QWE AI Academy You’ve got a multi-hour agentic job that keeps stalling on Opus 5.5 – messy multi-repo migration, long research loop, or a design that needs real engineering judgment. Claude Fable 5.2 and Mythos 5.2 landed November 9, 2026. This is not another announcement recap. It’s the path from “I swapped the model string” to calls that actually finish without silent 400s, surprise retention blocks, or a melted token budget.
Anthropic aims both at the hardest open-ended technical work, not everyday chat. They share one underlying model (Anthropic’s launch post). Fable keeps strong cybersecurity and biology safeguards and is generally available. Mythos is the same weights with those dual-use safeguards relaxed, limited to orgs in the expanded Life Sciences and Cyber Verification Programs.
What Claude Fable 5.2 and Mythos 5.2 actually are
Fable is the public Mythos-class deployment; Mythos is the verified-only twin. Shared specs as of the Nov 2026 platform overview: 1M-token context, 128k max output, adaptive thinking always on, knowledge cutoff June 2026 (Claude Platform Fable 5.2 overview). Public API ID: claude-fable-5-2. Mythos shows up as claude-mythos-5-2 only after approval.
$8 per M input, $40 per M output. That’s the list price at launch – about 20% under 5.1’s $10/$50. Fast mode? $16/$80, with output tokens per second up to ~2.5× when it sticks. Anthropic still says: start on Opus 5.5 for most coding and knowledge work. Escalate when high-effort Opus evals miss. Fable can sit as Advisor other models ask, or Orchestrator that farms easy steps downward.
Benchmarks aren’t a contract. Still: on FrontierCode 2.0, Fable 5.2 at High roughly matches Opus 5.5 at Max for about half the cost per task; Xhigh hits 49.1% in Anthropic’s table. Outputs land up to 30% faster than 5.1. Cheaper tokens, less waiting, longer agentic runs – that combo is what people are stress-testing this week.
Practical setup: call Fable in under ten minutes
Already on the Claude API or Claude Code? Migration is mostly a model-string swap. One gate first: your workspace must allow the Covered Model retention rules, or the request fails hard (detail in limitations).
- Install or update the Anthropic SDK / Claude Code (
npm install -g @anthropic-ai/claude-codeorclaude update). - Set the model to
claude-fable-5-2(Bedrock:anthropic.claude-fable-5-2or the global inference ID your region docs list). - Leave sampling alone: omit
temperature,top_p, andtop_k. Forcedtool_choiceofanyor a specific tool returns 400 on this family. - Set effort explicitly in production so API vs app defaults don’t surprise you.
import anthropic
client = anthropic.Anthropic()
msg = client.messages.create(
model="claude-fable-5-2",
max_tokens=16000,
# effort lives under output_config on current APIs
output_config={"effort": "high"},
messages=[{
"role": "user",
"content": "Plan a zero-downtime cutover for service X; list risks, then a 5-step runbook."
}],
)
print(msg.content)In Claude Code, /model claude-fable-5-2 or claude --model claude-fable-5-2 pins a session. On Bedrock playgrounds, pick Fable 5.2 after the account’s data-retention mode allows Covered Models. Mythos stays application-based – if you’re not verified, the ID never lights up.
Effort isn’t a free “make it smarter” slider. Think inspection depth on a code review: low skims; high actually walks the risky paths; max reads every footnote and still might argue with itself. You pay for that depth in output tokens whether or not the extra pass helped.
Advanced usage: effort dial and fast mode
Adaptive thinking never turns off. You only steer depth with effort (effort docs). API default for Fable 5.2 is high. Claude apps often sit at medium – so the same prompt feels lazy in chat and sharp in API. Official line: start high; move to xhigh only after you measure lift on long-horizon agentic work; save max for ceiling tests; drop medium/low for high-volume subagents.
| Effort | Best for | Cost/latency note |
|---|---|---|
| low | Fast subagents, extraction, tight loops | Cheapest thinking budget |
| medium | Interactive coding, everyday knowledge work | Good balance; often enough |
| high (API default) | Hard coding, research, unattended runs | High already captures most of the jump |
| xhigh | Long multi-hour jobs with proven lift | Many more tokens; use sparingly |
| max | Open scientific/math ceilings | Slow, expensive; weak ROI on easy tasks |
Fast mode is a different lever. Messages API: speed: "fast" (fast mode docs). Double rates. Separate pool – watch anthropic-fast-*-remaining. No Batch API. No Priority Tier. Microsoft Foundry wants a -fast deployment ID and no speed field; mix them and it errors. Streaming is where the OTPS bump shows. Check usage.speed on the response so you’re not guessing.
Pro tip: Freeze one task. Run medium vs high vs xhigh. Log tokens and wall time. Anthropic notes gains above high are uneven; community threads already flag max sometimes overthinking into worse outcomes.
Honest limitations (read before you burn budget)
Call Fable from a pure zero-data-retention workspace and you get a 400 – not a soft warning. Fable is a Covered Model: prompts and completions stay retained about 30 days for safety review on platforms that offer it (Covered Models / retention). Exceptions need express authorization, or you turn retention on for that workspace. Bedrock usually needs aws_review (or legacy share mode) before invoke works. If legal wants permanent ZDR with no carve-out, this family stays blocked until your Enterprise Frontier Safeguards path covers it.
Cyber and biology classifiers on Fable still fire. Refusal-style stops show up; API and Bedrock often leave fallback to you – design for stop_reason instead of treating it like a crash. Community reports still hit friction there after launch. Mythos drops those dual-use classifiers (why the verification gate exists) and still isn’t a hobby jailbreak toy.
Token burn is the quiet budget killer. High effort plus long loops chew output at $40/M – or $80/M fast. Consumer-plan threads from launch week report weekly limits vanishing on Fable-heavy days. Don’t default every chat to Fable when Sonnet or Opus already clears the bar.
Is the “same model, two product names” split the durable pattern for frontier capability, or a bridge until safeguards catch up? From the outside it’s hard to know – the verification wall around Mythos suggests gated dual-use access is a deliberate bet, not a weekend experiment.
FAQ
Should I switch every workload to Claude Fable 5.2 today?
No. Opus 5.5 first for most coding and knowledge work. Escalate when high-effort Opus evals still fail on long-horizon jobs.
How do I get Mythos 5.2?
You don’t self-serve it. Apply through Anthropic’s Life Sciences or Cyber Verification Programs (or your account team / Claude Security on Enterprise). Same price and specs as Fable once approved; safeguard posture and who can call it change. A security team hunting CVEs or a lab doing binder design is the intended audience – not a weekend reverse-engineering side project.
Why did my Fable request return 400 with no useful completion?
Three common causes: (1) workspace still on zero data retention for a Covered Model, (2) you sent temperature/top_p/top_k or forced tool_choice any/tool, (3) client/config mismatch on fast mode (Foundry -fast ID vs speed: "fast"). Fix retention or sampling first. Fast-mode 429s are different – the fast pool is empty; drop speed or wait on retry-after.
Pick one stuck multi-step task tonight, pin claude-fable-5-2 at effort high, log tokens and wall time against your old Opus run, then decide whether xhigh or fast mode earns its keep on that workflow alone.