🧭 Claude Fable 5.1 and Mythos 5.1 Launch: Cheaper Cache, Text Watermarks, and Tool-Choice Restrictions

Anthropic released Claude Fable 5.1 (claude-fable-5-1) and Claude Mythos 5.1 (claude-mythos-5-1) on September 1, 2026. Fable 5.1 is the successor to Claude Fable 5 for long-running agentic coding, knowledge work, and research; Mythos 5.1 is the companion model available to Project Glasswing participants. Both share the same pricing as their predecessors — $10 input / $50 output per MTok — with one significant cost improvement: prompt cache reads drop from 0.1× to 0.025× base input price, or $0.25 per MTok. On workloads that reuse large system prompts or reference corpora repeatedly, this can translate to a material reduction in effective per-task cost.

Key specifications

• Context window: 1M tokens by default
• Max output tokens: 128k
• Thinking: always-on adaptive thinking (no toggle required)
• Availability: Claude API · Amazon Bedrock · Claude Platform on AWS · Google Cloud · Microsoft Foundry
• Data retention: 30-day minimum required; not available under zero-data-retention unless expressly authorised by Anthropic

Breaking API changes you need to know

Two constraints apply to Fable 5.1 and Mythos 5.1 that do not apply to earlier Claude models:

• tool_choice types any and tool are not supported — both return a 400 error. Switch to auto or none. If you need schema-conformant tool inputs, use strict tool use or structured outputs instead.
• For accounts created on or after August 31, 2026: replaying thinking blocks after changes to the system prompt, tools, or message history returns a 400 error. Use the thinking-binding-controls-2026-08-01 beta header for advanced control over thinking-block reuse.

Content credentials and text watermarking

Both new models carry two new provenance signals. Text generated by Fable 5.1 and Mythos 5.1 embeds Anthropic's text watermark — a cryptographic signal detectable by verification tools but imperceptible to readers. Image and video files produced via the code execution tool and retrieved through the Files API carry C2PA Content Credentials, enabling downstream consumers to verify the AI-generated origin of the media. No changes to requests or response handling are required to benefit from either feature.

New thinking display option: "updates"

The thinking.display field now accepts a third value, "updates". When set, the API returns an empty thinking field plus progress updates as text during streaming. This gives streaming UIs a lightweight way to show "thinking in progress" without transmitting the full thinking transcript, reducing payload size on long reasoning steps.

🧭 Claude Code Self-Hosted Environments: Run Sessions Inside Your Own Network

Anthropic opened the public beta of self-hosted environments for Claude Code (August 6, 2026), giving organisations on Team and Enterprise plans the option to move Claude Code session execution off Anthropic's managed infrastructure and onto servers they provision and control. A single command — claude self-hosted-runner — turns your machines or containers into the compute layer where sessions run, whether started from the web, mobile, desktop, or a scheduled routine.

What stays on your infrastructure vs. what goes to Anthropic

• On your infra: repository checkouts, build artifacts, secrets, any files a session creates or modifies — these never leave your network.
• Sent to Anthropic: conversation data (prompts, responses, tool results) still traverse the Claude API for model inference. Self-hosting relocates execution, not inference.

Runner modes

Two deployment patterns are supported:

• Fixed runners: maintain a defined pool of always-on machines. Good for teams with predictable, steady-state session volume.
• On-demand runners: scale automatically by spinning up containers per session. Better for teams with bursty or unpredictable workloads — but requires an orchestrator component that your team owns.

# Install and configure the self-hosted runner
claude self-hosted-runner install \
  --org-id  your-org-id \
  --token    RUNNER_TOKEN \
  --mode     on-demand        # or: fixed

# Runner registers with Anthropic; sessions routed to it automatically
# Sessions start from any Claude Code client (web, desktop, mobile)

Self-hosted environments are disabled by default and unavailable for organisations that have Zero-Day Risk (ZDR) protection enabled.

🧭 Claude Sonnet 5 Introductory Pricing Is Now Permanent — No September Hike

When Claude Sonnet 5 launched in June 2026, Anthropic labelled its $2 input / $10 output per MTok rate as "introductory pricing" with a planned increase to $3/$15 per MTok taking effect on September 1, 2026. On August 10, Anthropic quietly confirmed that the introduction is now the standard: the September 1 price hike will not occur. The $2/$10 rate is permanent.

What this means for your cost projections

• Any budget model that assumed a 50% input and 50% output price increase on September 1 can now be revised — you're staying at $2/$10.
• Sonnet 5 retains its position as the best-value high-capability model in the Claude family: cheaper than Opus 5 ($15/$75) and Fable 5/5.1 ($10/$50), with a larger effective output budget per dollar than either.
• Teams weighing Sonnet 5 vs. Fable 5.1 for agentic workloads: Sonnet 5 is 5× cheaper on input and output per token; Fable 5.1 offers 1M-token context and always-on adaptive thinking. Combine them in a tiered pipeline — Sonnet 5 for classification and draft generation, Fable 5.1 only for the reasoning-heavy steps where it earns its cost.