Anthropic dropped Claude Fable 5 on June 9, 2026, and it is already live inside Microsoft Foundry. That means Azure developers can now call a frontier-grade, long-horizon reasoning model through the same enterprise fabric they already use for OpenAI models, internal deployments, and agent orchestration. No separate vendor contract. No custom integration layer.
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s newest Mythos-class model, purpose-built for autonomous, multi-step work. Unlike earlier Claude models that needed babysitting on complex tasks, Fable 5 is designed to plan across stages, track dependencies, and route around blockers on its own. According to Microsoft’s model catalog, “Claude Fable 5 works independently for longer than any prior generally available Claude model.”
The raw specs are worth knowing. It has a 1 million token context window, up to 128,000 output tokens per response, and adaptive thinking that is always on (no extended-thinking toggle to remember). Anthropic prices it at $10 per million input tokens and $50 per million output tokens according to llm-stats. On internal benchmarks, Fable 5 completed equivalent coding work with fewer tool calls and lower token consumption than prior models, per the GitHub Copilot changelog.
Where it shows up
Fable 5 is now available across all Azure regions with consumption-based pricing through Microsoft Foundry according to coverage. It powers two specific surfaces you should care about.
- GitHub Copilot agents. You can select Fable 5 as the model behind Copilot coding agents that work autonomously across repositories, pull requests, and CI pipelines. Available since June 9, 2026.
- Foundry Agent Service. Microsoft’s managed agent platform now supports Fable 5 for long-running autonomous agents, including hosted agents with durable state, file system access, and scheduled routines per the Foundry blog.
The Foundry Agent Service supports background agents that run without a user token present, authenticating via Microsoft Entra ID and managed identity per the hosted agents documentation, with DLP policies and VNet integration for governance.
A note on the availability wobble
Here is something the marketing blogs downplay. Fable 5 was briefly pulled from availability shortly after launch, before being partially restored in late June for vetted organizations. The sibling model Mythos 5 (same architecture, different safeguard configuration) was cleared for roughly 100 vetted US organizations on June 26 per fable5.app. If you are building production workflows on this model, check current availability in your region and have a fallback model configured in your Foundry routing layer. Do not assume uninterrupted access during this stabilization window.
What to do this week
If you are an Azure developer or platform engineer, here is where to start.
- Test Fable 5 in the Foundry model catalog. Head to ai.azure.com and run your existing agent prompts against it. Compare output quality and token costs against whatever Opus or GPT model you are currently using.
- Set up model routing with fallbacks. Foundry’s Model Router already supports Claude models alongside OpenAI and Microsoft internal models per coverage. Configure a priority chain (Fable 5 first, Opus 4.7 second, or whatever fits your latency budget) so your agents do not hard-stop if Fable becomes unavailable.
- Migrate long-running agent workloads first. This model’s differentiator is sustained autonomy, not single-shot Q and A. Start with multi coding tasks, data pipeline orchestration, or compliance report generation where the agent needs to hold context and make judgment calls across many steps.
- Budget for the output token cost. $50 per million output tokens is steep compared to smaller models per pricing data. For tasks that generate a lot of code or verbose plans, this adds up fast. Monitor your Foundry consumption dashboard weekly at first.
- Try GitHub Copilot agent mode with Fable 5. If your org already uses Copilot, switch an agent workflow to Fable 5 and compare completion rates on multi-file refactors or test generation tasks.
What this means
Microsoft Foundry adding Fable 5 is not just another model listing. It signals that Microsoft is willing to put Anthropic’s most autonomous model inside the same platform enterprises use for governance, identity, and production deployment. For years, “agentic AI” meant stitching together fragile chains of prompts and hoping the model did not drift. A model that can genuinely sustain multi-step plans across hours, running inside a platform with managed identity, VNet isolation, and DLP policies, changes the calculus. You stop asking “can an agent do this?” and start asking “which tasks do I not need a human loop on anymore?” That is the actual shift. The follow-up question, the one that matters for your roadmap, is how many internal workflows you designed around human checkpoints can survive direct contact with a model that does not need them.