Claude Fable 5

MA

Matthew Berman26 Aug 2026

Anthropic's premium model at $50 per million output tokens. Criticized for Anthropic retaining customer data, which is a platform risk for businesses.

MA

Matthew Berman4 Aug 2026

Closed-source US frontier model used as comparison; leads SWE-Bench Pro at 80 and visual agent/coding benchmarks. Priced at $10/M input and $50/M output, rumored at 7+ trillion parameters.

MA

Matthew Berman31 Jul 2026

Anthropic's top model, shown at $2.75 per task on Artificial Analysis. The creator says Anthropic had it internally months before public release.

AI

AI Search27 Jul 2026

Previous best Anthropic model, used as comparison; more restricted, rejecting biomedical requests.

MA

Matthew Berman24 Jul 2026

Anthropic's largest and best model, used as the comparison point. Opus 5 beats it on most benchmarks at half the price.

MA

Matthew Berman23 Jul 2026

Anthropic model that is even more expensive than GPT-5.6, costing $2.75 per task on Artificial Analysis versus about $1 for the others. It uses many more tokens, and the creator uses it for the planning stage.

MA

Matthew Berman15 Jul 2026

Anthropic's roughly 10 trillion parameter model, ranked first on the Artificial Analysis Intelligence Index at 60. It costs about $2.75 per task, and Anthropic is pulling it from subscriptions because of capacity constraints, though access was extended and it is expected to return.

MA

Matt Wolfe10 Jul 2026

Live demo

Used as a comparison point. The creator shows a Fable-built website that he judges better than the others. He says Fable's access was extended to July 12 and Anthropic reset the weekly limits. He finds it smarter at architectural decisions, while GPT-5.6 gets closer to the goal on big tasks.

AI

AI Search10 Jul 2026

Competitor model. It was the only one to find the frog and has a lower hallucination rate than GPT-5.6 Sol, but it refuses biology and cybersecurity questions and costs more.

AI

AI Search5 Jul 2026

Re-released after a ban was lifted, with new safety classifiers. Subscription use is limited to 50% of weekly limits and ends July 7; routine coding and debugging fall back to Opus 4.8. Benchmarks show mixed results versus the old version.

MA

Matt Wolfe3 Jul 2026

Live demo

Fable 5 was redeployed July 1st after being pulled by the US government, with extra safeguards and a classifier that flags more benign coding requests. The creator built a game, a shorts dashboard and a B-roll tool with it and says it feels about the same, though Bridgemind reported benchmark drops. Plan access ends July 7th, then usage credits.

MA

Matt Wolfe19 Jun 2026

US government export controls forced Anthropic to suspend Fable 5 and Mythos 5 for all users on June 12 over a jailbreak dispute. The creator liked Fable for one-shot app coding and says access may return within days.

AI

AI Search14 Jun 2026

Anthropic's flagship, claimed Mythos quality, heavily restricted in AI research, security and biology. It was found to secretly give weaker answers, which was later retracted. Access was then suspended for all customers following a US government directive.

MA

Matthew Berman13 Jun 2026

Anthropic's model launched last week, described as Mythos with a classification layer blocking cybersecurity, bio and nuclear questions. The creator calls it the best model he has used; access was suspended for foreign nationals under a US export control directive, so Anthropic disabled it for all customers. Pliny jailbroke it within hours, and Anthropic says the demonstrated jailbreak only found minor known vulnerabilities.

MA

Matt Wolfe12 Jun 2026

Live demo

Mythos-class model above Opus, announced June 9. Strong benchmarks and long autonomous work, but backlash over hidden safeguards (bio, cyber, frontier LLM development) that downgrade to Opus 4.8. Priced at $10/$50 per million tokens, double Opus 4.8. Creator one-shotted a Megabonk clone, an AI Tube YouTube clone and a B-roll generator.

AI

AI Search11 Jun 2026

Live demo

Anthropic's newest generally available model, said to beat Claude Mythos preview and Opus 4.8. Creator tested it on frog-finding, ray tracing, Earth digital twin, a DAW, a mecha game and a chemistry course; it did well on coding but is slow, expensive ($50/M output tokens), hallucinates more, and has safeguards that fall back to Opus 4.8 on biology/cyber topics.

MA

Matt Wolfe11 Jun 2026

Live demo

First Mythos-class model made safe for general use, priced at $10/$50 per million tokens. Strong on long one-shot coding tasks (Stripe 50M-line Ruby migration, game clones), but slow, token-hungry, and safety classifiers hand off to Opus 4.8. The creator tested it, including a one-shot Mega Bonk clone that worked well.

MA

Matthew Berman9 Jun 2026

Live demo

New Anthropic model family trained in the class of Mythos, with guardrails for general use. Priced at $10/M input and $50/M output tokens; leads benchmarks like SWE-Bench Pro (80%) and Terminal Bench. Creator found it very capable on long tasks but verbose, slow, and prone to asking many clarifying questions.