Why trust Cekura on this: we test voice and chat agents for a living, across simulated and real production calls, with native integrations for Retell, VAPI, LiveKit, Pipecat, and other major voice stacks.
This is a launch-day read of Claude Fable 5.2 from that seat: what it changes for agents, and what to check before you ship it.
Anthropic released Claude Fable 5.2 on November 9, 2026, and the first question on every agent team's mind is: does it belong in my stack, and what breaks if I switch?
It's Anthropic's top generally available model at $8 and $40 per million tokens, it's genuinely better at agentic work, and it is not the model you want answering every turn of a live phone call.
Here's where it fits, what changed, and how to prove it on your own calls before you move.
Every number below comes from Anthropic's launch post, the Fable 5.2 model page, and the fast mode docs, pulled on launch day. Re-verify anything load-bearing against your own account before you commit.
Claude Fable 5.2 at a Glance
-
Released: November 9, 2026. Model ID claude-fable-5-2.
-
Built for: "demanding reasoning and long-horizon agentic work," per Anthropic's model page. Think hours-long agent sessions and multistep research.
-
Price: $8 input, $40 output per million tokens. Cache reads are $0.20, five-minute cache writes $10, one-hour cache writes $16. Batch runs 50% off.
-
Fast mode: $16 input, $80 output, up to 2.5x higher output tokens per second.
-
Context: 1M tokens in, up to 128K out. Text and images in, text out.
-
Latency label: "Moderate," the same tier Anthropic gives Claude Opus 5.5. Sonnet 5.5 is "Fast," and Haiku 5.5 is "Fastest."
-
Thinking: adaptive and always on. You can't turn it off; you steer its depth with the effort setting, which defaults to high on the API.
-
Data retention: 30 days required for safety monitoring. No zero data retention unless Anthropic expressly authorizes it.
-
Platforms: Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.
Claude Mythos 5.2 is the same underlying model with fewer cybersecurity and biology safeguards, available only to organizations in Anthropic's Cyber and Life Sciences Verification Programs.
Is Fable 5.2 a Good LLM for Live Voice Calls?
Not as the model that answers every turn.
Anthropic's own advice is to start most work on Claude Opus 5.5, then move up to Fable for the hardest, most open-ended problems or down to Sonnet 5.5 and Haiku 5.5 for simpler work. For a live voice agent, most turns are simple and limited by speed, which puts them in that second group.
Three details decide it.
What Moderate Latency Means on a Call
In a cascaded pipeline, a caller's words go speech-to-text, then to the LLM, then to text-to-speech. The LLM's time to first token lands right in the middle of that silence while the caller waits.
Fable 5.2 carries a "Moderate" latency label, the same as Opus 5.5, which is fine for a chat box and long for a phone line. Our voice AI latency guide breaks down where that budget actually goes.
Fast Mode Speeds Up the Wrong Number for Short Turns
Fast mode helps, but read the fine print. The fast mode docs say its gains are "focused on output tokens per second (OTPS), not time to first token (TTFT)."
A voice turn is usually a sentence or two, so how fast tokens stream out matters less than how long the caller waits for the first one. Fast mode pays off on long streamed answers. On a crisp "Sure, what day works?" it barely moves the needle.
Thinking Adds Time Before the First Word
Adaptive thinking is always on, and the API default is high effort. The model reasons before it speaks, which is exactly what you want for a hard decision and exactly what you don't want when a caller is waiting on hold.
So where does Fable 5.2 earn its seat in a voice or chat stack?
-
As an Advisor. Anthropic lets Fable 5.2 consult for Haiku, Sonnet, or Opus models on harder problems. In a voice agent, that suits the moments where a short pause is expected anyway, like a complex eligibility check or a billing dispute.
-
As an Orchestrator. Fable plans multistep work and hands the simple parts to smaller models. Good for back-office agents that file, schedule, and follow up after the call ends.
-
For chat agents. People tolerate a longer wait in a chat window than on a phone line, so a Fable-backed chat agent is far more realistic than a Fable-backed voice turn.
-
Offline, around the call. Post-call analysis, LLM-as-a-judge scoring, writing test scenarios, and reviewing prompts all run off the live path, where Moderate latency costs you nothing.
If we were putting Fable 5.2 anywhere near a live call, it would be behind a faster model, triggered only on the turns that actually need it.
What Changed From Fable 5.1
If you're already on Fable 5.1, the upgrade is cheaper and noticeably stronger on agent tasks.
| Dimension | Fable 5.1 | Fable 5.2 |
|---|---|---|
| Input/output per 1M | $10/$50 | $8/$40 |
| Cache reads per 1M | $0.25 | $0.20 |
| Output speed | Baseline | Up to 30% faster |
| Fast mode | Not offered | $16/$80, up to 2.5x OTPS |
| Safety blocks (Anthropic's early testing) | Baseline | About 55% fewer |
| FrontierSWE v2 (agentic coding) | 56.3% | 69.9% |
| AutomationBench (business workflows) | 31.4% | 51.4% |
| OSWorld 2.1, partial (computer use) | 83.5% | 89.3% |
$8 and $40 is 20% below Fable 5.1's $10 and $50, per Anthropic's migration guide. Benchmarks come from Anthropic's own launch table.
Testers quoted by Anthropic say Fable 5.2 "says it doesn't know things instead of hallucinating," and Anthropic's behavioral audit found lower rates of hallucinated inputs, false claims of finished tasks, and undisclosed lazy work than on Opus 5.5, Sonnet 5.5, and Mythos 5.1.
For an agent that books appointments or quotes a balance, a model that admits it lacks data beats one that invents a time slot. That's also the exact kind of claim worth confirming against your own traffic before you trust a launch post.
Benchmarks That Map to Agent Work
Anthropic's table compares Fable 5.2 against Fable 5.1, Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon, at each model's highest effort. These are Anthropic's self-reported numbers.
| Benchmark | What it tests | Fable 5.2 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| AutomationBench | Business workflows across apps | 51.4% | 42.5% | 41.4% | 53.9% |
| WANDR (relaxed) | Agentic search, data collection | 82.1% | 72.3% | n/a | n/a |
| OSWorld 2.1 (partial) | Computer use | 89.3% | 87.2% | 79.5% | 81.5% |
| FrontierSWE v2 | Agentic coding | 69.9% | 62.3% | 65.5% | 61.2% |
None of these is a voice benchmark. They measure task completion, but don’t measure how an agent sounds, how long a caller waits, or how it recovers from an interruption. AutomationBench is the closest to what a back-office agent does, and Gemini 4 Carbon still edges it out.
That difference is the reason we test agents on real and simulated calls, where sound, timing, and recovery actually show up.
Our voice AI benchmarks score speech-to-text, text-to-speech, and agent platforms on reliability, response time, and interruption handling, On the Medicare workflow alone, pass³ ranged from 65.2% to 95.7% across six platforms, a 30-point spread.
Six API Changes That Can Break a Working Agent
Swap claude-opus-5-5 or claude-fable-5-1 for claude-fable-5-2 and change nothing else, and these are what bite first. All six come from Anthropic's model page and docs.
| Change | What breaks | What to do |
|---|---|---|
| Zero data retention returns a 400 | Fable 5.2 requires 30-day retention unless Anthropic authorizes otherwise, so an org on zero data retention gets a 400 on every request. | Confirm your retention setup before the swap. |
| Sampling params return a 400 | Anthropic says to omit temperature, top_p, and top_k, and some values return a 400. | Strip all three before the swap, including from shared request templates, or every call that still sends them will error. |
| Forced tool choice returns a 400 | Setting tool_choice to any or tool returns a 400 on Fable 5.2, so an agent that forces a lookup or handoff tool on a turn fails outright. | Switch to auto, say in the prompt when each tool must be called, and test that it still gets called. |
| Preserved thinking applies to every account | Editing or trimming earlier turns invalidates the model's thinking. Integrations that shorten voice history to save tokens are exposed. | Audit anything that rewrites conversation history mid-session. |
| Refusals fall back to another model | A blocked request comes back as a refusal. If you opt in to server-side fallback, it can be retried on Opus 4.8 or Opus 5 instead. | Log which model actually answered, and handle classifier refusals on the API. |
| Default effort is high | Opus 5.5 defaults to medium. A straight swap changes how long the model thinks, and with it your latency and cost. | Set effort deliberately per workload instead of taking the default. |
Anthropic’s migration guide lists a few more that only affect some setups: a Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor, the computer_20251124 tool returns a 400 outside Bedrock, and structured outputs aren’t supported in CMEK organizations.
A model swap changes response time, tool behavior, and failure modes all at once. That's why testing comes next.
Data Retention: Check This Before Healthcare or Finance Calls
Fable 5.2 uses 30-day data retention for safety monitoring by default. Anthropic monitors a window of usage data for its most capable models, and Fable 5.2 is on that list.
Eligible customers can keep using Fable 5.2 with zero data retention before they turn on Enterprise Frontier Safeguards. For everyone else, retention is 30 days.
Enterprise Frontier Safeguards, now in private beta with customers like Bridgewater, Salesforce, and Stripe, runs that monitoring on data stored in the customer's own infrastructure and is compatible with zero retention.
For an agent that handles patient records or payment details, that can decide the model on its own. If your stack requires zero retention and you aren't yet eligible for Enterprise Frontier Safeguards, Fable 5.2 stays off the live path until you are.
Treat HIPAA as a separate check. Anthropic's pages document retention and zero-retention eligibility, so confirm the terms in your own agreement with Anthropic before routing protected health information to any model.
Our guide to data privacy best practices for AI voice models covers the rest of the checklist.
Mythos 5.2 and Voice Agent Security
Mythos 5.2 is Anthropic's most capable cybersecurity model. On its internal CVE Rediscovery evaluation, Anthropic says Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1.
It’s available through Anthropic’s Cyber and Life Sciences Verification Programs, and Claude Enterprise customers can reach it through Claude Security.
Fable 5.2 goes the other way and blocks most security work, including penetration testing. If you red-team your voice agent with an LLM, expect Fable to refuse that job, unless you’ve set up fallback to another model.
Cost: What a Fable-Backed Agent Pays
Fable 5.2 costs twice what Opus 5.5 does per token, $8 and $40 against $4 and $20, with the same $0.20 per million for cache reads. Sonnet 5.5 runs $2 and $10, and Haiku 5.5 starts at $0.10 and $0.50.
Here's a rough turn with made-up but realistic token counts: a reply with 1,000 cached input tokens, 200 fresh input tokens, and an 80-token answer costs about half a cent on Fable 5.2.
That’s before thinking. Fable 5.2 always reasons first, so price your turns with the thinking tokens your own traffic produces.
The same turn on Haiku 5.5 costs a tiny fraction of that. Run a thousand of those turns a day across a few agents and the difference is your monthly bill.
Anthropic's effort charts show where Fable can still pay off. At High effort, it roughly matches Opus 5.5's best FrontierCode 2.0 score for about half the cost per task, and at Medium it matches Opus 5.5's top WANDR agentic-search score for about 40% of the cost.
The cheaper pattern for most agents is a split: a fast, low-cost model answers each turn, and Fable 5.2 handles the occasional hard decision or the after-call work. Price your real traffic both ways before you choose. The arithmetic rarely favors putting your most expensive model on your most frequent call.
How to Test Fable 5.2 Before You Switch Your Agent
Swapping the LLM in a production agent changes more than answer quality. It moves response time, script adherence, when the agent calls tools, and how it fails.
Our guide to testing new LLM model versions with real production calls walks through the full method, but the short version is five steps.
-
Freeze a baseline. Run your current model on a fixed scenario set and record pass rates, tool-call accuracy, and response time.
-
Fix the six API changes in a branch, then run Fable 5.2 at Medium and at High effort.
-
Compare on the call metrics that count: task completion, tool-call correctness, hallucinated details, interruption handling, and time to first audio.
-
Test the split setup too, with Fable 5.2 as an Advisor behind your current model, triggered only on hard turns.
-
Watch production after rollout for at least a week, scoring live calls for drop-offs and workflow adherence.
This comparison is what Cekura is built to run. With Cekura, teams can:
-
Run pre-production simulations: fire thousands of caller scenarios against the old and new LLM setups and compare the results side by side.
-
Stress the infrastructure: replay interruptions, background noise, and poor audio, where slow responses and missed turns show up first.
-
QA production calls: score live traffic for drop-off points, sentiment, workflow adherence, and compliance failures.
-
Track tool and refusal behavior: catch forced-tool-use errors and unexpected fallbacks across model versions.
-
Regression-test every change: rerun the full suite on each prompt swap or model update before it ships.
Native integrations work out of the box for Retell, VAPI, ElevenLabs, LiveKit, Pipecat, Bland, and more. Cekura adds a testing and monitoring layer on top of the stack you already run.
Cekura supports SOC 2, HIPAA, and GDPR compliance, covering transcript redaction, role-based access, and audit trails.
Book a demo to test Claude Fable 5.2 on your own calls.
Frequently Asked Questions
What Is Claude Fable 5.2?
Claude Fable 5.2 is Anthropic's top generally available model, released November 9, 2026, for demanding reasoning and long-running agent work.
It costs $8 per million input tokens and $40 per million output tokens, has a 1M-token context window, and runs on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.
Is Claude Fable 5.2 Good for Voice Agents?
Claude Fable 5.2 is better suited to planning, hard decisions, and after-call work than to answering every turn of a live call.
Anthropic lists its latency as "Moderate," and its fast mode raises output tokens per second, not time to first token. Run it against your real call traffic before you put it in the live path.
What Is the Difference Between Claude Fable 5.2 and Mythos 5.2?
Claude Fable 5.2 and Mythos 5.2 are the same underlying model at the same price. Fable 5.2 has strong cybersecurity and biology safeguards and is available to everyone, while Mythos 5.2 is limited to organizations in Anthropic's verification programs.
Does Claude Fable 5.2 Support Tool Calling?
Claude Fable 5.2 supports tool use, but forced tool use (setting tool_choice to any or a specific tool) returns a 400 error. Agents that force a function call need to change that pattern before switching.
Is Claude Fable 5.2 HIPAA Compliant?
Claude Fable 5.2 uses 30-day retention for safety monitoring unless your organization is eligible for zero data retention. Check your own agreement with Anthropic for HIPAA terms before routing patient data to it, since retention rules decide this as much as the model does.
How Much Does Claude Fable 5.2 Cost Per Call?
Per-call cost depends on your prompt and response length. At list prices, Claude Fable 5.2 runs $8 per million input tokens and $40 per million output tokens, twice Opus 5.5, with cache reads at $0.20 per million that help agents resending the same system prompt each turn.