On this page
The Cromus Mythos Benchmark Framework
Measuring Claude Fable 5 & Mythos 5.1 (Capybara-tier) Workflow ROI
Access Update: Claude Fable 5 GA, Mythos 5.1 Trusted Access (September 1, 2026)
Claude Fable 5 reached general availability on June 9, 2026, while Claude Mythos 5.1 is a restricted trusted-access configuration of the same underlying model, both priced at $10 / $50 per million input/output tokens (roughly 2x Claude Opus 4.8). In its September 1, 2026 announcement, Anthropic said Mythos 5.1 access is limited to vetted trusted programs for cybersecurity and life sciences; it is not generally available through the public API.
The pricing and capability details below remain accurate for planning; Fable 5 is callable on the Claude API today, while Mythos 5.1 remains restricted to trusted-access programs and is not generally callable through the public API. Source: Anthropic's Claude Fable and Mythos 5.1 announcement.
This framework provides a methodology for measuring the workflow-level return on investment of Claude Mythos (Capybara tier) before and after deployment — independent of benchmark scores, API pricing sheets, or vendor claims.
It is designed for:
- Enterprise AI teams evaluating Mythos adoption
- Founders building on top of frontier models
- Operations leaders responsible for AI workflow spend
It is not a marketing document. It is a measurement tool.
What This Means for You
The Mythos 5.1 announcement is a textbook illustration of the point this framework keeps making: benchmark scores and token prices are worthless if you can't actually call the model. The Capybara tier is split between a globally available model (Fable 5) and an access-gated one (Mythos 5.1), with Mythos 5.1 limited to trusted programs for cybersecurity and life sciences rather than general API availability. Model availability is a variable, not a constant, and your workflow architecture has to treat it that way.
Concretely, this affects three groups:
- Teams that hard-coded Fable 5 or Mythos 5.1 need a continuity plan. Fable 5 is callable, but any pipeline that names a single model as a fixed dependency is one access-policy change away from breaking — and Mythos 5.1-dependent pipelines stay stalled unless the organization qualifies for a trusted-access program.
- Organizations that need Mythos 5.1 remain gated — Anthropic limits it to vetted trusted programs for cybersecurity and life sciences, not general API availability. This is a compliance and access-governance problem, not just an engineering one — it determines who can run a given workflow at all.
- Anyone who built ROI projections on Capybara-tier pricing should re-baseline whenever tier availability shifts, validating against models they can actually deploy today (Fable 5 now, or Claude Opus 4.8 as a fallback).
The defensible response is to stop architecting around a single frontier model and start treating the model tier as a swappable parameter. That is exactly what the four-step framework below is built for:
- Use Workflow Classification to confirm which steps genuinely require a top-tier model versus which were over-classified and can run on Frontier or Balanced tiers today.
- Re-run Baseline Cost per Workflow with a currently available model as the substitute, so you have a working plan rather than a stalled one.
- Score each substitution scenario in Croms™ so a forced model switch becomes a configuration change — not a rebuild.
If your SKILL.md keeps the model tier as a declared, governed parameter rather than a buried assumption, an event like this is an afternoon of re-pointing — not a fire drill. That portability is the whole point of measuring at the workflow level instead of the token level.
What Is Claude Mythos (Capybara)?
Claude Mythos (internally codenamed Capybara) is Anthropic's most powerful model tier, sitting above Opus. The tier now ships as two distinct models: Claude Mythos 5.1, restricted to trusted-access programs for cybersecurity and life sciences and not generally callable through the public API, and Claude Fable 5, generally available since June 9, 2026 as the flagship on the Claude API — both with dramatically higher scores on coding, academic reasoning, and cybersecurity tasks than Claude Opus 4.8.
Capybara is a new model tier — larger and more intelligent than Opus, and more expensive to serve and use. Both Capybara-tier models (Mythos 5.1 and Fable 5) share confirmed pricing of $10 per million input tokens and $50 per million output tokens ($1 per million cached input) — roughly 2x Claude Opus 4.8's $5 / $25 — with a 1M-token context window, up to 128K output tokens, and adaptive thinking always on.
For more details, see Anthropic's model overview.
The Wrong Question Everyone Is Asking
Every conversation about Claude Mythos starts with the same question: "How much will it cost per token?"
This is the wrong question. Token pricing tells you nothing about workflow ROI. A model that costs 5x more per token but eliminates 80% of human review time and 90% of error correction could be radically cheaper at the workflow level. Conversely, a model with impressive benchmarks could produce outputs that require more downstream processing, making it more expensive than the model it replaces.
The Right Questions: Workflow Cost, Not Token Cost
Token pricing is the wrong unit of analysis for enterprise AI deployment. The relevant unit is total cost per workflow outcome — including model API spend, human review time, error correction cycles, downstream rework, and opportunity cost of delayed outputs.
This framework measures cost at the workflow level using Croms — Cromus's unit of preventable AI workflow waste.
1 Croms = any execution cost, latency penalty, structural inefficiency, or failure risk that pre-execution intelligence would eliminate.
Every workflow is scored across four weighted dimensions of waste:
| Dimension | Weight | What It Measures |
|---|---|---|
| Cost Waste | 35% | Overspend from unoptimized model selection, redundant calls, and uncached operations |
| Latency Overhead | 25% | Execution time lost to sequential processing that could be parallelized |
| Failure Risk | 25% | Probability of runtime failures from missing validation, retries, and governance gates |
| Structural Gaps | 15% | Architectural inefficiencies in workflow design, modularity, and clarity |
The lower your Croms, the more efficient and governed your workflow.
What Are Croms?
Croms are Cromus's proprietary metric for measuring preventable AI workflow waste. Every workflow is scored, and the result tells you exactly how much waste is present and where it lives.
| Severity | Croms Range | Interpretation |
|---|---|---|
| Low | < 100 | Well-optimized workflow |
| Moderate | 100–299 | Room for improvement |
| High | 300–599 | Significant waste detected |
| Critical | 600+ | Urgent optimization needed |
Croms are calculated across four weighted dimensions: Cost Waste (35%), Latency Overhead (25%), Failure Risk (25%), and Structural Gaps (15%).
How Mythos Changes the Croms Calculation
A frontier model like Mythos creates a paradox: it may dramatically reduce failure risk and latency overhead (fewer errors, fewer retries, more capable first-pass outputs) while simultaneously increasing cost waste (~2x Opus 4.8 API pricing on tasks that don't require frontier capability).
The framework below gives you the methodology to measure whether Mythos actually reduces your total Croms or simply shifts waste from one dimension to another.
The Four-Step Framework
Step 0 — Workflow Classification
Before any model substitution analysis, map every AI touchpoint in your organization and classify by model tier.
Workflow Classification Matrix
For each AI-assisted workflow, assign a complexity tier:
| Tier | Description | Example Tasks | Representative Model |
|---|---|---|---|
| Mythos/Capybara | Breakthrough reasoning, novel multi-domain synthesis | Advanced cybersecurity analysis, complex research across domains, tasks where Opus output is measurably insufficient | Claude Mythos |
| High-Reasoning Frontier | Multi-step reasoning, synthesis across sources, novel problem-solving | Code generation, research synthesis, complex legal/medical analysis | Claude Opus |
| Mid-Tier Balanced | Structured generation, summarization, classification with nuance | Drafting, content creation, data interpretation | Claude Sonnet |
| Commodity Lightweight | Extraction, formatting, simple Q&A, classification with clear rules | Form parsing, tagging, translation, simple lookup | Claude Haiku |
Workflow Inventory Template
For each workflow, document:
| Field | Description |
|---|---|
| Workflow Name | Human-readable identifier |
| Complexity Tier | Mythos / Frontier / Balanced / Lightweight |
| Current Model | Model currently deployed |
| Volume | Calls per day / week / month |
| Average Input Tokens | Baseline prompt size |
| Average Output Tokens | Baseline response size |
| Human Review Rate | % of outputs requiring human review before use |
| Error/Rework Rate | % of outputs requiring correction or regeneration |
| Downstream Impact | What breaks or slows if this workflow underperforms |
Step 1 — Baseline Cost per Workflow
Calculate the true fully-loaded cost of each workflow today.
Baseline Cost Formula
Baseline Cost per Workflow =
Model API Cost
+ (Human Review Rate × Review Time × Hourly Rate)
+ (Error Rate × Correction Time × Hourly Rate)
+ Downstream Rework Cost
+ Prompt Engineering Overhead (amortized per execution)
+ Monitoring and Observability CostMost organizations track only the first line. The true baseline is typically 5–10x the API bill.
Croms Baseline Scoring
Score each workflow across the four weighted dimensions of waste:
- Cost Waste (35% weight) — Is the current model more capable than the task requires? Are there redundant calls, uncached operations, or unoptimized model selection?
- Latency Overhead (25% weight) — Is processing sequential where parallelization is possible? How much execution time is lost to avoidable wait states?
- Failure Risk (25% weight) — Are validation gates, retry handlers, and governance checks in place? What is the probability of runtime failure in this workflow?
- Structural Gaps (15% weight) — Is the workflow architecture modular and clearly designed? Are there inefficiencies in how steps are organized and connected?
Total Croms Score is the weighted output across all four dimensions. Severity bands:
| Severity | Croms Range | Interpretation |
|---|---|---|
| Low | < 100 | Well-optimized workflow |
| Moderate | 100–299 | Room for improvement |
| High | 300–599 | Significant waste detected |
| Critical | 600+ | Urgent optimization needed |
Mythos Substitution Scenarios
Once you have a baseline, model three scenarios for each Frontier or Mythos-candidate workflow:
Scenario A: Full Mythos Substitution
Replace the current model entirely with Mythos.
- Recalculate API cost using confirmed Fable 5 (Mythos-tier) pricing: $10 / $50 per million input/output tokens — roughly 2x Opus 4.8
- Estimate quality improvement: reduce human review rate and error rate by projected improvement factor
- Recalculate fully-loaded workflow cost
- Recalculate Croms score
Scenario B: Partial Mythos Substitution
Deploy Mythos only for the steps within the workflow requiring frontier capability. Route simpler steps to Sonnet or Haiku.
- Map workflow steps to capability requirements
- Assign model per step
- Recalculate blended API cost
- Estimate quality impact of hybrid routing
- Recalculate fully-loaded workflow cost and Croms score
Scenario C: Prompt Optimization + Current Model
Keep the current model. Invest in prompt engineering, workflow redesign, and output validation to reduce Croms generation without upgrading the model.
- Estimate engineering time investment
- Project quality improvement from optimization
- Recalculate fully-loaded workflow cost and Croms score
Step 2 — Croms
Score each workflow — and each substitution scenario — across the four weighted dimensions of waste.
| Workflow | Current Croms | Scenario A Croms | Scenario B Croms | Scenario C Croms |
|---|---|---|---|---|
| [Name] | X | X | X | X |
Decision Rule:
- If Scenario A produces the lowest Croms score AND the fully-loaded cost is acceptable: adopt Mythos fully
- If Scenario B produces the lowest Croms score: adopt Mythos selectively
- If Scenario C produces the lowest Croms score: optimize before upgrading
- If all scenarios produce similar Croms scores: the bottleneck is not the model
Step 3 — Total Cost of Workflow Ownership (TCWO)
The final metric for CFO-level decision-making.
TCWO Formula
TCWO = (Fully-Loaded Workflow Cost × Annual Volume)
+ Implementation Cost (migration, testing, prompt engineering)
+ Ongoing Monitoring Cost
- Downstream Value of Quality ImprovementTCWO Comparison Table
| Scenario | Annual API Cost | Annual Human Review | Annual Rework | Implementation | TCWO |
|---|---|---|---|---|---|
| Current State | $ | $ | $ | — | $ |
| Full Mythos | $ | $ | $ | $ | $ |
| Partial Mythos | $ | $ | $ | $ | $ |
| Optimized Current | $ | $ | $ | $ | $ |
Mythos may win on quality. It may lose on cost efficiency. TCWO is how you find out before you commit.
Full methodology: cromus.ai/total-cost-of-workflow-ownershipCromus.ai automates this framework for your entire AI stack.
Run Croms audits, simulate costs across 96 active verified generative models, compare substitution scenarios, and monitor workflow performance continuously — all from one platform. TypeSafe's decision model is tracked separately.
What Project Glasswing Tells Us
Anthropic's decision to gate Mythos behind Project Glasswing before its general release was the clearest signal yet that frontier capability rewards governance, not just budget. The organizations that earned early access shared one thing in common: documented workflow governance, measurable cost accountability, and the infrastructure to deploy responsibly.
Now that Fable 5 is generally available, that discipline is what separates orgs that capture its upside from those that simply inflate their bill. That's exactly what this benchmark measures: a low Croms™ score and a clean TCWO profile aren't just efficiency metrics — they're the profile of an org ready to deploy Capybara-tier capability without runaway cost.
Glasswing Readiness Checklist
Before rolling Capybara-tier capability (Claude Fable 5) into production, your org should have these governance layers in place. This checklist maps to the discipline that keeps a 2x-priced model from quietly inflating your AI spend.
Croms Monitoring Cadence
| Frequency | What to Measure |
|---|---|
| Daily | Error rate, output quality flags, human review volume |
| Weekly | Croms score per workflow, cost per outcome |
| Monthly | TCWO recalculation, model performance drift |
| Quarterly | Full workflow inventory review, model tier reassignment |
Trigger Conditions for Model Tier Reassignment
Trigger a workflow review when:
- Croms score increases by more than 50 points in a single week
- Human review rate increases by more than 10 percentage points
- Per-outcome cost increases by more than 15% without corresponding quality improvement
- A new model tier becomes generally available
Mythos-Specific Considerations
Confirmed at general availability (Claude Fable 5, June 2026):
Known
- Ships as two Capybara-tier models above Opus: Claude Fable 5 (GA flagship) and Claude Mythos 5.1 (restricted to trusted-access programs for cybersecurity and life sciences), at identical pricing
- Dramatically higher performance on coding, academic reasoning, and cybersecurity tasks than Opus 4.8
- Pricing: $10 / $50 per million input/output tokens ($1 cached input) — ~2x Opus 4.8
- 1M-token context window, up to 128K output tokens, adaptive thinking always on
- Generally available on the Claude API as of June 9, 2026
Still to confirm
- Production rate limits and sustained throughput
- Fine-tuning availability
- Enterprise agreement terms and volume discounts
- Regional availability and data-residency options
Croms Scoring Reference
| Severity | Croms Range | Interpretation |
|---|---|---|
| Low | < 100 | Well-optimized workflow |
| Moderate | 100–299 | Room for improvement |
| High | 300–599 | Significant waste detected |
| Critical | 600+ | Urgent optimization needed |
Croms are calculated across four weighted dimensions: Cost Waste (35%), Latency Overhead (25%), Failure Risk (25%), and Structural Gaps (15%).
Frequently Asked Questions
About This Framework
This framework is published by Cromus.ai as an open standard for workflow intelligence measurement. We believe the industry needs a shared methodology for evaluating AI model ROI at the workflow level — independent of vendor benchmarks and API pricing sheets.
Cromus.ai builds the infrastructure to run this analysis continuously, at scale, across an organization's entire AI workflow stack.
Workflow Intelligence as a Service.
Cromus.ai · Questions, feedback, or collaboration: connect via cromus.ai
Framework Version 1.1 — June 2026 (updated for the Claude Fable 5 general-availability release). Licensed for free use with attribution.
Find out if Fable 5 is worth it for your workflows.
Run your workflows through the Cromus demo and get your Croms™ score before you point a 2x-priced model at them.