Claude Fable 5 Real-World Test Shows Bigger Lead on Complex Tasks (but pricey)

Anthropic announced two new models, Claude Fable 5 and Claude Mythos 5, on the same day. Both share the same underlying model; Fable 5 is the safety‑filtered Mythos‑class version for general users, while Mythos 5 removes some safeguards for a limited set of partners in the Project Glasswing program.

The company wraps the base model with an independent classifier that redirects requests in the domains of cybersecurity, biochemistry, and model distillation to Opus 4.8, explicitly informing users of the switch. Official data claim over 95% of sessions avoid any fallback, but the author observed frequent fallback on medical queries, such as a Kaggle medical‑imaging task.

Benchmarking on SWE‑Bench Pro shows Fable 5 achieving 80.3% versus Opus 4.8’s 69.2%, GPT‑5.5’s 58.6% and Gemini 3.1 Pro’s 54.2%, a gap of 11–20 points. On the Diamond‑level FrontierCode suite, Fable 5 scores 29.3% compared to Opus 4.8’s 13.4%, and Terminal‑Bench 2.1 reaches 88.0%.

Anthropic emphasizes that “the longer and more complex the task, the larger Fable 5’s lead over other models.” A Stripe internal test reportedly migrated a 50‑million‑line Ruby codebase in one day, a job that would take a human team over two months.

In visual tasks, earlier Claude versions needed extensive tooling to beat Pokémon Red, whereas Fable 5 solved it using only raw screenshots. The model also reconstructed a web‑app source code from a single screenshot, demonstrating strong perception‑to‑code abilities.

Long‑context memory tests show Fable 5 playing Slay the Spire with file‑based persistent memory, achieving three times the performance improvement of Opus 4.8.

Mythos 5’s biochemistry case studies report a ten‑fold speedup in protein‑design steps, yielding strong candidates for 9 out of 14 targets, and a week‑long genomics project that produced a model 100× smaller yet more effective than a newly published Science paper.

Customer testimonials include Cursor claiming SOTA on CursorBench, Replit noting near‑saturation on its Vibe‑coding benchmark, and Lovable highlighting a shift from hundreds of prompts to a single‑shot workflow.

The author performed two hands‑on evaluations. First, he asked Fable 5 to reproduce Shopify’s Editions Winter 2026 site (≈6 k px, 12 sections, 27 Rive animations). The model inspected the live site, identified the Hydrogen + Tailwind v4 stack, extracted a 394 KB turbo‑stream payload containing a 939 KB JSON representation of all CMS data, and recreated the 3D Three.js scene using a custom 250‑line WebGL2 shader. The final replica rendered all sections, played every Rive animation, and passed TypeScript checks without errors.

Second, the author tasked Fable 5 with reproducing the DDIM sampler from the ICLR 2021 paper. The model drafted a specification focusing on Table 1’s claim that DDIM with 20–100 steps approximates the quality of a 1000‑step DDPM. Using a cloud RTX 4090 (48 GB), it reduced the standard 50 k‑sample FID to a 10 k‑sample proxy, documented assumptions (e.g., constant c in the second‑order sequence), and rewrote the sampler from the original formulas, including a unit test verifying η = 1 reduces to DDPM. It also handled infrastructure issues such as mirrored HuggingFace mirrors, corrupted Inception weights, and SSH session persistence.

Results: DDIM outperformed DDPM at 10, 50, and 100 steps (FID 24.95 vs 54.62, 11.88 vs 14.74, 11.16 vs 12.10). The absolute gap narrowed from 29.7 to 0.9, matching the paper’s trend. However, absolute FID values were 6–14 points higher than reported. The model investigated this discrepancy by calibrating a 10 k‑sample FID (2.1), running nine passing unit tests, and hypothesizing that the HuggingFace checkpoint used a non‑EMA weight variant, later confirming with a 1000‑step anchor experiment.

A validation agent independently audited the report, downgrading a claim that DDPM was three times worse than DDIM at 10 steps because the measured ratio was only 2.19×, and recorded the outcome.

Safety guardrails triggered five to six times during the two tasks. ExploitBench scores show Mythos 5 at 78.0% versus Opus 4.8’s 40.0% in vulnerability detection, confirming the high‑risk focus of the safeguards.

Cost analysis: the Shopify reconstruction cost $48 (based on $10/50 per million tokens), and the DDIM replication cost $60, totaling $108. The author notes that from today until June 22, Fable 5 is free for Pro, Max, Team, and seat‑based enterprise plans; after June 23 it will require usage credits.

Overall, the author’s hands‑on experience aligns with Anthropic’s claim: on longer, more complex tasks, Fable 5 demonstrates a markedly larger advantage over competing models.