TL;DR: Claude Fable 5 came back globally on July 1 and still leads every coding benchmark that matters. But the new cybersecurity classifier reroutes flagged requests — including routine debugging — to the weaker Opus 4.8, and starting after today (July 7) it’s credit-only at $10/$50 per million tokens. Keep Sonnet 5 as your daily driver; reach for Fable 5 only on the hardest tasks.

	Claude Fable 5	Claude Opus 4.8	Claude Sonnet 5
Best for	Hardest planning + architecture	Balanced daily agentic work	Cost-efficient default backend
Price (in / out per M)	$10 / $50	$5 / $25	$2 / $10 (intro, → $3 / $15 Sep 1)
SWE-bench Pro	80.3%	69.2%	63.2%
Terminal-Bench 2.1	88.0%	82.7%	80.4%
The catch	Classifier can silently reroute you to Opus 4.8 mid-task	Slightly behind on gnarliest tasks	New tokenizer = ~30% more tokens per task

Honest take: Fable 5 is objectively the strongest coding model you can buy today, but the classifier means you’re not always getting Fable 5 — you may get Opus 4.8 without asking. For 90% of Cursor and Cline work, Sonnet 5 is the smarter default. Turn to Fable 5 only when a task has already beaten the cheaper models, and use a flat Cursor Pro plan rather than metered API credits.

What actually happened

On June 12, 2026, the US government issued an export-control directive suspending all access to Claude Fable 5 and Mythos 5. The trigger, per Anthropic’s own account and CNBC’s reporting, was a report from Amazon researchers describing a jailbreak: prompt Fable 5 to read a specific codebase and identify software vulnerabilities, and it would comply in a way regulators considered a national-security risk. The order took effect immediately, and because Anthropic had no reliable way to verify user nationality in real time, it went dark for everyone — including Anthropic’s own foreign-national employees.

On June 30, the US Commerce Department lifted the order. Anthropic redeployed Fable 5 on July 1, roughly an 18-day blackout. As of today it’s live again on Claude.ai, the Claude Platform (API), Claude Code, and Claude Cowork. AWS Bedrock, Google Cloud (Vertex), and Microsoft Foundry access is being re-enabled “as quickly as possible” — not all cloud endpoints were back on day one, so verify before you wire a production pipeline to one.

The model itself is unchanged. What changed is the guardrail wrapped around it — and for developers, that guardrail is the whole story.

The classifier reroute is the real headline

Anthropic didn’t just flip Fable 5 back on. It retrained a cybersecurity classifier and layered it in front of the model. The classifier blocks the specific Amazon-reported jailbreak in over 99% of cases. That part works.

The problem is what happens when it fires. When the classifier flags a request as risky, the request doesn’t fail — it’s silently handed to Claude Opus 4.8 instead, and you get a notification that the handoff happened. Anthropic states this plainly in its own redeployment post: the stricter classifier “comes at the cost of flagging benign requests more often during routine coding and debugging tasks.”

Read that again. Anthropic is telling you, in writing, that routine debugging will sometimes get bounced off the model you’re paying a premium for.

Why does normal work trip a cyber classifier? Because the flagged behavior — read a codebase, find the flaws, propose fixes — is structurally identical to ordinary debugging and code review. The classifier can’t reliably tell “fix this null-pointer crash” from “find the exploitable bug.” So it errs toward flagging, and infrastructure code, Rust syscall work, and plain code reviews all get caught in the net.

Independent testing bore this out fast. BridgeMind ran 12 TypeScript coding tasks through Fable 5 after the July 1 relaunch: only 3 of the 12 actually reached Fable 5. The other 9 were rerouted to Opus 4.8 and, in the harness that expected Fable 5’s output, scored zero — a reported ~70% collapse in debugging performance versus the pre-ban model. That’s not a subtle regression. On that sample, you had a 3-in-4 chance of not getting the model you selected.

How to tell which model actually answered

This is the single most important habit for anyone running Fable 5 right now: verify the model that responded, don’t assume. In Claude Code and the API, the response carries the model that actually served it. Check it.

# Claude Code: confirm which model handled the last turn
$ claude --model claude-fable-5 -p "Refactor the auth middleware for clarity"
# ...output...
# If the classifier fired, you'll see an inline notice like:
#   ⚠  This request was handled by claude-opus-4-8 (cyber-safety reroute)

# Via the API, inspect the response model field rather than trusting your request:
$ curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{"model":"claude-fable-5","max_tokens":1024,
       "messages":[{"role":"user","content":"Audit this parser for bugs"}]}' \
  | jq '.model'
# "claude-opus-4-8"   ← you asked for Fable 5, you got Opus 4.8

In Cursor and Cline the reroute is quieter — you selected Fable 5 in the model picker, but a flagged turn is answered by Opus 4.8 under the hood. There’s currently no clean per-turn indicator in the Cursor UI, which is exactly why the API model field is your source of truth for anything that matters.

The one workaround, and its limits

Anthropic opened a Cyber Verification Program: vetted security professionals can apply to use the models for legitimate offensive-security work without the cyber safeguards. If your work legitimately looks like vulnerability research, this is the intended path.

It’s not a clean fix. Anthropic says it expects to occasionally decline eligible applications by mistake, and — critically — that approved users may still hit blocks on legitimate work. That’s not theoretical: there’s an open anthropics/claude-code issue (#67107) titled “Fable 5 keeps downgrading to Opus 4.8 even when approved for Cyber Verification Program.” Anthropic says it will keep refining the classifier to cut false positives but has set no timeline.

For a normal application developer who isn’t doing security research, the program is largely irrelevant. Your practical options are: accept occasional reroutes, or route the affected task type to a model that doesn’t have the classifier in front of it (Sonnet 5, or a local backend).

The pricing cliff hits today

During the relaunch window, Pro, Max, Team, and select Enterprise plans got Fable 5 included for up to 50% of weekly usage limits — through July 7. That’s today. After this window, Fable 5 reverts to being billed against usage credits at the full API rate: $10 per million input tokens and $50 per million output tokens.

That’s exactly double Opus 4.8 ($5/$25) and five times the intro rate for Sonnet 5 ($2/$10, holding until August 31). It is the most expensive model Anthropic has ever priced for general use.

Here’s what that means for a real Cursor agentic run. A moderate multi-file task — say ~50K tokens of context in, ~8K tokens of edits and reasoning out — costs roughly:

Model	Input cost	Output cost	Total per task
Claude Fable 5	$0.50	$0.40	~$0.90
Claude Opus 4.8	$0.25	$0.20	~$0.45
Claude Sonnet 5 (intro)	$0.10	$0.08	~$0.18

Run 30 such tasks a day and Fable 5 is ~$27/day in credits versus ~$5.40 on Sonnet 5. If you’re on Cursor Pro’s flat $20/month plan, Fable 5 access there is far more economical than BYOK API credits — assuming your workload stays within Cursor’s plan limits. For sustained heavy use, the flat plan is the only sane way to touch Fable 5; metered credits will empty your wallet in a week.

Does it still win on benchmarks? Yes — decisively

None of the above changes the raw capability. Anthropic’s published post-relaunch numbers keep Fable 5 at the top of the coding leaderboard:

Benchmark	Fable 5	Opus 4.8	Sonnet 5	GPT-5.5
SWE-bench Pro	80.3%	69.2%	63.2%	—
SWE-bench Verified	95.0%	—	—	—
Terminal-Bench 2.1	88.0%	82.7%	80.4%	83.4%
OSWorld-Verified	85.0%	83.4%	81.2%	—
FrontierCode Diamond	29.3%	13.4%	—	—

Cursor confirmed Fable 5 is available again and leads every model on CursorBench (72.9% on CursorBench 3.1 at max effort) — while noting it’s also the most expensive per task. The FrontierCode Diamond result is the eye-opener: 29.3% versus 13.4% for the next model is the widest gap Fable 5 posts anywhere, and it’s on the hardest problems. That’s precisely where paying for Fable 5 is defensible — genuinely difficult, novel work that the cheaper models fail outright.

The catch, one more time: those benchmark numbers describe Fable 5 answering. If the classifier reroutes your task to Opus 4.8, your effective score is Opus 4.8’s — and you paid Fable 5 rates for the privilege of finding out.

Where it breaks

• Security-adjacent code gets rerouted. Writing a rate limiter, sanitizing input, reviewing auth flows, touching syscalls — all can read as “cyber” to the classifier. Expect Opus 4.8 to answer these more often than not.
• No pre-flight indicator in Cursor/Cline. You find out you were rerouted after the fact, if at all. Build the habit of checking the API model field on anything important.
• Credit burn is brutal. At $10/$50, letting an agent loop autonomously on Fable 5 is the fastest path to a four-figure month. Cap it.
• Cloud endpoints lag. If your stack runs on Bedrock or Vertex, confirm Fable 5 is actually re-enabled there before you depend on it.

If your work is privacy-sensitive or you simply refuse to hand routine debugging to a classifier’s discretion, the fully local route sidesteps the entire problem — no reroute, no credit meter, no export-control risk. See our companion coverage on building a resilient stack in Fable 5 and Mythos 5 got killed by a government order, the running cost breakdown in Claude Fable 5 is now credit-only, and how Sonnet 5 stacks up as a daily driver in our Claude Sonnet 5 coding review. For the GPU side of a local fallback, runaihome.com’s hardware guides cover what VRAM you actually need, and open-weight alternatives worth keeping on hand are tracked at aifoss.dev.

The verdict

Fable 5 is back, and on pure capability it’s still the best coding model money can buy — 80.3% on SWE-bench Pro and a 2× lead on the hardest FrontierCode problems aren’t marketing, they’re measured. But “best model” and “best model you’ll actually get” have diverged. The cybersecurity classifier means a meaningful share of routine coding sessions get silently answered by Opus 4.8, at Fable 5 prices, with no reliable warning in Cursor’s UI.

So the disciplined play is unchanged from before the ban, just for a new reason. Run Sonnet 5 as your default Cursor and Cline backend — it’s five times cheaper, it doesn’t carry the cyber classifier, and it clears the bar for the overwhelming majority of day-to-day work. Keep Fable 5 for the hard 20%: architecture planning, gnarly multi-file refactors, the bug that’s already beaten two cheaper models. Access it through a flat Cursor Pro plan, not metered credits, and check the model field when the answer matters. Treat the reroute as a feature you’re routing around, not one you’re relying on.

FAQ

Is Claude Fable 5 available again? Yes. It was redeployed globally on July 1, 2026 after the US Commerce Department lifted the June 12 export-control order. It’s live on Claude.ai, the API, Claude Code, and Claude Cowork; Bedrock, Vertex, and Foundry are being re-enabled.

Why does Fable 5 sometimes answer as Opus 4.8? A new cybersecurity classifier blocks a specific jailbreak in over 99% of cases. When it flags a request — including some routine debugging and security-adjacent code — the request is rerouted to Claude Opus 4.8 and you’re notified. Independent testing (BridgeMind) saw 9 of 12 TypeScript tasks rerouted.

How much does Fable 5 cost now? After the July 7 window, it’s credit-only at $10 per million input tokens and $50 per million output tokens — double Opus 4.8 and the priciest Anthropic model for general use. A moderate Cursor task runs about $0.90.

Should I switch my Cursor backend to Fable 5? For most work, no. Sonnet 5 ($2/$10 intro) is far cheaper, avoids the classifier, and handles the majority of tasks. Use Fable 5 only for the hardest problems, ideally via a flat Cursor Pro plan.

Can I turn off the reroute? Not generally. Vetted security professionals can apply to the Cyber Verification Program to run offensive-security work without the safeguards, but approved users still report reroutes (claude-code issue #67107). Anthropic says it’s reducing false positives with no set timeline.

Sources

• Redeploying Claude Fable 5 — Anthropic (accessed Jul 7, 2026)
• More details on Fable 5’s cyber safeguards and our jailbreak framework — Anthropic
• Anthropic says Trump admin has lifted export controls on Claude Fable 5 and Mythos 5 — CNBC
• Anthropic is bringing back Claude Fable 5 globally — VentureBeat
• Anthropic Restores Claude Fable 5 After U.S. Lifts Jailbreak-Linked Export Controls — The Hacker News
• Claude Fable 5 Debugging Scores Drop 70%: Safety Classifier Reroutes Tasks — TechTimes
• Anthropic’s Claude Fable 5 Is Back With New Usage Limits And Safeguards — Search Engine Journal
• Claude Fable 5 Benchmarks Explained — Kingy AI
• Real-time cyber safeguards on Claude — Claude Help Center
• BUG: Fable 5 keeps downgrading to Opus 4.8 — anthropics/claude-code issue #67107

---

Last verified: July 7, 2026. Pricing and availability for AI models change constantly — confirm current rates on the Claude Platform pricing page before committing spend.