Anthropic has pushed its newest Mythos-class model onto the Claude Developer Platform, pairing a one-million-token context window with up to 128,000 output tokens and a reasoning mode that no longer switches off. The move, landing in the middle of June 2026 alongside Anthropic's "Code with Claude" developer event in Tokyo, is aimed squarely at the people building autonomous agents on top of Claude rather than at chat users — and it arrives bundled with a set of platform features designed to keep those agents running long after the developer has logged off.
The headline number is the context window. Both Claude Fable 5, the generally available Mythos-class model, and Claude Mythos 5, its less-restricted sibling, ship with a 1M-token context window by default and support up to 128k output tokens per request. According to Anthropic's model documentation, the two models share the same underlying weights; the difference is the safety layer. Fable 5 carries classifiers that can decline certain requests, while Mythos 5 omits them and is offered only in limited availability to approved customers through a program Anthropic calls Project Glasswing. Developers seeking Mythos 5 access are directed to contact their Anthropic, AWS, or Google Cloud account team.
What unifies the family is reasoning that is always on. "On Claude Fable 5 and Claude Mythos 5, thinking is always enabled and cannot be disabled," Anthropic states in its adaptive-thinking documentation, describing adaptive thinking as the only available mode. In practice that means every API call now budgets for internal reasoning tokens automatically, scaling the depth of deliberation to the difficulty of the task instead of forcing developers to toggle an "extended thinking" flag. Pricing for both models sits at $10 per million input tokens and $50 per million output tokens — a premium tier that signals these are meant for the most demanding reasoning and long-horizon agentic work.
A platform built to run agents unattended
The model is only half the story. The bigger shift for builders is a cluster of Developer Platform updates that turn Claude from something you call into something that runs on its own. Chief among them is the expansion of Managed Agents, Anthropic's hosted runtime for Claude plus the Model Context Protocol, which entered public beta with two notable additions during the Tokyo event on June 10.
The first is scheduled deployments. Developers can now attach an agent to a cron schedule on the Claude Platform; each time the schedule fires, the agent spins up a fresh session and completes its task. Teams can pause, resume, archive, or trigger extra runs on demand — effectively putting agents on autopilot for recurring work like nightly data validation or periodic monitoring. Browserbase, an early user, pointed a scheduled deployment at its public catalog of browser skills to validate the catalog on a recurring cadence.
The second is "vault" environment-variable credentials, a feature built to solve the awkward problem of giving an agent access to a service without handing it the keys. With vaults, a customer registers an API key together with an environment-variable name and an allowlist of domains the key may reach. Command-line tools running inside the agent's sandbox can then make authenticated calls through that variable — but the agent never sees the real secret. "The agent never sees the real key — the sandbox holds only a placeholder, and the actual secret is attached at the network boundary," Anthropic explains, with injection happening only on requests to domains the customer has explicitly approved. Notion used the mechanism to roll out its Notion CLI to agents while keeping sensitive tokens away from the model, satisfying its security team's guidelines.
Rounding out the release are richer session and thread webhook events. The platform's `session.thread_*` webhooks now carry a `session_thread_id` field, giving developers finer-grained visibility into long-running, multi-turn agent sessions — the kind of observability that becomes essential once agents are firing on a schedule with no human watching each run.
The long-context, long-running race
The launch lands in a market where the competitive fault lines have shifted from raw benchmark scores to who can keep an agent coherent and trustworthy over long stretches of work. A 1M-token window is no longer a differentiator on its own — Anthropic's own Claude Sonnet 4 reached that mark back in 2025, and both OpenAI and Google have pushed comparable or larger windows into their flagship offerings. What Anthropic is betting on is the combination: a huge context paired with always-on reasoning and, critically, a hosted runtime that handles scheduling, secrets, and session state so developers do not have to assemble that scaffolding themselves.
That positioning reads as a direct response to the agent-platform push from OpenAI's Responses and Agents tooling and Google's Vertex agent stack. Where rivals have emphasized model capability and developer SDKs, Anthropic is leaning into managed infrastructure — cron, credential vaults, sandboxed CLIs, and webhook telemetry — as the moat. The vault design in particular is a security pitch as much as a convenience one, aimed at enterprise teams that have been reluctant to let autonomous agents hold live API keys. By keeping the secret at the network boundary, Anthropic is trying to remove one of the last objections security reviewers raise before approving agents for production.
The split between Fable 5 and Mythos 5 is its own quiet statement. By reserving the unclassified Mythos 5 for vetted Glasswing customers while shipping the safety-wrapped Fable 5 broadly, Anthropic is trying to serve frontier-capability demand without loosening its public safety posture — a balancing act competitors will be watching closely.
What to watch next
The open questions are about adoption and cost. Always-on adaptive thinking improves output quality but spends reasoning tokens on every call, so the real-world bill for agents running on schedules could climb fast at $50 per million output tokens; expect developers to scrutinize how aggressively the model deliberates on routine tasks. Watch, too, whether Managed Agents graduates from public beta with enterprise guarantees around uptime and audit logging, and whether Project Glasswing access to Mythos 5 broadens or stays tightly gated. Finally, the proof of the vault-and-cron model will be in production incidents that don't happen — the measure of agent infrastructure is how rarely it makes the news.
"The agent never sees the real key — the sandbox holds only a placeholder, and the actual secret is attached at the network boundary."— Anthropic, Managed Agents vault documentation