops constraint: Mythos-tier traffic is retained 30 days for trust and safety—not eligible for zero-data-retention (ZDR) agreements. Enterprise compliance reviews need a separate pass.
Difference 6: 2× pricing and new billing rules
Fable 5 is priced at $10 / million input tokens, $50 / million output tokens—roughly 2× Opus 4.8 ($5 / $25), but under half Mythos Preview ($25 / $125).
Price is a task filter: right for multi-hour migrations, audits, and whole-repo test repair; wrong for high-frequency short Q&A (Opus 4.8 Fast or Sonnet is saner there).
On subscriptions, Anthropic bundled Fable 5 free in Pro/Max/Team from June 9–22, then switched to credit consumption from June 23—signaling Fable as premium compute, not the default chat model.
- Do CI / agent scripts handle
stop_reason: refusal? - Is fallback to
claude-opus-4-8configured? - Are billing alerts set on task duration, not request count?
- Do sensitive customers accept the 30-day Mythos-tier retention policy?
Difference 7: Claude Code workflow paradigm shift
For most developers, Fable 5’s front door isn’t claude.ai—it’s Claude Code, the repo-level agent in your terminal. Compared to the Opus 4.8 era, reorder the workflow like this:
| Stage | Opus 4.8 habit | Fable 5 habit |
|---|---|---|
| Task grain | Single file / module / one PR | Cross-directory migration, multi-day acceptance, self-looping tests |
| Orchestration | Hand-written Dynamic Workflows and sub-agents | Built-in long chains + task budgets (beta) |
| Thinking mode | Optional adaptive / thinking off | Adaptive only; raw CoT not returned |
| Memory | Mostly session context | Memory tool + file-backed long-term state |
| Human role | Step-by-step diff review | Set goals, set gates, accept before merge |
Recommended starter on Cloud Mac:
cd ~/work/your-ios-app tmux new -s fable-migration claude # In session (exact commands depend on your install): # /model → claude-fable-5 # /effort → xhigh (large migration) # Example task prompt: # "On branch agent/fable-checkout, migrate the Checkout module sync API to async/await. # Draft a plan first; after I confirm, execute autonomously. After each phase run # swift test --filter CheckoutTests. When all pass, output git diff --stat. Do not push." # Ctrl+B D to detach; laptop can sleep
Same logic as iOS CI cache tuning and speeding up xcodebuild: Fable 5 will hammer the same build repeatedly—DerivedData and Pods with a fixed home beat bare GitHub-hosted macOS on time and quota.
Fable 5 or Opus 4.8—which when?
You don’t have to pick one forever—match model to task horizon and risk:
- Choose Fable 5 for whole-repo migrations, cross-module consistency refactors, and multi-day jobs where the model must run tests and fix its own failures; budget like engineer-days, not chat messages
- Choose Opus 4.8 for daily pairing, Dynamic Workflows experiments, security/bio topics Fable may refuse (or the stable path after fallback), and cost-sensitive high-volume calls
- Choose Cursor local + remote combo for small IDE edits in Cursor; put Fable 5 on Cloud Mac for the big rocks—same split as the Opus 4.8 era, except “big” got bigger
Why this has to run on Cloud Mac
Fable 5’s “long autonomy” demands on the runtime are as hard as Opus 4.8 Dynamic Workflows—often harder:
- Sessions can’t die—hour- and day-scale tasks don’t survive laptop sleep;
tmuxon a 24/7 Cloud Mac is baseline - Real macOS toolchain—
xcodebuild, Simulator, SPM / CocoaPods only run on Mac to validate Swift agent edits - Disk and RAM—1M context plus parallel tests eats memory; M4 24GB and a large disk aren’t luxury on monorepos
- Same machine as CI—agent edits → warm build on-box → failure logs back into the agent tightens the loop
The model ID may move from claude-fable-5 to the next Mythos generation, but the host and workflow don’t need a rebuild—that’s Cloud Mac as the agent’s physical layer.
FAQ
What’s the biggest difference from Opus 4.8?
Long-horizon autonomous agent capability. Chat gaps are modest; repo-scale, multi-day, self-verifying work gaps are large.
Is Fable 5 as strong as Mythos 5?
Same foundation. Fable adds safety classifiers and may fall back to Opus 4.8 in some scenarios; Mythos 5 is Glasswing and other trusted paths only.
API model IDs?
claude-fable-5 (public), claude-mythos-5 (restricted).
Can I use it right now?
Global access was paused from June 12, 2026; restoration timing is per Anthropic announcements. As of this writing (June 15), check official status; Opus 4.8 and other models are unaffected.
Does this conflict with Cursor?
No. Local IDE + remote Claude Code is a common split; Fable 5 owns the overnight big jobs.
Closing
What did Anthropic actually change? One sentence: they handed Mythos-tier long-running agents to everyday developers for the first time. The seven differences above—tier, autonomy, benchmark step change, context spec, safety fallback, pricing, Claude Code paradigm—all point the same way:
The AI coding race is shifting from “who answers smarter once” to “who finishes the job unattended.”
80.3% vs 69.2% on a chart is a signal; the real acceptance test is handing it overnight and waking up to a diff you can merge. Until then, get a Cloud Mac that stays online—Fable 5 and Opus 4.8 both need it; Fable 5 needs it more.
Overnight Fable 5 migrations need a Mac that doesn’t sleep
Vuncloud Cloud Mac M4: tmux long runs for Claude Code, on-box xcodebuild verification, persistent DerivedData cache—the physical execution layer for long-horizon agents.