Claude Fable 5 & Mythos 5: The Mythos-Class Breakthrough That Redefines the Frontier

Executive Summary

On June 9, 2026, Anthropic did something it has never done before: shipped a single frontier model as two distinct products. Claude Fable 5 is the new generally available flagship — a Mythos-class model made safe for public use with safety classifiers that route sensitive queries to Claude Opus 4.8. Claude Mythos 5 is the same underlying model with those safeguards lifted in select areas, restricted to vetted cyberdefenders and infrastructure providers through Project Glasswing.

The benchmark numbers are not incremental — they are generational. Fable 5 / Mythos 5 reaches 80.3% on SWE-Bench Pro (vs. Opus 4.8's 69.2%, GPT-5.5's 58.6%, Gemini 3.1 Pro's 54.2%), 29.3% on FrontierCode Diamond (vs. Opus 4.8's 13.4%, GPT-5.5's 5.7%), and 1932 Elo on GDPval-AA (vs. Opus 4.8's 1890, GPT-5.5's 1769). On Terminal-Bench 2.1, it hits 88.0% — surpassing GPT-5.5's previously dominant 82.7%.

But the story is more nuanced than the headline numbers suggest. The starred benchmarks (cybersecurity, biology, HLE) report Mythos 5 scores; Fable 5 falls back to Opus 4.8 on those domains, effectively capping its public capability in exactly the areas where the model is most dangerous. At $10/$50 per million tokens — double Opus 4.8's pricing — Fable 5 is a premium tier, not a drop-in upgrade.

Key finding: The Fable 5 / Mythos 5 release fractures the closed-source frontier into three distinct tiers: Mythos (gated, unrestricted capability), Fable (public, safeguarded), and the previous generation (Opus 4.8, GPT-5.5, Gemini 3.5 Flash). For agentic coding and knowledge work, Fable 5 is the new undisputed leader. For cybersecurity and biology, the model you can buy is deliberately constrained — and the model you can't buy is the one with the headline numbers. This is the first time a frontier lab has explicitly admitted that capability and safety can no longer be tuned with one dial for one audience.

This article analyzes Fable 5 / Mythos 5 across benchmarks, safeguards, pricing, and strategic positioning — placing it in context with the Frontier Trinity Comparison Opus Gpt Gemini Benchmark Showdown 2026 06 01, the Claude Opus Benchmark Evolution 41 To 48 Complete Trend Analysis 2026 05 29 evolution track, and the open-weight challengers covered in Qwen36 27b Dense Beats Moe Agentic Coding Analysis 2026 06 03, Minimax M3 Open Weight Challenger Analysis 2026 06 03, and Gemma 4 12b Encoder Free Laptop Multimodal Analysis 2026 06 04.

---

1. The Release: One Model, Two Products

1.1 What Shipped

Dimension	Claude Fable 5	Claude Mythos 5
Underlying Model	Same Mythos-class model	Same Mythos-class model
Availability	Generally available (API, claude.ai, AWS, GCP, Azure)	Restricted (Project Glasswing partners only)
API Model ID	claude-fable-5	No public API ID
Safeguards	Active (cyber, biology/chemistry, distillation → Opus 4.8 fallback)	Lifted in select areas
Pricing	$10/M input, $50/M output	$10/M input, $50/M output
Data Retention	30-day retention required	30-day retention required
Trigger Rate	<5% of sessions fall back to Opus 4.8	N/A (no fallback)

1.2 Why the Split?

Anthropic's framing is explicit: Mythos-class models have reached a threshold where their capabilities in cybersecurity and biology could be misused to cause serious damage if released without safeguards. The solution was not to dumb down the model, but to ship it twice — once with guardrails for the public, once without for trusted partners.

This is a structural admission that the frontier has outpaced the ability to tune a single model for both maximum capability and maximum safety across all use cases. As Anthropic stated: "The number you read about in the cybersecurity and biology benchmarks belongs to a model you cannot buy."

1.3 The Fallback Mechanism

When Fable 5's safety classifiers detect a request in a safeguarded domain (cybersecurity, biology/chemistry, or distillation), the response is not refused — it is handed to Claude Opus 4.8, which answers in Fable 5's place. Users are informed when this occurs. For the overwhelming majority of business work (coding, analysis, content, research, agentic workflows), the handoff never fires.

Critical implication: On the starred benchmarks (ExploitBench, BioMysteryBench, HLE), the published Fable 5 / Mythos 5 score is actually the Mythos 5 score. A Fable 5 deployment on those tasks performs closer to Opus 4.8.

---

2. Benchmark Analysis: The Five-Model Matrix

2.1 Complete Comparison

Benchmark	Fable 5 / Mythos 5	Mythos Preview	Opus 4.8	GPT-5.5	Gemini 3.1 Pro
SWE-Bench Pro	80.3%	77.8%	69.2%	58.6%	54.2%
FrontierCode (Diamond, xhigh)	29.3%	—	13.4%	5.7%	—
Terminal-Bench 2.1	88.0%*	—	82.7%	83.4% (Codex CLI)	70.7% (Gemini CLI)
GDPval-AA (ELO)	1932	—	1890	1769	1314
GDP.pdf (no tools)	29.8%	—	22.5%	24.9%	16.7%
Blueprint-Bench 2	38.6%	—	14.5%	36.2%	26.5%
AutomationBench	17.4%	—	15.5%	12.9%	9.6%
OSWorld-Verified	85.0%	85.4%	83.4%	78.7%	76.2%
Legal Agent Benchmark	13.3%	—	10.4%	2.1%	0.0%
HLE (no tools)	59.0%*	56.8%	49.8%	41.4%	44.4%
HLE (with tools)	64.5%*	64.7%	57.9%	52.2%	51.4%
BioMysteryBench (hard)	46.1%*	29.6%	40.0%	—	—
BioMysteryBench (human solved)	83.9%*	82.6%	80.4%	—	—
ExploitBench (Cap%)	78.0%*	69.0%	40.0%	34.0%	—
HealthBench Professional	66.0%*	64.7%	56.9%	51.8%	—

*Starred benchmarks show larger divergence between Fable 5 and Mythos 5 due to safeguards. Fable 5 performs closer to Opus 4.8 on these.

2.2 The Clean Reads: Where Fable 5 Leads Uncontested

On non-starred benchmarks — the ones most teams will actually experience — Fable 5's leads are substantial:

• SWE-Bench Pro: +11.1pp over Opus 4.8, +21.7pp over GPT-5.5. This is the single most important number for agentic coding. The gap over GPT-5.5 (58.6%) is so large it effectively ends OpenAI's coding lead, which was the defining narrative of the Frontier Trinity Comparison Opus Gpt Gemini Benchmark Showdown 2026 06 01 analysis.
• FrontierCode Diamond: 29.3% vs. 13.4% (Opus 4.8) and 5.7% (GPT-5.5). This benchmark tests difficult coding tasks held to production-codebase standards. Fable 5 more than doubles Opus 4.8 and is 5× GPT-5.5.
• GDPval-AA: 1932 Elo vs. 1890 (Opus 4.8) and 1769 (GPT-5.5). Knowledge work and financial reasoning.
• Legal Agent Benchmark: 13.3% vs. 10.4% (Opus 4.8) and 2.1% (GPT-5.5). The gap over GPT-5.5 is essentially total dominance.

2.3 The Starred Caveats: What You Can't Buy

On the starred benchmarks, the published numbers belong to Mythos 5, not Fable 5:

• ExploitBench: 78.0% (Mythos 5) vs. 40.0% (Opus 4.8, which is effectively what Fable 5 falls back to). Anthropic separately reports Fable 5 made 0% progress on offensive cyber tasks in blocking mode.
• BioMysteryBench: 46.1% (Mythos 5) vs. 40.0% (Opus 4.8 fallback).
• HLE: 59.0% (Mythos 5) vs. 49.8% (Opus 4.8 fallback).

If you are evaluating Fable 5 for deployment, treat the starred figures as the ceiling of the restricted tier, not the performance you will see.

---

3. Real-World Validation: Beyond Benchmarks

3.1 Software Engineering

During early testing, Stripe reported that Fable 5 compressed months of engineering into days. In a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months by hand.

On Cognition's FrontierCode evaluation, Fable 5 scores highest among frontier models even at medium effort — not just at maximum effort. This suggests the model is more token-efficient than past Claude models, delivering higher quality with less reasoning budget.

3.2 Knowledge Work

On Hebbia's Finance Benchmark for senior-level reasoning, Fable 5 has the highest score of any model, with substantial gains in document-based reasoning, chart and table interpretation, and problem solving. IMC noted that Fable 5 aced their trading-analysis evaluations nearly across the board.

One analytics team reported Fable 5 was the first model to break 90% on their core benchmark of complex, long-running analytical tasks — a 10-point jump over Opus 4.8.

3.3 Vision

Fable 5 is the new state-of-the-art for vision tasks. It can extract precise numbers from detailed scientific figures and rebuild a web app's source code from screenshots alone. Most strikingly, it completed Pokémon FireRed using only raw game screenshots — no maps, no navigation aids, no extra game-state information. Previous Claude models needed complex helper harnesses to attempt the same task.

3.4 Memory and Long-Context

Fable 5 stays focused across millions of tokens in long-running tasks. When playing Slay the Spire with persistent file-based memory, its performance improved three times more than for Opus 4.8, and it reached the game's final act three times more often.

3.5 Scientific Research (Mythos 5)

The Mythos 5 capabilities in science are where the model shows its most dramatic potential:

• Drug design: Accelerated aspects of the process by ~10×. With protein design and bioinformatics tools but no human assistance, Mythos 5 matched or beat skilled human operators. Nine of 14 protein targets yielded strong candidates currently under investigation.
• Novel hypotheses: In blinded head-to-head comparisons, scientists preferred Mythos 5's molecular biology hypotheses ~80% of the time over Opus-class models. One hypothesis about an E. coli protein mechanism was independently corroborated by another lab.
• Genomics: Conducted novel research over a week of largely autonomous work, assembling single-cell data for millions of cells across 138 animal species. Its trained model outperformed a recent Science publication despite being 100× smaller.

---

4. Comparison with the Frontier Trinity

The Frontier Trinity Comparison Opus Gpt Gemini Benchmark Showdown 2026 06 01 analysis documented a key finding: no single model family held a clear overall lead. Opus 4.8 led on math and trustworthiness, GPT-5.5 led on coding and terminal workflows, Gemini 3.5 Flash led on tool orchestration.

Fable 5 collapses that trifecta into a single leader.

Dimension	Previous Leader	Fable 5 / Mythos 5
Agentic Coding (SWE-Bench Pro)	Opus 4.8 (69.2%)	80.3% (+11.1pp)
Terminal Workflows	GPT-5.5 (82.7%)	88.0% (+5.3pp)
Math / Multidisciplinary (HLE)	Opus 4.8 (49.8%)	59.0% (+9.2pp)
Knowledge Work (GDPval-AA)	Opus 4.8 (1890)	1932 (+42 Elo)
Tool Use (AutomationBench)	Opus 4.8 (15.5%)	17.4% (+1.9pp)

The only benchmark where Fable 5 does not cleanly lead is OSWorld-Verified (85.0% vs. Mythos Preview's 85.4%) — a tie within statistical noise.

4.1 What This Means for the Trinity Narrative

The Trinity analysis identified three strategic theses:

1. Anthropic: "Make the model trustworthy for autonomous work"
2. OpenAI: "Make the model build software"
3. Google: "Make the model do everything"

Fable 5 validates Anthropic's thesis while simultaneously stealing OpenAI's coding crown and Google's tool-use lead. The divergence narrative is over — at least for the closed-source tier. Anthropic has converged the three specializations into one model.

---

5. Comparison with Open-Weight Challengers

The open-weight landscape covered in recent articles presents a different picture. Let's place Fable 5 against the three most capable open-weight models:

Model	SWE-Bench Pro	SWE-Bench Verified	Key Strength	Cost
Fable 5 / Mythos 5	80.3%	~95% (CursorBench SOTA)	All-round frontier	$10/$50 per M tokens
Qwen3.6-27B	—	77.2%	Dense efficiency, perfect tool calling	Free (Apache 2.0)
MiniMax M3	59.0%	—	1M context, multimodal	$0.60/$2.40 per M tokens
Gemma 4 12B	—	—	Encoder-free laptop multimodal	Free (Apache 2.0)

5.1 The Gap Remains Real

Even against the most capable open-weight models, Fable 5's lead is substantial:

• vs. Qwen3.6-27B: Qwen3.6 achieved 77.2% on SWE-Bench Verified (impressive for a 27B dense model), but Fable 5's ~95% on equivalent benchmarks represents a qualitative gap. Qwen3.6 is the most efficient frontier-adjacent model; Fable 5 is the frontier itself.
• vs. MiniMax M3: M3's 59.0% on SWE-Bench Pro is impressive for an open-weight model, but it trails Fable 5 by 21.3 points — a gap larger than the one between GPT-5.5 and Gemini 3.1 Pro in the Trinity comparison.
• vs. Gemma 4 12B: Gemma 4 12B solves a different problem (laptop-deployable multimodal). It's not competing on the same benchmark surface.

5.2 The Cost Equation

Here's where the open-weight story still matters:

• Fable 5: $10/$50 per million tokens. A complex coding session with 500K input + 200K output tokens costs ~$15.
• MiniMax M3: $0.60/$2.20 per million tokens (promo pricing). Same session costs ~$1.04.
• Qwen3.6-27B / Gemma 4 12B: Free to self-host (hardware costs aside).

For teams where 80% of closed-source quality at 1/10th the cost is acceptable, the open-weight models remain the rational choice. Fable 5 is for the work where the last 20% of quality matters more than the token bill.

---

6. The Safeguard Architecture

6.1 How It Works

Fable 5's safeguards are not prompt-level refusals — they are separate AI classifier systems that monitor the conversation in real-time. When a classifier detects potential misuse, the main model (Fable 5) is prevented from responding, and the query is routed to Opus 4.8 instead.

Three domains are covered:

1. Cybersecurity: Exploitation, offensive cyber tasks, agentic hacking (reconnaissance, discovery, lateral movement). Fable 5 made 0% progress on offensive cyber tasks in blocking mode.
2. Biology and chemistry: Currently a broad net that Anthropic acknowledges is overly broad. Narrowing is planned so legitimate biomedical research is not caught.
3. Distillation: Extraction attacks designed to siphon the model's behavior to train competing models.

6.2 Robustness

• External bug bounty: 1,000+ hours of testing, no universal jailbreaks found
• UK AI Safety Institute: Made progress toward a jailbreak within an initial testing window
• External partner testing: Zero harmful single-turn responses across 30 public jailbreak techniques
• Internal evaluation: Fable 5's safeguards show greater resistance to jailbreaks than any previous generally accessible model

Anthropic's goal: make any remaining jailbreaks "slow and costly enough to detect and stop before they are used at scale."

6.3 The Data Retention Policy

A new requirement for all Mythos-class traffic: 30-day data retention on both first-party and third-party surfaces. Anthropic states this data will not be used to train new Claude models. This is a significant departure from the no-retention policy on earlier models and raises questions for privacy-sensitive deployments.

---

7. Pricing and Economic Implications

7.1 The Rate Card

Model	Input (per M tokens)	Output (per M tokens)	Prompt Caching Discount
Fable 5 / Mythos 5	$10	$50	90% on cached input
Opus 4.8	$5	$25	90% on cached input
Opus 4.7	$5	$25	90% on cached input
GPT-5.5	~$12.50	~$50	Varies
MiniMax M3	$0.60	$2.40	N/A

Fable 5 is double Opus 4.8's pricing. Unlike the Opus 4.7→4.8 upgrade (same price), this is a tier decision with a real cost attached.

7.2 The Subscription Rollout

• June 9–22: Fable 5 included at no extra cost on Pro, Max, Team, and Enterprise subscription plans
• June 23+: Usage credits required on subscription plans
• Future: Will be restored as standard inclusion when capacity allows (no firm date)

7.3 When Fable 5 Earns the Premium

The doubled cost is justified when:

• The work genuinely benefits from sustained autonomy and self-verification
• A framework migration across thousands of files (Stripe's 2-month → 1-day case)
• Multi-day research-and-build tasks that would otherwise need a sprint
• Complex knowledge work where the cost of a missed detail outweighs the token bill

When Opus 4.8 remains the right default:

• Routine, high-volume, or latency-sensitive work
• Classification, summarization, drafting, interactive chat
• Workloads near cybersecurity or biology (where Fable 5 falls back to Opus 4.8 anyway)

---

8. Strategic Implications

8.1 The Three-Tier Frontier

Fable 5 / Mythos 5 creates a new hierarchy:

8.2 What This Means for the Industry

1. Capability-safety decoupling is now explicit. No longer can labs claim a single model is both maximally capable and maximally safe for all audiences. The split is structural.
2. The coding crown has changed hands. GPT-5.5's dominance on SWE-Bench and Terminal-Bench is over. Fable 5 leads by wide margins on both.
3. The open-weight gap is real but narrowing. Qwen3.6-27B at 77.2% SWE-Bench Verified is impressive, but Fable 5's ~95% represents a qualitative leap that open-weight models haven't matched yet.
4. Pricing tiering is the new competition. With base token costs converging, the competitive lever is feature pricing (caching, batch, tool-use) and capability-tier pricing. Fable 5 at 2× Opus 4.8 sets a precedent.
5. The biology safeguard is a wild card. Until Anthropic narrows the overly broad biology/chemistry net, legitimate life-sciences work may be forced to use Opus 4.8 — or wait for the trusted access program.

8.3 Connection to Prior Analysis

This release validates several themes from recent journal articles:

• From Frontier Trinity Comparison Opus Gpt Gemini Benchmark Showdown 2026 06 01: The trifecta narrative (Opus = math, GPT = coding, Gemini = tools) is resolved. Fable 5 leads on all three dimensions.
• From Claude Opus Benchmark Evolution 41 To 48 Complete Trend Analysis 2026 05 29: The Opus evolution trajectory (capability → agentic specialization → reliability) culminates in the Mythos-class leap. The 41-day cycle from 4.7 to 4.8 was a warm-up for this release.
• From Qwen36 27b Dense Beats Moe Agentic Coding Analysis 2026 06 03: The dense-vs-MoE debate continues, but Fable 5 (architecture undisclosed) shows that the closed-source frontier remains out of reach for open-weight models on the hardest coding tasks.
• From Minimax M3 Open Weight Challenger Analysis 2026 06 03: M3's 59% SWE-Bench Pro was the open-weight ceiling. Fable 5's 80.3% raises that ceiling by 21 points — a gap M3 would need a generational upgrade to close.

---

9. Key Takeaways

1. Fable 5 is the new frontier leader on agentic coding, knowledge work, vision, and general reasoning — but only for non-safeguarded domains.
2. The starred benchmarks are Mythos 5 numbers, not Fable 5 numbers. Treat them as the ceiling of the restricted tier.
3. At 2× Opus 4.8 pricing, Fable 5 is a strategic tier decision, not a default upgrade. Route by task, not by model.
4. The open-weight gap remains large on the hardest tasks, but models like Qwen3.6-27B and MiniMax M3 deliver 70-80% of closed-source quality at a fraction of the cost.
5. The three-tier frontier (Mythos → Fable → previous generation) is a structural shift that will define the industry for the rest of 2026.
6. The biology safeguard is temporarily overly broad — watch for narrowing in the coming weeks if your work touches life sciences.

---

10. References & Resources

• Anthropic: Claude Fable 5 and Claude Mythos 5
• Anthropic: Fable 5 & Mythos 5 System Card
• Anthropic: Claude Fable 5 Product Page
• Project Glasswing
• Digital Applied: Fable 5 & Mythos 5 Analysis
• Frontier Trinity Comparison Opus Gpt Gemini Benchmark Showdown 2026 06 01
• Claude Opus Benchmark Evolution 41 To 48 Complete Trend Analysis 2026 05 29
• Qwen36 27b Dense Beats Moe Agentic Coding Analysis 2026 06 03
• Minimax M3 Open Weight Challenger Analysis 2026 06 03
• Gemma 4 12b Encoder Free Laptop Multimodal Analysis 2026 06 04

---

11. Future Directions

Several questions remain open:

• Architecture: Anthropic has not disclosed Fable 5's parameter count, architecture, or training methodology. Is it a scaled-up Opus 4.8, or a fundamentally new design?
• Biology safeguard narrowing: When will the overly broad biology/chemistry net be tightened? This is the single biggest blocker for life-sciences adoption.
• Mythos 5 trusted access: How broad will the trusted access program become? Will it extend to academic researchers, or remain limited to government and infrastructure partners?
• Open-weight response: Will Qwen, MiniMax, or other open-weight teams release models that close the 20-point SWE-Bench Pro gap? The dense-efficiency thesis from Qwen3.6-27B suggests it's possible.
• Competitor response: How will OpenAI and Google respond? GPT-5.5's coding lead is gone. Gemini 3.5 Flash's tool-use lead is gone. The next releases from both will need to address this directly.
• Mythos-class successors: Anthropic has signaled "more capable models arriving in the coming months." The Fable 5 launch may be the beginning of a new capability curve, not the end of it.

---

The frontier has fractured. The model you can buy is not the model with the headline numbers. And for the first time, that's not a bug — it's the product.