Claude icon

Claude Mythos 5

Mythos 5Verified

Access Fable 5-level capability with safeguards lifted in some areas through Project Glasswing — the successor to Claude Mythos Preview, delivering the strongest cybersecurity capabilities of any model in the world Accelerate scientific discovery with ~10× faster protein design (9 of 14 strong drug candidates) and autonomous single-cell genomics across 138 species, with novel hypotheses preferred by scientists ~80% of the time over Opus-class models Access in limited availability only for approved cyberdefenders and infrastructure providers, with a trusted biology access program planned — same 1M-token context window and $10/$50 per-million-token pricing as Claude Fable 5

Reviewed by ToolWorthy Editors·updated today·Mythos 5 released yesterday

Pricing:From $10/per use
Visit Site
Jump to section

Featured alternatives

Doubao icon

Doubao

Z.ai icon

Z.ai

Odysseus icon

Odysseus

DocsBot AI icon

DocsBot AI

Taskade icon

Taskade

Earkick icon

Earkick

Pros & Cons

Editor-reviewed

Pros

  • Fable 5's frontier capability with safeguards lifted in some areas, for sanctioned research
  • "Strongest cybersecurity capabilities of any model in the world," per Anthropic
  • Demonstrated scientific impact — ~10× protein design speedup, autonomous genomics across 138 species
  • Hypotheses preferred ~80% of the time over Opus-class models for early-stage research
  • Same 1M-token context, 128k output, and $10/$50 pricing as Fable 5

Cons

  • Not generally available — limited to vetted Project Glasswing customers only
  • Safeguards lifted in some areas, so it is intentionally restricted and requires approval to access
  • Covered Model status means 30-day data retention with no zero-data-retention option
  • Alignment assessment notes misalignment behavior "similar to Opus 4.8"
  • Most teams cannot use it and should default to Claude Fable 5 instead

Overview

Claude Mythos 5 is the limited-access counterpart to Claude Fable 5, released June 9, 2026. It is the same underlying model as Fable 5 with the same capability set, but deployed with its safeguards lifted in some areas rather than the full classifiers that gate the generally available version. Because of that, Mythos 5 is not generally available: it is offered in limited availability to approved customers through Project Glasswing, and is the direct successor to Claude Mythos Preview.

Access is initially granted to a small group of vetted cyberdefenders and infrastructure providers, with Anthropic planning a trusted access program for biology to accelerate biomedical research — customers work through their Anthropic, AWS, or Google Cloud account team to be approved. Teams without Glasswing access are directed to Fable 5, the generally available Mythos-class model. Mythos 5 carries the same 1M-token context window, 128k max output tokens, and $10/$50 per-million-token pricing as Fable 5, and shares its adaptive-thinking behavior. Anthropic's alignment assessment reports Mythos 5's misalignment behavior as "similar to Opus 4.8," and it is designated a Covered Model with 30-day data retention.

What's New

Frontier Cybersecurity Capability

Anthropic describes Mythos 5 as having the strongest cybersecurity capabilities of any model in the world. The Glasswing access model exists precisely because lifting safeguards in some areas exposes capability that Anthropic gates behind vetting — Mythos 5 is positioned for sanctioned defensive security research and infrastructure hardening by approved organizations, not open deployment.

~10× Faster Drug Design

In molecular design work, Mythos 5 accelerated protein design workflows by roughly 10× and generated 9 of 14 strong drug candidates in an evaluated set — moving the model from research curiosity toward a practical accelerant for biomedical discovery pipelines run by approved labs.

Autonomous Genomics Research

Mythos 5 conducted autonomous single-cell analysis across 138 species. In one result, a custom machine-learning model it produced outperformed a published model from a Science paper while being roughly 100× smaller — a concrete demonstration of frontier-level scientific reasoning applied end to end rather than as a chat assistant.

Novel Hypothesis Generation

For early-stage research, scientists preferred Mythos 5's novel hypotheses roughly 80% of the time over Opus-class models — reinforcing its positioning as a research collaborator for the hardest discovery work rather than a general-purpose assistant.

Performance & Positioning

Dimension Claude Mythos 5
Underlying model Same as Fable 5, with safeguards lifted in some areas
Cybersecurity "Strongest of any model in the world" (Anthropic)
Protein / drug design ~10× faster; 9 of 14 strong candidates
Genomics Autonomous single-cell analysis across 138 species
Hypothesis generation Preferred ~80% of the time vs. Opus-class
Availability Limited availability via Project Glasswing only
Alignment Misalignment "similar to Opus 4.8" (Anthropic assessment)

Because Mythos 5 and Fable 5 are the same model under the hood, raw capability benchmarks track Fable 5 (frontier-leading FrontierCode score, and stronger long-context memory use than Opus 4.8 in Slay the Spire testing). The distinction is governance, not performance: Mythos 5 trades general availability for safeguards lifted in some areas under a vetted-access program.

Migration Guide

The supported upgrade path is from Claude Mythos Preview to Mythos 5; Anthropic publishes step-by-step migration instructions for existing Glasswing customers. Key points when moving over:

  • Same integration model as Fable 5. Adaptive thinking is always on (no disabled mode), the raw chain of thought is never returned, and you control depth with the effort parameter.
  • Access is the gate, not the API. Mythos 5 uses the claude-mythos-5 identifier, but availability is controlled through Project Glasswing approval — contact your Anthropic, AWS, or Google Cloud account team.
  • Covered Model retention. Mythos 5 carries 30-day data retention and is not available under zero data retention; confirm this fits your compliance posture before migrating sensitive workloads.

Pricing & Plans

Plan Price
API – Input $10 per million tokens
API – Output $50 per million tokens
Access Project Glasswing limited availability — approval required
Cloud availability Through Anthropic, AWS, or Google Cloud account teams
Alternative Claude Fable 5 — the generally available Mythos-class model

Access note: Pricing matches Fable 5, but Mythos 5 cannot be purchased on demand. It is provisioned only to approved Project Glasswing customers. If you need Mythos-class capability without the approval process, Fable 5 delivers the same underlying model with safety classifiers in place.

Best For

  • Approved cybersecurity research teams conducting sanctioned defensive security and infrastructure-hardening work
  • Infrastructure providers vetted into Project Glasswing for capability that is gated from open deployment
  • Selected biomedical and genomics labs — protein design, single-cell analysis, hypothesis generation — if approved under Anthropic's planned trusted biology access program
  • Organizations with an Anthropic, AWS, or Google Cloud account team that can sponsor Glasswing approval
  • Existing Claude Mythos Preview customers upgrading to the production Mythos 5 model

Everyone else should use Claude Fable 5, which delivers the same underlying model with safety classifiers and full general availability.

FAQ

What is Claude Mythos 5?

Claude Mythos 5 is Anthropic's June 9, 2026 limited-access frontier model — the same underlying model as Claude Fable 5 but with its safeguards lifted in some areas. It is the successor to Claude Mythos Preview and is offered only in limited availability through Project Glasswing — initially to approved cyberdefenders and infrastructure providers, with a trusted biology access program planned.

How is Mythos 5 different from Claude Fable 5?

They are the same model under the hood. Fable 5 is generally available with full safety classifiers and is what nearly all teams should use. Mythos 5 has its safeguards lifted in some areas and is restricted to vetted Project Glasswing customers. The difference is governance and access, not raw capability.

How do I get access to Claude Mythos 5?

Mythos 5 is not generally available. Access is granted to approved customers through Project Glasswing; you request it through your Anthropic, AWS, or Google Cloud account team. Customers without access use Claude Fable 5, the generally available Mythos-class model.

How much does Claude Mythos 5 cost?

Pricing matches Fable 5 at $10 per million input tokens and $50 per million output tokens. However, Mythos 5 cannot be purchased on demand — it is provisioned only to approved Project Glasswing customers.

Is Mythos 5 a successor to Mythos Preview?

Yes. Anthropic publishes a dedicated migration guide for moving from Claude Mythos Preview to Mythos 5 for existing Glasswing customers. Mythos 5 uses the claude-mythos-5 API identifier.

Version History

Fable 5

Released on June 9, 2026

View Update
+What's new
3 updates
  • Run the most demanding reasoning and long-horizon agentic work on Anthropic's most capable widely released model — one Fable 5 session compressed a 50M-line Ruby migration from two months of manual work into a single day
  • Stay focused across a 1M-token context window while solving complex vision problems end-to-end — including completing Pokémon FireRed from pixels alone — with persistent file-based memory that helped 3× more than it helped Opus 4.8 in Slay the Spire testing
  • Deploy with general-availability safeguards that fall back to Opus 4.8 on sensitive prompts in under 5% of sessions, priced at $10 per million input and $50 per million output tokens

Mythos 5

Current Version

Released on June 9, 2026

+What's new
3 updates
  • Access Fable 5-level capability with safeguards lifted in some areas through Project Glasswing — the successor to Claude Mythos Preview, delivering the strongest cybersecurity capabilities of any model in the world
  • Accelerate scientific discovery with ~10× faster protein design (9 of 14 strong drug candidates) and autonomous single-cell genomics across 138 species, with novel hypotheses preferred by scientists ~80% of the time over Opus-class models
  • Access in limited availability only for approved cyberdefenders and infrastructure providers, with a trusted biology access program planned — same 1M-token context window and $10/$50 per-million-token pricing as Claude Fable 5

Opus 4.8

Released on May 28, 2026

View Update
+What's new
3 updates
  • Catch more issues with stronger honesty and judgment — Opus 4.8 is roughly 4× less likely than Opus 4.7 to let code flaws pass unremarked, and is the first model to break 10% all-pass on the Legal Agent Benchmark
  • Drive browser and computer-use agents further with 84% on Online-Mind2Web and more efficient tool-calling that reaches the same result in fewer steps across long agentic sessions
  • Upgrade from Opus 4.7 at the same $5/$25 per-million-token pricing, with better context retention and style consistency across extended sessions

Opus 4.7

Released on April 16, 2026

View Update
+What's new
3 updates
  • Handle harder production coding agents with stronger autonomy and instruction following — CursorBench rises from 58% to 70%, while Rakuten-SWE-Bench resolves 3× more production tasks than Opus 4.6
  • Analyze dense visual content like chemical diagrams, technical screenshots, and patents with 3× higher image resolution (up to 2,576 pixels) and XBOW visual acuity climbing from 54.5% to 98.5%
  • Tune reasoning effort with the new xhigh control level between high and max, giving teams finer latency/quality tradeoffs while instruction-following becomes more literal across tool use

Mythos Preview

Released on April 7, 2026

+What's new
3 updates
  • Discover zero-day vulnerabilities in major operating systems and web browsers autonomously — including a 27-year-old bug in OpenBSD — by applying frontier code reasoning to security-critical software analysis at scale
  • Construct multi-step exploit chains combining up to four vulnerabilities into functional attacks, including JIT heap sprays, ROP chains, privilege escalation, and remote code execution against fully patched targets
  • Achieve 181 successful Firefox JavaScript exploits versus 2 for Opus 4.6, with 595 crash discoveries and 10 full control-flow hijacks in OSS-Fuzz testing — capabilities that emerged from improvements in code reasoning rather than explicit security training

Mythos Preview

Released on April 7, 2026

View Update
+What's new
3 updates
  • Preview a gated frontier model for cybersecurity research, with Project Glasswing offering early access and major gains in zero-day discovery, exploit development, and reverse engineering
  • Analyze complex program behaviors and surface edge-case issues through multi-step reasoning workflows designed for research and testing environments
  • Improve automated code analysis performance in fuzzing and debugging scenarios, helping developers detect crashes and stability issues more efficiently

Sonnet 4.6

Released on February 17, 2026

View Update
+What's new
3 updates
  • Write better code and complete complex multi-step tasks at Opus-level intelligence — developers in Claude Code prefer Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time
  • Navigate spreadsheets, fill out multi-step web forms, and complete browser-based workflows with significantly improved computer use capabilities and stronger resistance to prompt injection attacks
  • Process entire codebases, long contracts, or dozens of research papers in a single request with the new 1M token context window in beta

Opus 4.6

Released on February 5, 2026

View Update
+What's new
3 updates
  • Plan and sustain complex agentic tasks for longer, operate more reliably in larger codebases, and catch more mistakes with improved code review and debugging skills for professional engineering workflows
  • Work with up to 1 million tokens in a single conversation — perfect for analyzing large document sets, long contracts, and extended multi-step agentic task sequences without hitting context limits
  • Achieve top scores on Terminal-Bench 2.0 agentic coding and Humanity's Last Exam, while outperforming frontier competitors by ~144 Elo points on economically valuable professional knowledge work

Opus 4.5

Released on November 24, 2025

+What's new
3 updates
  • Execute long-horizon agentic workflows with improved reasoning and planning — delivering higher pass rates on held-out tests while using up to 65% fewer tokens than earlier models for equivalent results
  • Apply frontier reasoning to everyday work tasks including deep research, multi-step code migrations, and creating slides and spreadsheets — meaningfully better at common professional knowledge work
  • Generate more reliable code solutions with better error handling and stronger self-correction, delivering 50–75% reductions in tool calling and build errors on complex automated engineering tasks

Haiku 4.5

Released on October 15, 2025

+What's new
3 updates
  • Run fast, low-cost coding and agent workloads with near-frontier intelligence, matching Sonnet 4 on coding, computer use, and agent tasks at lower cost
  • Power latency-sensitive customer support, classification, and sub-agent workflows with faster responses and stronger quality than earlier Haiku models
  • Use a current Haiku model across Claude.ai, Claude Code, API, Bedrock, Vertex AI, and Microsoft Foundry for high-volume workloads that need speed and cost control

Sonnet 4.5

Released on September 29, 2025

+What's new
3 updates
  • Balance intelligence and speed with near-Opus-level performance on coding, analysis, and writing at Sonnet pricing — making Opus-class capability accessible for everyday complex workloads at lower cost
  • Generate more accurate responses across coding, writing, and analytical tasks with improved reasoning that matches or exceeds predecessor models on most real-world professional benchmarks
  • Handle larger context windows efficiently while maintaining consistent response quality throughout long conversations — ideal for document-heavy workflows and extended agentic task sessions

Sonnet 4.5

Released on September 29, 2025

+What's new
3 updates
  • Balance intelligence and speed with Opus-level performance at Sonnet pricing for everyday complex tasks
  • Generate more accurate responses across coding, writing, and analytical work with improved reasoning
  • Handle larger context windows efficiently while maintaining consistent quality throughout conversations

Opus 4.1

Released on August 5, 2025

+What's new
3 updates
  • Upgrade Opus 4 coding and agent workflows with better real-world code changes, stronger reasoning, and 74.5% on SWE-bench Verified at the same Opus 4 pricing
  • Track details more reliably in deep research and data analysis, especially when following instructions across large codebases, files, and multi-step agentic searches
  • Improve multi-file refactors and everyday debugging by pinpointing exact corrections in large codebases without making unnecessary changes or introducing new bugs

Opus 4

Released on May 22, 2025

+What's new
3 updates
  • Run long-horizon coding agents for complex, multi-step software work, with Opus 4 leading SWE-bench and Terminal-bench and sustaining focused work over thousands of steps
  • Build more capable agents with new API features including code execution, MCP connector support, Files API, and improved tool use for multi-tool workflows
  • Use frontier reasoning for research, technical analysis, and enterprise workflows with the highest Claude 4 capability tier, while planning around Opus-class cost and ASL-3 safeguards

Sonnet 4

Released on May 22, 2025

+What's new
3 updates
  • Upgrade everyday coding and reasoning workflows from Sonnet 3.7 with stronger instruction following, better software engineering, and faster, lower-cost responses than Opus
  • Use the same Claude 4 platform features, including code execution, MCP connector support, Files API, and improved tool use, for practical agent and app-building workflows
  • Run production workloads at Sonnet pricing with a strong balance of speed, cost, and intelligence for coding assistants, document analysis, and multi-step business tasks

3.7 Sonnet

Released on February 24, 2025

+What's new
3 updates
  • Use one model for quick answers or extended thinking, with API controls over thinking budget so teams can trade latency, cost, and depth on harder reasoning tasks
  • Delegate coding work through the first Claude Code research preview, letting developers ask Claude from the terminal to inspect files, edit code, run tests, and fix bugs
  • Build stronger front-end and coding workflows at the same $3/$15 Sonnet pricing, with extended thinking available on paid Claude surfaces and major cloud platforms

3.5 Sonnet (Upgraded)

Released on October 22, 2024

+What's new
3 updates
  • Control computers, interact with software UIs, and automate desktop workflows through the first publicly available Claude computer use capability for agentic browser and desktop automation tasks
  • Write better code with significantly improved SWE-bench Verified scores versus the original 3.5 Sonnet, making it more reliable for real-world software engineering and complex programming tasks
  • Maintain strong vision and image input capabilities while delivering major improvements in coding quality and computer use for practical software engineering and workflow automation use cases

3.5 Haiku

Released on October 22, 2024

+What's new
3 updates
  • Get fast, cost-effective responses approaching higher-tier model performance for high-volume pipelines, customer support, and content moderation tasks at a fraction of the per-token cost
  • Process high-volume tasks at faster speeds than larger Claude models, enabling real-time customer support, content moderation, and classification pipelines without meaningfully sacrificing accuracy
  • Run fast coding, tool use, data extraction, and content-moderation workloads in a lightweight, budget-friendly 200K-context model — text-only in the Claude API — built for high-throughput production at scale

3.5 Sonnet

Released on June 21, 2024

+What's new
3 updates
  • Outperform Claude 3 Opus on complex tasks including coding and data analysis while maintaining the response speed and lower per-token cost that makes Sonnet practical for production-scale workflows
  • Generate high-quality code with graduate-level reasoning capabilities for sophisticated programming and analytical tasks, matching or exceeding the performance of previous-generation flagship models
  • Handle nuanced instructions across writing, summarization, and Q&A tasks while producing more natural, contextually appropriate responses compared to earlier Claude generations

3 Haiku

Released on March 13, 2024

+What's new
3 updates
  • Process tasks at lightning speed with near-instant response times — ideal for high-volume customer interactions, live chat support, and real-time classification pipelines where latency matters
  • Handle complex multi-step instructions across summarization, Q&A, and coding at a lower cost point than Claude 3 Sonnet while maintaining strong quality for most production use cases
  • Analyze images and documents quickly with multimodal vision capabilities, bringing image understanding to high-throughput workflows in a fast, resource-efficient production model

3 Opus

Released on March 4, 2024

+What's new
3 updates
  • Tackle the most complex reasoning, analysis, and writing tasks with the highest comprehension and fluency of the Claude 3 family — suited for research, legal, financial, and technical domains
  • Analyze images alongside text for comprehensive multimodal understanding, enabling visual Q&A, document extraction, and chart interpretation at the highest quality level in the Claude 3 lineup
  • Generate sophisticated content with improved accuracy, meaningfully fewer unnecessary refusals, and stronger instruction following for nuanced, long-form creative and analytical tasks

3 Sonnet

Released on March 4, 2024

+What's new
3 updates
  • Balance capability and cost with strong performance across writing, analysis, summarization, and coding — a reliable workhorse for most business and enterprise production workflows at scale
  • Process both text and images seamlessly with new multimodal understanding capabilities, enabling document analysis, visual Q&A, and image-grounded reasoning in a single API request
  • Generate reliable, consistent responses for data processing, customer support, and content creation workflows at scale — suited for both high-volume automated pipelines and direct user interactions

2.1

Released on November 21, 2023

+What's new
3 updates
  • Process up to 200K tokens — approximately 500 pages — in a single conversation, enabling comprehensive analysis of lengthy legal documents, research reports, and multi-chapter texts
  • Get more accurate responses with 2× fewer hallucinations than Claude 2.0 when analyzing complex, multi-source information, reducing the risk of fabricated facts in high-stakes professional workflows
  • Integrate Claude with external tools, databases, and APIs through the new beta tool use capability, enabling structured data retrieval and action execution in early agentic task workflows

2

Released on July 11, 2023

+What's new
3 updates
  • Handle longer conversations with a 100K token context window — enough to analyze entire books, lengthy legal documents, or multiple research papers in a single coherent session
  • Write better code with 71.2% performance on the Codex HumanEval benchmark, up from 56% in Claude 1.3, reflecting a meaningful jump in functional code generation and debugging capabilities
  • Receive safer responses with improved Constitutional AI training — approximately 2× reduction in harmful outputs compared to prior versions while maintaining helpfulness and response accuracy

1

Released on March 14, 2023

+What's new
3 updates
  • Access Anthropic's first publicly available AI assistant — built with Constitutional AI training to be helpful, honest, and harmless across conversational and text processing use cases
  • Engage in natural multi-turn conversations for summarization, search, creative writing, Q&A, and coding tasks — with Constitutional AI making responses more predictable and steerable for developers
  • Direct Claude's personality, tone, and behavior to fit your use case — Claude's design emphasizes steerability, allowing more precise customization than earlier AI systems of the same generation

Top alternatives

Related categories

From the blog

View all →

Track Claude in ToolWorthy Weekly

Important tool updates, better alternatives, and selected AI signals in one weekly brief.

Weekly only. Unsubscribe anytime.