Claude Fable 5 explicado: benchmarks, preços e a diferença para o Mythos 5
A Anthropic lançou o Claude Fable 5 em 9 de junho de 2026 — um modelo da classe Mythos tornado seguro para uso geral. Analisamos os benchmarks de lançamento, o fallback de segurança para o Opus 4.8, os preços e a comparação com GPT-5.6 e GPT-6 Astra.
Claude Fable 5 Explained: Anthropic's Mythos-Class Model Made Safe for General Use
On June 9, 2026, Anthropic launched Claude Fable 5 alongside Claude Mythos 5. Both are Mythos-class models — but Fable 5 is the first one Anthropic has made broadly available, with a new layer of safety classifiers wrapped around the same underlying weights. It is the most capable Claude model the company has released to everyone.
This explainer breaks down what Fable 5 actually is, how it differs from Mythos 5, what the launch benchmarks say (and where they need caveats), how the safeguard fallback works, and what it costs. All figures are attributed to their sources and were checked on November 20, 2026.
What Is Claude Fable 5?
Claude Fable 5 is a Mythos-class model that Anthropic made safe for general use. You call it through the Claude API with the model id claude-fable-5. Anthropic positions it a tier above the Opus class: it launched at double the price of Claude Opus 4.8, and reports state-of-the-art results on nearly all tested capability benchmarks — with especially strong results in software engineering, knowledge work, vision, and scientific research.
The name encodes the safety story. Fable comes from the Latin fabula, “that which is told,” akin to the Greek mythos. Fable 5 and Mythos 5 use the same underlying model; they differ in how safeguards are applied and who can access each version.
Fable 5 vs Mythos 5: Same Weights, Different Guardrails
| Dimension | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Underlying model | Mythos-class | Same Mythos-class weights |
| Access | Generally available (API, Claude apps, Claude Code, major clouds) | Restricted: Project Glasswing trusted partners, with staged expansion |
| Safeguards | Blocking classifiers for cyber, bio/chem, and distillation queries | Some safeguards lifted for approved use cases |
| Pricing (per 1M tokens) | $10 input / $50 output | Same |
Mythos 5 is deployed through Project Glasswing, a trusted-access program run with the US government for cyberdefense and critical-infrastructure security. One number worth clarifying: the 78.0% ExploitBench score quoted at launch belongs to Mythos 5, not Fable 5. Anthropic has not published a Fable 5-specific ExploitBench figure — if you see 78% attributed to Fable 5, that is the Mythos number.
The Launch Benchmarks (Read the Footnotes)
Anthropic's launch table compares the Mythos 5 / Fable 5 pair against Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. Selected rows:
| Benchmark | Mythos 5 / Fable 5 | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|---|
| SWE-Bench Pro (agentic coding) | 80.3% | 69.2% | 58.6% |
| FrontierCode Diamond (xhigh) | 29.3% | 13.4% | 5.7% |
| GDPval-AA (knowledge work) | 1932 | 1890 | 1769 |
| OSWorld-Verified (computer use) | 85.0% | 83.4% | 78.7% |
| Humanity's Last Exam (no tools)* | 59.0% | 49.8% | 41.4% |
| Terminal-Bench 2.1* | 88.0% | 82.7% | 83.4% |
Two caveats matter. First, the table shows the higher of the Mythos 5 / Fable 5 pair — the two usually land within 1–3 points of each other. Second, starred (*) rows are where Fable 5's blocking safeguards kick in and performance falls back closer to Opus 4.8. For Terminal-Bench specifically, independent evaluators place Fable 5 between 83.4% and 84.3%, below the starred headline.
The durable finding across sources: Fable 5 leads on repo-level coding. When OpenAI shipped GPT-5.6 a month later, its own eval tables still showed Fable 5 ahead on SWE-Bench Pro (80% vs 64.6% for GPT-5.6 Sol). More on that split below.
The Safeguard Fallback: A Feature, Not a Refusal
Fable 5 ships with separate classifier systems that detect potential misuse — including jailbreak attempts — before the main model responds. When a request touches cybersecurity, biology/chemistry, or distillation, the response is automatically handled by Claude Opus 4.8 instead, and the user is told when this happens.
- More than 95% of Fable sessions involve no fallback at all (Anthropic's early data). In those sessions, Fable 5 performs effectively the same as Mythos 5.
- False positives on benign requests occur in under 5% of sessions; Anthropic says it will tune this down post-launch.
- Under automated red-teaming, offensive cyber task completion dropped from 56.6% on Opus 4.8 to 5.4% on Fable 5.
- An external bug bounty produced no universal jailbreaks across 1,000+ hours of testing.
Alongside the safeguards, Anthropic introduced a 30-day data-retention policy for Fable 5, Mythos 5, and future models at similar capability levels: traffic is retained for 30 days, not used to train new Claude models, with logged human access and deletion afterward in almost all cases.
Pricing and Availability
| Surface | Availability at launch (June 9, 2026) |
|---|---|
Claude API (claude-fable-5) | Fully available; $10 input / $50 output per 1M tokens |
| Consumption-based Enterprise | Fully available |
| Pro / Max / Team / seat-based Enterprise | Included at no extra cost through June 22; usage credits required from June 23 |
| Claude Code & major cloud platforms | Available at launch |
At $10/$50, Fable 5 costs double Claude Opus 4.8 ($5/$25) — premium-tier pricing for frontier capability. One update worth noting: by September 2026, our GPT-6 Astra review listed Claude Fable 5.1 as Anthropic's current flagship, matching the $10/$50 pricing with cached input at $0.25 per 1M tokens and SWE-bench Pro at 81.2%. If you are evaluating the family today, check which iteration your surface actually serves.
Choosing a Reasoning Effort
Fable 5 exposes configurable reasoning effort, and the cost curve is not linear. Appwrite's CursorBench analysis found:
- Low (64.2%) is already roughly on par with GPT-5.5 at its highest effort for code writing.
- Low → Medium is the meaningful jump: 64.2% → 69.8%.
- Medium → Max adds little score while cost climbs quickly.
The practical default for most developers is Medium effort — high scores with costs under control. Reserve Max for genuinely hard, long-horizon tasks.
Fable 5 vs GPT-5.6 and GPT-6 Astra: The Frontier Is Split
A month after Fable 5 launched, OpenAI shipped GPT-5.6 Sol, Terra, and Luna (July 9, 2026). The head-to-head is not a sweep in either direction:
- Fable 5 leads repo-level coding: SWE-Bench Pro 80% vs Sol's 64.6% — a gap OpenAI's own tables show, alongside its claim that ~30% of SWE-Bench Pro tasks are broken.
- GPT-5.6 Sol leads agentic work: Terminal-Bench, BrowseComp, OSWorld 2.0, and Agents' Last Exam.
- GPT-6 Astra (September 3, 2026) succeeded Sol and pushed reasoning and cybersecurity further — but independent indexing still scored it level with Sol. See our Astra review for the system-card details.
Anyone telling you one lab swept the board is selling something. Match the model to the workload.
What Fable 5 Means for Creators
Fable 5 is not an image or video model — it is a reasoning engine that sits upstream of creative pipelines. Three launch findings matter for creators:
- Vision state-of-the-art: it extracts precise numbers from scientific figures and can rebuild a web app's source from screenshots alone — useful for storyboard-to-brief and reference-frame analysis workflows.
- Long-context memory: Anthropic reports it stays focused across millions of tokens and improves outputs using its own notes — production bibles and multi-episode outlines fit in a single session.
- Long-horizon autonomy: early testers reported codebase-wide migrations compressed from months to days. The same persistence applies to multi-step creative agent runs — script, storyboard, prompt iteration — feeding tools like our AI video generator and AI image generator.
Sources and Verification
Figures above were checked on November 20, 2026 against:
- Appwrite: Anthropic just launched Claude Fable 5 and Claude Mythos 5 (June 9, 2026) — launch table, safeguards, pricing, availability.
- Vellum: GPT-5.6 Sol vs Terra vs Luna (August 3, 2026) — cross-vendor benchmark comparison, ExploitBench attribution, independent Terminal-Bench range.
- DataCamp: Claude Opus 5 vs Fable 5 (July 29, 2026) — June 9 release date and Mythos-class positioning.
- Internal: GPT-6 Astra Review (September 4, 2026) — Fable 5.1 pricing and SWE-bench Pro figures.
Vendor benchmarks are claims, not guarantees; starred and paired scores are labeled as such above. This article is a documentation summary, not a hands-on benchmark.
Perguntas Frequentes
What is Claude Fable 5?
What is the difference between Fable 5 and Mythos 5?
How much does Claude Fable 5 cost?
What happens when Fable 5's safeguards trigger?
Is Claude Fable 5 better than GPT-5.6 Sol?
Continue lendo
A equipe editorial do ZNIX explica modelos de IA com fontes primárias, verificações de disponibilidade datadas e exemplos identificados. Resumos da documentação são diferenciados de testes práticos.