Anthropic Launches Claude Mythos 5, API Pricing Cut by More Than Half from Preview
Alina Collins
Anthropic released its dual-tier flagship — Claude Fable 5 for all users, Claude Mythos 5 for trusted users only — while cutting API prices over 50% from preview: $10 per million input tokens, $50 per million output tokens, a sharper commercial signal than any benchmark.
Why split one model into two versions?
Fable 5 and Mythos 5 share the same base model. The only difference is how tight the safety rails are: Fable 5 routes risky queries through a risk classifier — a filter that auto-detects dangerous prompts — while Mythos 5 ships with no filter at all.
This means → frontier AI models are entering a permission-tiered era: ordinary users get the safety-netted build; only vetted users access the unrestricted original.
In plain terms = same engine, but Fable has a speed limiter installed and Mythos does not — which one you drive depends on Anthropic's trust level for you.
Where exactly does Fable 5 outperform?
On SWE-bench Pro, the software-engineering benchmark, Fable 5 scored 80.3% versus OpenAI GPT-5.5's 58.6% — a gap of more than 20 percentage points.
On Cognition's Frontier Code evaluation, Fable 5 ranked first among all frontier models at medium reasoning intensity, with its edge growing as tasks got longer and more complex.
Anthropic's flagship real-world case comes from Stripe: Fable 5 completed a full migration across a 50-million-line Ruby codebase in one day — work that would take an engineering team over two months.
How big is the leap in vision?
On GDPpdf — a visual document reasoning benchmark, no external tools — Fable 5 / Mythos 5 scored 29.8%, above Opus 4.8's 22.5%, GPT-5.5's 24.9%, and Gemini 3.1 Pro's 16.7%.
Anthropic also demoed Fable 5 beating *Pokémon FireRed* autonomously, using only raw game screenshots with no map tools.
This reflects a shift in visual understanding — from "recognize what's in an image" to "reason through complex visual information." The game demo is a flex; the real commercial value sits in document processing and code review.
What does the API price cut really signal?
Both new flagships are priced at over 50% below the preview version: $10 per million input tokens, $50 per million output tokens.
This means → for developers building on the API, the cost barrier dropped even as performance jumped — a more immediate commercial signal than any benchmark score.
In plain terms = a stronger engine at a lower fuel price — Anthropic is using pricing to capture developer ecosystem share.
Where is the uncertainty around Mythos 5?
Anthropic describes Mythos 5 as having "the world's most elite cybersecurity and bioscience capabilities," with safety restrictions lifted in domains like network security.
Yet the company has not published separate standard benchmark scores for Mythos 5, leaving the market without independent verification of its actual capability boundary.
This reflects the core tension of the dual-version architecture: how to balance capability transparency with safety control — the stronger the model, the less its maker wants to disclose, yet the market needs exactly those details to price it.
Content is for reference only, not financial advice.