Claude Mythos Preview: Developer Guide for 2026

On April 7, 2026, Anthropic released Claude Mythos Preview — not to the public, but to a curated list of 12 major technology companies and 40+ additional organizations. The reasoning was simple: Mythos is capable enough to autonomously discover and exploit zero-day vulnerabilities in production software. Anthropic decided the world needed defenders armed with it before attackers could replicate it.

That decision is both a milestone and a message. It tells developers that AI capability has crossed a threshold where release strategy matters as much as the capability itself. This guide covers what Mythos can actually do, how to understand its benchmarks, who has access today, and what developers should be watching as Anthropic charts a path toward eventual broader availability.

Why Claude Mythos Changes the Benchmark Conversation

Every major AI release arrives with a table of benchmark scores. Most of the time, the improvements are incremental — a point here, two points there. Claude Mythos is different.

On SWE-bench Verified — the standard benchmark for real-world software engineering task completion — Mythos scored 93.9%. That is currently the highest reported score on SWE-bench Verified of any model. For context, Anthropic's own Claude Opus 4.6 sits in the mid-70s on the same benchmark. The jump represents not an incremental improvement but a qualitative change in what autonomous coding agents can reliably handle.

The USAMO score is particularly striking: Mythos moves from roughly 60% (Opus 4.6) to 97.6% on 2026 competition mathematics problems. On competition math, that means Mythos is almost never wrong. That kind of reliability on adversarial reasoning tasks is what makes it relevant to security work — complex attack chain reasoning requires the same deep, multi-step inference that competition math demands.

CyberGym is the benchmark most relevant to Mythos's actual deployment context. Designed specifically to measure AI models' ability to reproduce and exploit known software vulnerabilities, it is one of the few benchmarks where a 16-point gap between models has direct real-world implications. Mythos at 83.1% means that in roughly 5 out of 6 attempts on known CVEs, it can successfully reproduce the exploit. Opus 4.6 at 66.6% succeeds about 2 out of 3 times. In defensive security workflows — finding vulnerabilities before attackers do — that difference compounds.

Project Glasswing: How Anthropic Gated the Most Powerful Model

Rather than a standard API launch, Anthropic built a controlled consortium around Mythos. Project Glasswing is described as an effort to "secure the world's most critical software for the AI era," and its partner list reflects that mandate.

Launch partners include:

• Amazon Web Services
• Apple
• Broadcom
• Cisco
• CrowdStrike
• Google
• JPMorgan Chase
• The Linux Foundation
• Microsoft
• NVIDIA
• Palo Alto Networks

Beyond the named 12, Anthropic has extended access to over 40 additional organizations that build or maintain critical software infrastructure. Anthropic committed $100M in usage credits to seed the initiative.

The rationale is explicit in Anthropic's documentation: Mythos is capable enough that they want defenders using it before attackers can build equivalents. During internal testing, Mythos autonomously discovered and exploited zero-day vulnerabilities across major operating systems and browsers — including a 27-year-old patched bug in OpenBSD. Where Opus 4.6 developed working JavaScript shell exploits on Firefox vulnerabilities only twice out of several hundred attempts, Mythos succeeded 181 times.

That number — 181 vs. 2 — is the clearest illustration of why Anthropic chose a gated release.

Where Claude Mythos Is Technically Available

For authorized organizations, Mythos Preview is accessible through three cloud platforms:

Amazon Bedrock (US East — N. Virginia region) is the primary distribution channel. AWS announced Mythos Preview availability on April 13, 2026 as part of their weekly AWS roundup. Access is through an allowlist managed by AWS and Anthropic jointly.

Google Cloud Vertex AI offers Mythos Preview on an invitation-only basis. Google published dedicated documentation on the Vertex AI integration, and it supports the same API patterns as other Claude models on the platform.

Microsoft Azure AI Foundry added Mythos support as part of their expanded partnership with Anthropic. Access conditions mirror Bedrock and Vertex: you need to be part of a Glasswing-approved organization or apply directly through Anthropic.

Pricing (for approved organizations): $25 per million input tokens, $125 per million output tokens. That is approximately 3× the price of Opus 4.6. Anthropic has not announced standard public pricing because there is no standard public availability.

For developers who want to understand Mythos's capabilities before access becomes available to them, the model documentation is live on the Claude API docs under claude-mythos-preview-20260407.

What Claude Managed Agents Means for Developers Right Now

One day after the Mythos announcement, Anthropic launched something developers can access today: Claude Managed Agents, which entered public beta on April 8, 2026.

Managed Agents is the infrastructure layer for running Claude as an autonomous agent in production. Instead of building your own agent loop, tool execution engine, and runtime, you get a managed environment where Claude handles:

• Secure sandboxing — code execution, file reads, and web browsing happen in isolated environments
• Long-running sessions — agents can operate autonomously for hours, with persistent state across disconnections
• Scoped permissions — granular tool access controls rather than all-or-nothing
• Built-in tracing — observability over what the agent did and why

The pricing model is standard Claude API token rates plus $0.08 per session-hour. For most agentic workloads, the session-hour cost is minimal compared to token costs; it's essentially the infrastructure overhead charge.

Managed Agents is relevant to this guide because it represents the infrastructure pathway that Mythos will eventually run on for broader deployments. The multi-agent coordination feature — where agents spawn sub-agents for parallel complex tasks — is currently in research preview within Managed Agents, with early access through the Claude Platform console.

If you are building autonomous coding workflows today using Claude Sonnet 4.6 or Opus 4.6, Claude Managed Agents is worth evaluating now. When Mythos access expands, the migration path will be a model swap within the same infrastructure.

Common Mistakes Developers Make When Interpreting Mythos

Treating benchmark scores as performance guarantees. 93.9% on SWE-bench means Mythos fails roughly 6% of evaluated tasks. For production autonomous coding agents, that failure rate matters — especially in unattended workflows. Benchmark scores establish capability ceilings; actual reliability depends on your task distribution, prompt quality, and error handling.

Assuming Glasswing access equals general developer access. Project Glasswing organizations are using Mythos for critical infrastructure security work — not general application development. There is no self-serve sign-up, and Anthropic has not announced a timeline for general availability.

Conflating Mythos with an "unsafe" model. The gated release is not because Mythos is unaligned or unreliable. Anthropic's public documentation is clear that this is a calculated release strategy for capability-matched deployment. The model itself follows the same Constitutional AI principles as other Claude models.

Ignoring Claude Managed Agents. Many developers are waiting for Mythos while overlooking the fact that Managed Agents — available today — solves the infrastructure problems that limit production agent deployments regardless of which model you use. Start there. For more on agentic infrastructure, see our guide to Claude Code advanced workflows.

Expecting Mythos pricing to stay at $25/$125. That pricing is for research preview partners under constrained access. Historical pattern from other Anthropic model launches suggests that pricing adjusts as capacity scales and competition increases. The frontier model premium tends to compress over 12-18 months.

How Mythos Fits Into the Broader AI Model Landscape

Mythos arrives in a crowded frontier model moment. Within the same two-week window in April 2026, developers also processed Qwen3's hybrid thinking modes, Grok 4's multi-agent architecture release, and GLM-5.1's SWE-bench showing. The benchmarks are moving fast across the board.

What distinguishes Mythos is not just raw benchmark performance but the specificity of its capability focus. Anthropic built and evaluated Mythos explicitly for cybersecurity, long-running autonomous agents, and complex multi-step reasoning. The CyberGym score is not a side effect of general intelligence gains — it reflects targeted training and evaluation investment in security-relevant capabilities.

This matters for developers thinking about model selection. Mythos is not "Claude, but better." It is Claude, optimized for a specific and demanding class of tasks. For those tasks — autonomous security research, vulnerability analysis, complex codebase reasoning at scale — it represents a generational step. For general-purpose application development, Opus 4.6 or Sonnet 4.6 remain the practical choices at current access levels.

The comparison with Grok 4's multi-agent architecture is instructive: xAI bet on multi-agent coordination as the differentiating architecture, while Anthropic bet on raw model capability at the security reasoning frontier. Both bets are live simultaneously.

FAQ: Claude Mythos Developer Questions

Q: Can I apply for Project Glasswing access as an individual developer?

Not currently. Project Glasswing access is extended to organizations building or maintaining critical software infrastructure. Anthropic has not published an individual developer application pathway. The closest available path is through an authorized cloud provider (AWS Bedrock or Google Vertex AI) if your organization qualifies.

Q: Will Claude Mythos eventually be available on claude.ai or the standard API?

Anthropic has stated their "eventual goal is to enable users to safely deploy Mythos-class models at scale," but has not committed to a timeline or confirmed it will reach standard API availability in its current form. Capability-restricted or fine-tuned variants may arrive before a full Mythos release.

Q: How does Claude Mythos compare to GPT-5.4 for coding tasks?

SWE-bench Verified provides a direct comparison point: Mythos at 93.9% represents the highest publicly reported score on that benchmark as of April 2026. Direct head-to-head evaluations across all benchmarks have not been published by either company, and independent third-party evaluations are limited by Mythos's restricted availability.

Q: What model ID do I use for Claude Mythos Preview in API calls?

The model identifier is claude-mythos-preview-20260407, consistent with Anthropic's standard model versioning pattern. This ID is used in both Amazon Bedrock and Google Vertex AI integrations for authorized users.

Q: Is Claude Managed Agents required to use Mythos?

No. Managed Agents is separate infrastructure — an agent hosting layer that works with any Claude model. Mythos can be accessed directly via the standard Claude API for authorized partners. Managed Agents is a productivity multiplier for running agents in production, not a prerequisite for Mythos access.

Q: What's the context window for Claude Mythos Preview?

Anthropic's documentation references support for long-context tasks consistent with other frontier Claude models, but specific context window sizes for Mythos have not been independently confirmed in public documentation at this time. [DATA NOT AVAILABLE] for official Mythos-specific context window spec.

Key Takeaways

• Claude Mythos Preview launched April 7, 2026 with the highest reported SWE-bench score (93.9%) and a 16-point lead over Opus 4.6 on CyberGym (83.1% vs 66.6%)
• Access is gated through Project Glasswing — a consortium of 12+ major technology companies plus 40+ additional organizations focused on critical infrastructure security
• Three cloud access points exist for authorized organizations: AWS Bedrock (US East), Google Vertex AI, and Microsoft Azure AI Foundry — all invitation-only
• Pricing for authorized partners is $25/$125 per million input/output tokens — approximately 3× Opus 4.6 pricing
• Claude Managed Agents (public beta, April 8, 2026) is available to all developers today — it's the infrastructure backbone that future Mythos deployments will run on, and worth adopting now
• Benchmark scores are capability ceilings, not guarantees — understand your task distribution before making model selection decisions based on headline numbers
• No public timeline exists for general Mythos availability — Anthropic's position is that broader access requires the industry to develop defensive infrastructure first