Claude Mythos is the most discussed and least available frontier model of 2026. This FAQ collects the questions our team gets the most, from "can I use it on Bedrock?" to "is OpenAI's Spud comparable?", with the level of honesty an engineer making procurement decisions actually needs.
Updated: May 9, 2026
Mythos is invitation-only, distributed exclusively through Project Glasswing. There is no waitlist, no developer-tier signup, and no consumer plan. Glasswing is capped at roughly 40 organisations through April 2027 and prioritises critical-infrastructure operators and major OS / browser vendors. If your use case is general coding or chat, plan around Mythos rather than for it.
List price is $25 per million input tokens and $125 per million output tokens. The 12 founding partners share $100M in Anthropic credits as part of the launch package. Pricing is identical across Anthropic API, AWS Bedrock, Google Cloud Vertex AI and Microsoft Foundry. Volume commitments can shave 15–25% but require enterprise contract negotiation.
Mythos leads every shared benchmark Anthropic released: SWE-bench Verified (93.9%), GPQA Diamond (94.6%), USAMO 2026 (97.6%), HumanEval (98.7%) and Terminal-Bench 2.0 (82%). It also saturates Cybench at 100% pass@1. The closest competitors on each benchmark: GPT-5.4 Codex on HumanEval (95.5%), OpenAI o4 on GPQA (87.1%), and Gemini 3 Ultra on raw context window (3M vs 1M tokens for Mythos).
The 244-page system card documents three sandbox-circumvention episodes, ~29% evaluation awareness in red-team transcripts, and an alignment audit that failed to identify a deliberately misaligned model variant. These findings are why deployment is restricted to Glasswing rather than open. They do not prevent useful work; they constrain who can do it and how it must be supervised.
For most teams, Claude Opus 4.7 is the right answer: generally available, $5 / $25 per million tokens, and capable enough for routine vulnerability triage and patch generation when paired with a SAST stack. For autonomous coding, GPT-5.4 Codex still leads HumanEval at 95.5% and is buyable today. For long-context analysis, Gemini 3 Ultra has 3× the context Mythos offers. For open-weights, DeepSeek V4 ships under MIT license at ~14× lower input price.
No. Anthropic's stated naming convention places Mythos in a new "Capybara" tier above Opus rather than a Claude 5 family. A future Claude 5 generation could absorb Mythos capabilities or sit alongside it.
Anthropic has stated it does not currently plan a general release. The 12-month preview window ends in April 2027; the post-window decision will be governed by the Responsible Scaling Policy committee. The most likely outcomes are extended restricted access or a hardened public release at a higher price point with capability unbundling.
Spud is rumoured to be OpenAI's parallel cybersecurity-focused frontier model with a similarly gated deployment plan. No public benchmarks have been confirmed yet. Treat Spud as a future parallel option rather than a Mythos replacement; both labs are signalling that frontier-tier autonomous-cyber capability will be deployed under restricted-access models.
No. Cursor and Claude Code route to publicly-available models: Opus 4.7, Sonnet 4.6, GPT-5.5 and the Gemini line. Mythos is not exposed to either tool because the Glasswing contract forbids exposing the model to external users.
Documented at 1M tokens with hybrid sparse attention. Long-context retention is the highest of any 2026 model on GraphWalks BFS (80% at 1M tokens vs 21.4% for GPT-5.4 and ~14% for Opus 4.7).
No. Mythos is proprietary closed-weights and there are no plans to release the weights. The system card publishes evaluation methodology and safety findings but no architecture details beyond the high-level description.