Claude Fable 5.2 is now available on Bifrost

Anthropic just released Claude Fable 5.2, and Bifrost supports it from day one. It costs 20% less than Fable 5.1, does more at lower effort levels, and writes clearer reports after long agent runs.

If you already route Claude through Bifrost, change one model string to anthropic/claude-fable-5-2. Your routing rules, fallbacks, virtual keys, budgets, guardrails and logs keep working as they are.

Why Fable 5.2

Anthropic positions Fable 5.2 as a cheaper, sharper Fable 5.1 for agentic work. Per Anthropic's launch guidance:

• Lower price. $8 input and $40 output per million tokens, down from $10 and $50.
• More from low effort. At low, Fable 5.2 scores close to what Fable 5.1 did at higher effort levels. Good fit for fast subagents and high-volume traffic.
• Better long-run reports. Early testers rated its writing above Fable 5.1 and on par with Claude Opus 5.5 for clarity.
• Strong orchestrator. Early testers saw good results using Fable 5.2 to plan and direct work across other models.
• Self-review at high effort. At xhigh and max, it uses subagents to check its own work, which Fable 5.1 did not do on the same tasks.
• Built for async. Anthropic recommends it for long-running, unattended tasks.

A practical setup on Bifrost: send planning and the hardest agentic runs to Fable 5.2 at high, and route high-volume subagent calls to Fable 5.2 at low or to a cheaper model.

Pricing

Fable 5.2 is 20% cheaper than Fable 5.1 on both input and output.

Bifrost pulls model pricing from its pricing datasheet on a schedule, every 24 hours by default. To get Fable 5.2 rates into budgets and cost tracking right away, open Models > Model Settings in the Bifrost dashboard and click Force Sync Now.

Supported providers

Fable 5.2 is available through the Anthropic API and the three major clouds.

Add more than one provider to a virtual key and Bifrost load balances across them and fails over when one is down or rate limited.

What changes from Fable 5.1

Fable 5.2 is not a drop-in swap for every workload. Check these four things before you move traffic.

• Effort levels behave differently. Fable 5.2 supports low, medium, high (the API default), xhigh and max. Do not carry over your Fable 5.1 setting. Start at high, then re-test medium and your current setting. Use xhigh only for your hardest long-running work where you have measured a gain. Keep max for the rare task where time and tokens do not matter.
• Thinking is always on. Adaptive extended thinking cannot be disabled, and there is no thinking-budget parameter. Instructions like "answer directly without deliberating" no longer stop it from thinking at medium and above. Lower the effort level instead.
• Earlier thinking must be passed back unchanged. Fable 5.2 uses preserved thinking. If your harness edits the system prompt, the tools, or an earlier message between requests, the API returns a 400 error. Keep message history append-only, or use compaction or context editing. To drop invalid blocks instead, set thinking.block_binding.prefix_mismatch_behavior to "drop_block" with the beta header anthropic-beta: thinking-binding-controls-2026-08-01, and measure the extra output tokens first. Details are in Anthropic's preserved thinking docs.
• Refusals carry a category. A request declined by safety classifiers returns stop_reason: "refusal" with stop_details.category set to cyber, bio, frontier_llm, reasoning_extraction or general_harms. Remove any prompt text that asks the model to write out its reasoning in the response, since that now triggers reasoning_extraction.

Bifrost makes the rollout safe. Put Fable 5.1 and Fable 5.2 behind the same virtual key, shift traffic by weight while you re-run your evals, and compare cost, latency and output tokens per model in Bifrost's logs before you cut over.

Get started

Point the Anthropic SDK at Bifrost's /anthropic route and set the model string. The example sets effort explicitly and appends the assistant turn unchanged, so preserved thinking stays valid on the next request.

Prefer the OpenAI format? Call the same model through Bifrost's OpenAI-compatible endpoint.

New to Bifrost? Start the gateway with npx -y @maximhq/bifrost, add your Anthropic key or cloud credentials, and send your first request in under a minute.

Ready to build reliable AI applications?

Bifrost gives every team one gateway for Claude and 20+ other providers, with failover, budgets, guardrails and full observability built in. Book a demo or star the open-source repo on GitHub.