Claude, 5th Generation: Fable 5, Opus 5, Mythos 5 — The Good, the Bad, and the Cost

This summer, Anthropic released three new models within six weeks—and set a precedent for AI export controls. Here’s what you need to know in practice.

The Three Models at a Glance

Also available are Sonnet 5 ($2/$10 MTok, introductory price through August 31, then $3/$15) and Haiku 4.5 ($1/$5 MTok) for everyday use and high-volume tasks.

Mythos 5 and Fable 5 are the same model. The difference: Fable 5 has security classifiers that automatically reroute cyber and biology queries to weaker Opus models. Mythos 5 does not have these—and is therefore accessible only to vetted partners.

The Good

1. LLM Arena: Anthropic Dominates the Top

In the Agent Arena (as of July 28, 2026, 1.4 million sessions, 44 models), Anthropic holds the top three spots—and five of the top 10:

And in the individual metrics: Opus 5 leads in Confirmed Success (17.56%) and Praise vs. Complaint (25.63%). Fable 5 leads in Steerability (12.93%)—it handles user corrections most reliably.

2. Opus 5: Frontier-Level Quality at Half the Price

Opus 5 is the standout value-for-money option. It costs exactly the same as Opus 4.8 ($5/$25), but delivers:

• Frontier-Bench v0.1: New state-of-the-art, more than twice as good as Opus 4.8 — at a lower cost per task
• CursorBench 3.2: Within 0.5% of Fable 5 on Max-Effort, but at half the price
• ARC-AGI 3: Three times the score of the next-best model
• OSWorld 2.0 (Computer Use): Outperforms Fable 5’s best score at just one-third the cost
• AutomationBench (Zapier): 1.5× higher pass rate than any other model, even at the lowest effort level

Cursor co-founder Sualeh Asif sums it up: "Claude Opus 5 delivers near-Fable 5 intelligence at Opus speed and cost."

3. Project Glasswing: 10,000+ critical vulnerabilities found

Mythos 5 identifies vulnerabilities in mission-critical software as part of Project Glasswing—a collaboration with AWS, Apple, Google, Microsoft, NVIDIA, and others. As of May 2026: over 10,000 high- or critical-severity vulnerabilities. Expanded since June to 150+ organizations in 15+ countries.

4. Agentic Coding: Days Instead of Months

Several early-access customers report dramatic productivity gains:

• Zapier: Opus 5 achieved 100% on AutomationBench—earlier models failed completely
• Lovable: 22% better Agentic Coding performance than Opus 4.7, with significantly less variance
• Zapier (Internal): An engineer at a trading firm built a complete market data feed for a new exchange in a single session using Opus 5 — previous models failed even with detailed plans

5. Opus 5: The Safest Model to Date

According to Anthropic, Opus 5 is the best-aligned model they’ve ever built—it adheres more closely to the Constitution than Opus 4.8, Sonnet 5, or Fable 5, exhibits the lowest rate of deceptive behavior, and is the hardest to jailbreak. It has no data retention requirement (unlike Fable/Mythos 5, which has a 30-day requirement).

The Downsides

1. 23 Days of Downtime in June

Fable 5 and Mythos 5 were launched on June 9. Three days later—on June 12—the U.S. government imposed export controls on Mythos 5. Fable 5 was also affected because it is based on the same underlying model. Both were unavailable until July 1.

For developers who had already migrated to the new models, this meant 23 days of falling back to Opus 4.8—and the question of whether production pipelines can be built on a model that could disappear at any time.

2. Fable 5: Safety Classifiers Block Productive Work

Fable 5 automatically routes cyber and biology queries to weaker Opus models—and not always accurately. Developers who want to use Fable 5 for legitimate security automation or code analysis unexpectedly end up on Opus 4.8. The fallback experience is documented (Why Claude switched models), but it cannot be disabled.

Opus 5, on the other hand, intervenes 85% less frequently—and has a Cyber Verification Program (CVP) for legitimate security work.

3. Mythos 5: Not for You (Unless You’re Vetted)

Mythos 5 remains limited to a small circle of partners. Although Project Glasswing has grown to include over 150 organizations, the model remains inaccessible to everyone else. Anyone who needs unfiltered Frontier performance must make do with Fable 5 and a fallback—or wait.

4. Fable 5: Only 10.26% Confirmed Success

A surprising data point from the Agent Arena: Fable 5 achieves only 10.26% for “Confirmed Success” (how often users confirm the task as complete)—while Opus 5 Max stands at 17.56% and Kimi K3 at 14.23%. This means: Fable 5 may be smarter, but it completes tasks to the user’s satisfaction less often. This may be an “overthinking” issue with shorter tasks.

5. 30-Day Data Retention for Fable/Mythos

Unlike Opus 5 (no data retention), Fable 5 and Mythos 5 require acceptance of 30-day data retention for safety monitoring. This is a real obstacle for regulated industries or GDPR-sensitive workloads.

The Costs: What Do You Really Pay?

API Pricing Comparison

* Sonnet 5 introductory price through August 31, 2026; thereafter $3/$15 MTok.

Prompt Caching delivers up to 90% savings on input tokens—the biggest cost-saving lever for repeated system prompts or long conversations.

Opus 5 Effort Levels

Opus 5 has adjustable effort levels. At low effort, it’s more cost-effective (fewer reasoning tokens); at maximum effort, it reaches nearly Fable-5 level. This enables:

• High-Value Tasks (code review, architecture): Max Effort, Frontier quality
• Everyday Tasks (documentation, refactoring): Medium Effort, lower costs
• High-Volume Tasks (classification, extraction): Low Effort or switch to Sonnet/Haiku

Fast Mode for Opus 5

With Fast Mode, Opus 5 runs about 2.5× faster—at double the price ($10/$50 MTok). Useful for interactive coding sessions where latency is more important than cost.

Plan Pricing (Consumer)

Price Decline Compared to Opus 4.1

A comparison with Opus 4.1 ($15/$75 MTok, November 2025) shows the price decline: Opus 5 costs 67% less than Opus 4.1 and is significantly more powerful. Fable 5’s $50/MTok corresponds to GPT-4’s output price in March 2024—making it cheaper when adjusted for inflation.

Arena Comparison: Anthropic vs. OpenAI vs. Kimi vs. Grok

For comparison, here are the top models in the Agent Arena (Net Improvement, as of July 28, 2026):

The gap between Fable 5 (12.58%) and GPT 5.6 Sol (10.02%) is 2.56 percentage points—a significant lead. And the gap to Grok 4.3 (-8.03%, ranked 38th) is a world apart.

However, the trend is moving toward smarter, more affordable workhorse models: Sonnet 5 (#7, 8.57%) already has more sessions (24,657) than Fable 5 (23,807) and is likely to rise even further in the coming weeks once the rollout begins.

Which model is right for whom?

Choose Opus 5 if you...

• Use coding agents in your daily work (Claude Code, Cursor, Lovable, Devin)
• Automate enterprise workflows that don’t justify a Fable budget
• Work with data privacy-sensitive information (no data retention)
• Don’t feel comfortable with $50/MTok output

Choose Fable 5 if you...

• Need multi-day agent runs that Opus 5 can’t reliably sustain
• Are conducting frontier research where every percentage point of accuracy counts
• Have vision tasks involving complex diagrams, charts, or PDFs
• Have the budget for $50/MTok output

Choose Mythos 5 if you...

• Are part of Project Glasswing or another vetted program
• Are conducting cybersecurity research at the vulnerability exploitation level
• Need biology research without a safety fallback

Choose Sonnet 5 if you...

• Want high volume with solid coding results ($2/$10 MTok)
• Need fast, cost-effective agent runs
• Can’t justify the price jump from Haiku to Opus

Conclusion

With its 5th generation, Anthropic has built three models that dominate the LLM arena—but the real story isn’t Fable 5, it’s Opus 5: Frontier-quality at half the price, with no data retention, and better validation rates than Fable 5 in practice.

The 23-day downtime in June was a wake-up call for anyone who relies on a single model for their production pipelines. Anyone building systems today should plan for a fallback to Opus 5 (or Sonnet 5)—not because the models fail often, but because export controls on Frontier models are no longer just a theoretical possibility.

For most mid-market workloads, the math is simple: Opus 5 (Max) for agentic coding, Sonnet 5 for day-to-day operations, Haiku 4.5 for bulk tasks. Fable 5 is the Ferrari—impressive, but in everyday use, you’ll stick with the Opus.

Further Reading

• Claude Opus 5 — Announcement (July 24, 2026)
• Claude Fable 5 & Mythos 5 — Announcement (June 9, 2026)
• Claude Models & Prices
• Agent Arena Leaderboard
• Project Glasswing
• Claude Opus 5 System Card
• Fable 5 / Mythos 5 System Card