Claude Fable 5.2 and Mythos 5.2: Pricing, Access, and Cost per Task

Key Takeaways

• Same model, two doors: Claude Fable 5.2 and Claude Mythos 5.2 are one model, with Fable open to paid users and Mythos to vetted groups.
• List price: Both cost $8 per million input and $40 per million output tokens (small chunks of text), 20% below Fable 5.1 by our math.
• Real cost: Effort level (how hard the model thinks) and tokens per task set your bill more than the rate card does.
• Where to start: Anthropic recommends starting on Opus 5.5 and stepping up to Fable or Mythos only when the work demands it.
• What to measure: Track cost per completed task by team and feature, because a per-token price can't tell you what a result costs.

What Are Claude Fable 5.2 and Mythos 5.2?

Claude Fable 5.2 and Claude Mythos 5.2 are the same underlying Anthropic model for hard technical work, released November 9, 2026. Fable 5.2 adds cybersecurity and biology safeguards and is open to paid plans and the API, while Mythos 5.2 lifts them for vetted organizations.

Anthropic's launch announcement says it plainly: "Claude Fable and Claude Mythos are the same underlying model." What changes is who gets in and which tasks each version will take on. Anthropic says Fable is available to all users, with extra usage this launch week for subscribers on eligible paid plans.

What Fable 5.2 Is Built For

Fable 5.2 is aimed at long, open-ended technical work. Think agentic coding across large codebases, scientific research, financial analysis, and long-running agents that coordinate subagents.

An agent is an AI model that works through a task in steps, using tools and checking its own output. A subagent is a helper agent that a lead agent starts to handle one piece of a bigger job.

Anthropic's examples show the scale it has in mind. A team of 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in 30 hours. In another test, Fable 5.2 coordinated over 1,000 subagents for eight days to build a prototype operating system.

Both examples mean thousands of model calls running for hours or days. Fable 5.2 can also act as an Advisor that smaller Claude models consult, or as an Orchestrator that hands simple tasks to smaller models.

What Mythos 5.2 Adds, and Who Can Get It

Mythos 5.2 is the same model without the safeguards that block most cybersecurity tasks and research-level biology. It isn't a bigger or smarter model, and it won't beat Fable on general work.

Mythos 5.2 is built for security engineering, such as finding vulnerabilities and red teaming (attacking your own systems to find weak spots). It also suits life sciences work like biomolecular design. Anthropic reports Mythos 5.2 found 14.4% more known vulnerabilities than Mythos 5.1 on its internal CVE Rediscovery evaluation.

You can't just switch it on. Organizations apply through the Cyber Verification Program or the Life Sciences Verification Program, and Claude Enterprise customers can use Mythos 5.2 through Claude Security. Mythos fast mode is limited to Verification Program members.

Fable 5.2 vs. Mythos 5.2 vs. Opus 5.5 at a Glance

The table compares Fable 5.2, Mythos 5.2, and Opus 5.5 by access, safeguards, list price, context and output limits, and best fit.

The Claude Opus 5.5 model page lists input pricing at $4 per million tokens and output pricing at $20 per million tokens. The same page lists Opus 5.5 on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

That puts Fable 5.2 at twice the per-token price of Opus 5.5. For the full Opus rate card, see our Claude Opus 5.5 pricing breakdown.

How Much Do Fable 5.2 and Mythos 5.2 Cost?

Fable 5.2 and Mythos 5.2 share one price list. Every rate in this section applies to both.

The List Price and What Changed From 5.1

A token is a small chunk of text, often part of a word, and AI models charge by the token. Prices are quoted per million tokens, split into input (what you send) and output (what the model writes).

Anthropic lists Fable 5.2 at $8 per million input tokens and $40 per million output tokens. The Fable 5.1 announcement put Fable 5.1 at $10 per million input tokens and $50 per million output tokens. By our math, the 5.2 list price is 20% lower per token, though Anthropic itself only calls it "a lower price."

Anthropic also says 5.2 generates outputs up to 30% faster than 5.1. The context window, meaning how much text the model can hold in view at once, is 1M tokens, with up to 128K tokens of output.

A lower rate card does not guarantee a lower bill. If 5.2 uses more tokens on your tasks, the savings shrink or vanish. For how pricing looked at the first launch, see our original Fable 5 and Mythos 5 review.

Fast Mode, Batch, and Caching Rates

Three options change the per-token rate. Fast mode speeds up token generation by up to 2.5 times, per Anthropic, at double the standard price.

Batch means you send many requests together and accept slower results in exchange for a discount. Prompt caching stores a repeated part of your prompt, such as a long system prompt or a codebase, letting later calls reuse it cheaply.

The table compares Fable 5.2's standard, fast, batch, and cache rates per million tokens.

The Claude Fable 5.2 model page lists cache reads at $0.20 per million tokens and a 50% Batch API discount on input and output. The Microsoft Foundry announcement lists Fable 5.2 Fast at $16 input and $80 output, with batch marked "NA." You can't combine fast mode with batch.

Anthropic lists fast mode on Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry. Amazon Bedrock isn't on that list. Fast mode buys speed and saves nothing, while a cache read costs 40 times less than fresh input.

Why the Per-Token Price Isn't Your Real Cost

What you pay for a finished task is the price per token times the tokens used, plus everything around the model call. Orchestration, retries, evaluations, and the people reviewing output all sit outside the token line.

Tokenomics, the economics of token-based AI, splits this into production (your rate), consumption (tokens per task), and value (cost per finished task). Change one link and the next one moves.

Effort Levels Change How Many Tokens a Task Burns

The effort level is a setting that controls how hard the model thinks before it answers. Anthropic's adaptive thinking docs say the tokens Claude uses while thinking are billed as output tokens. On Fable 5.2, thinking is always on.

Medium is the default in Claude apps and Claude Code. Anthropic's API docs list high as Fable 5.2's default, which means the same prompt can cost more through the API.

In Anthropic's tests, Fable 5.2 matched Opus 5.5's best score for less per task on three benchmarks:

• FrontierCode 2.0, High effort: Fable 5.2 scored 45.3% against Opus 5.5's 45.5% for "roughly half the cost per task."
• WANDR, Medium effort: Fable 5.2 matched Opus 5.5's top score on Perplexity's data-collection benchmark for about 40% of the cost per task.
• Humanity's Last Exam, Low effort: Fable 5.2 matched Opus 5.5's top score for less than one-tenth of the cost per task.

Fable 5.2 costs twice as much per token as Opus 5.5. To cost half as much per task, it must finish with about one-quarter of the tokens, by our math. At 40% of the cost, it needs about one-fifth of the tokens.

These results compare Fable against Opus 5.5 at its best (Max) effort, on Anthropic's chosen benchmarks. Your workloads may not behave the same way.

Set effort per workload, and run your own evals (scored tests built from your real tasks) before you switch. Measure tokens per completed task at each effort level, then pick the cheapest level that clears your quality bar.

Blocked Queries Fall Back to Other Models

Fable 5.2's safeguards block high-risk cybersecurity work, like penetration testing and exploit generation, plus most research biology. Anthropic says blocked cybersecurity queries get an answer from Opus 4.8, and blocked biology queries from Opus 5.

That spend shows up under a different model. If you only watch Fable line items, you'll miss part of what a Fable-powered feature costs.

The AWS announcement for Fable 5.2 says that on Amazon Bedrock, your code has to send the retry itself. If the refused request hit the prompt cache, the refusal includes a fallback credit token. Send it with the retry and the cached input bills as cache reads.

Anthropic reports about 55% fewer blocks than Fable 5.1, based on early testing in claude.ai, Claude Code, and Cowork. API workloads may see a different rate.

Log every fallback and retry as its own flag, next to the team and feature that sent it. Our guide to tracking AI costs per feature covers how to tag requests at that level.

Multi-Agent Runs Multiply Token Use

Fable 5.2 is pitched for orchestrating many subagents at once. Each subagent carries its own context and burns its own tokens, which means a 1,000-agent run is 1,000 meters running at once.

Anthropic published no cost for its 1,000-agent runs. Plan as if fan-out, one task spawning many agents, is your largest swing factor.

Cap how many subagents can run in parallel, and set a token budget per run. Use Fable as the Orchestrator and route simple subtasks to smaller, cheaper models.

The subagent spike patterns in Claude Code show how fast parallel agents can run up a bill. For budgets and approval gates per agent, see our guide to agentic AI cost governance.

Data Retention and Safeguards Have Their Own Costs

A covered model is one that Anthropic requires to keep your data for safety monitoring. Zero data retention (ZDR) means the provider keeps no copy of your prompts or responses after it answers.

Fable 5.2 is a covered model. It requires 30-day data retention and isn't available under ZDR unless Anthropic expressly authorizes it. On Amazon Bedrock, Amazon personnel perform human safety review within the AWS boundary.

Enterprise Frontier Safeguards runs that monitoring on data stored in your own infrastructure, making it compatible with ZDR. Anthropic calls it a private beta, while the AWS and Microsoft launch posts say eligible customers can use Fable 5.2 with ZDR until it's available. Confirm access with your account team.

The Enterprise Frontier Safeguards announcement says your cloud provider bills you for that storage, plus reads, writes, and data egress fees. Anthropic doesn't charge for the safeguards themselves.

Those charges land on your cloud bill, far from your AI line items. Egress means fees for moving data out of a cloud, and it rarely gets tagged to the AI feature that caused it.

Where Can You Access Fable 5.2 and Mythos 5.2?

Fable 5.2 runs on every major platform, but the bill arrives in different places:

• Claude apps: Fable 5.2 is available on eligible paid plans, billed by Anthropic.
• Claude Platform API: Developers call the model as claude-fable-5-2, billed on the Anthropic invoice.
• Amazon Bedrock: Fable 5.2 runs through US Geo and Global cross-region inference profiles, plus Claude Platform on AWS.
• Google Cloud: Fable 5.2 runs through Claude Models on Google Cloud, in Anthropic-managed and Google-managed versions.
• Microsoft Foundry: Fable 5.2 runs on Foundry, including fast mode on the Anthropic-hosted version.
• Mythos 5.2: Access runs through the Verification Programs, and Foundry customers contact their Microsoft account team.

Prices can differ by platform and region, and this guide uses Anthropic's list price. Check each cloud's price page, and remember that cloud usage lands on your cloud invoice, apart from any direct Anthropic spend.

What Does This Mean in Practice?

The right choice depends on your quality bar, your latency needs, and the kind of work you run:

1. If Opus 5.5 already clears your quality bar: Stay there, because Fable 5.2 costs twice as much per token.
2. If Opus 5.5 at higher effort still falls short: Test Fable 5.2 at Medium or Low effort first, because lower effort burns fewer tokens.
3. If latency matters more than cost: Use fast mode at double the rate, and note that it isn't listed for Amazon Bedrock.
4. If the work isn't time-sensitive: Use batch at half the standard rate, because fast mode and batch can't combine.
5. If you run security or biology work: Expect blocks and fallbacks on Fable 5.2, and apply for Mythos 5.2 instead.
6. If you run multi-agent jobs: Set a budget per run before you turn Fable 5.2 loose, because fan-out multiplies tokens.

Your finance team will ask one question about any of these choices. What does one completed task cost us?

Track Cost Per Task Across Fable, Mythos, and Opus With Finout

Finout is an enterprise-grade FinOps platform for cloud and AI spend, built for the agentic era. Finout ties each Anthropic dollar to the team and feature behind it.

AI Cost Management brings your Anthropic spend into the same view as your cloud costs. Finout's Anthropic integration pulls in Claude Platform spend with a daily refresh and historical backfill. Input, output, and cache tokens sit next to every dollar, by model.

Billy answers plain-language questions, such as which team's Fable usage rose this week. Answers come from your live cost data, with read-only access that follows your permissions.

FinOps Agents take on the follow-up work. Detection Agent scans your AI and cloud spend for anomalies and waste. Investigation Agent traces each finding to its cause and its owner.

Finout's MCP server lets your own agents, including Claude, query the same governed cost data.

Under those AI tools, the core platform handles the bill itself:

• MegaBill: Shows Bedrock, Foundry, and direct Anthropic spend next to your cloud, Kubernetes, and SaaS costs.
• Virtual Tags: Delivers 100% cost allocation of Fable and Mythos spend by team and feature, without code changes.
• Anomaly Detection: Sends Slack or email alerts when a runaway agent run pushes spend past your threshold.
• Financial Plans: Budgets the step up from Opus 5.5 and tracks it against actual spend.
• Unit Economics Widget: Tracks cost per completed task on a dashboard over time.

Want to see what one Fable task costs your team? Book a demo