Claude Fable 5 is Anthropic's Mythos-class flagship, and at the July 2026 snapshot it was the highest-ranked model on the Arena leaderboard: #1 on the Artificial Analysis Intelligence Index at ~60, about a point clear of GPT-5.6, at $10 in / $50 out per million tokens. If you only learn one model name this year, this is the one. One date stamp before the numbers: Artificial Analysis's v4.3 index refresh on September 30 has since reshuffled the top of the board, so read every ranking below as July history, not current standings.
The Name and the Tier
Let's clear up the naming first, because it trips everyone up. "Fable 5" sounds like a codename, but it's real: Anthropic's Mythos-class flagship, the first model in the Claude 5 family, released June 9, 2026. Capability-wise it sits above Opus 4.8, which was itself the previous top of the stack. A lot of people still assume "Opus" means "the biggest one." It doesn't anymore. Fable 5 is.
The positioning is deliberate. Anthropic didn't just scale up Opus; they opened a new tier above it. Mythos-class is built for the hardest and longest work: multi-hour agentic loops where the model can't lose the thread at step 12. It carries a 1M+ token context window with 128K max output, and on Anthropic's internal evals it's a full generation ahead of Opus. (A telling launch detail: with its file-based memory, Fable 5 plays Slay the Spire about 3x better than Opus 4.8. A silly benchmark, but a real signal for long-horizon planning.)
So: vendor Anthropic, shipped June 2026, top of the Claude 5 family, designed to be the smartest model money can buy. That's the one-sentence version. The interesting part is whether "smartest" holds up under the benchmarks.
What the Model Actually Does
Numbers first. These are the July 2026 Arena and Artificial Analysis snapshots, not Anthropic's slide deck.
- SWE-Bench Pro: 80.3%, #1. Real software-engineering tasks. Fable 5 leads the runner-up by roughly eleven points. A cliff, not a margin.
- SWE-bench Verified: 0.950, #1. The verified subset, same story.
- FrontierCode Diamond: 29.3%. Beats Opus 4.8 on the hardest coding eval. In Arena's front-end coding duels it won 72% of head-to-heads, finishing 98 Elo ahead of #2.
- Artificial Analysis Intelligence Index: #1. About a point ahead of GPT-5.6 Sol. On the newer GDPval-AA real-work index it broke Elo 1932, leaving Opus 4.8 well behind.
- Eight industry benchmarks, all #1. Artificial Analysis rolled out six new domain indices (finance/accounting, legal, medical, strategic operations, engineering, economics) plus the existing agent and coding indices. Fable 5 (with Opus 4.8 as fallback for ~5% of sensitive turns) took first on all eight.
- Agent Arena: #1, record margin. It posted an 11.2% net lift, the largest gap the arena has ever recorded, driven by an 18.2-point edge in confirmed task completion and a 30.6-point edge in satisfaction-to-complaint ratio.
The honest caveat: this is an English-first, reasoning-and-code model. Its Chinese is competent but reads like a brilliant non-native speaker: technically perfect, slightly off in cadence. And at ~63-65 tokens/second it's the slowest of the frontier tier. For real-time chat that stings. The real-world proof point I keep coming back to: one widely-circulated report has Stripe using Fable 5 to migrate a 50-million-line Ruby codebase in a single day, work their own team had estimated at over two months. That's the kind of job this model exists for.
The Bill and the Access Saga
Here's the part that makes finance teams wince. $10 per million input tokens, $50 per million output tokens: the steepest rate in Anthropic's lineup, double Opus 4.8 ($5/$25) and about 3.3x Sonnet 4.6 ($3/$15) on output. There's a reason GLM-5.2 gets pitched in developer circles as "1/39 the price of Fable 5."
Access has been a soap opera. Fable 5 shipped June 9, got pulled three days later by a US export-control order, came back July 1, ran at 50% capacity for a stretch, and as of July 20 Anthropic split access by tier: Max and Team Premium subscribers (the $200+/mo plans) keep it permanently but capped at 50% of standard weekly usage; Pro and Team Standard lost free access and got a one-time $100 credit. The API, meanwhile, has been fully available throughout at the $10/$50 rate.
Is it worth it? Depends entirely on the task. On a hard, high-value job (debugging a multi-file service, analyzing a contract, writing long-form where voice matters), the quality gap more than pays for the token cost. On a thousand simple CRUD summaries, you're burning money for nothing. Fable 5 is a scalpel, not a sledgehammer; pricing it like a sledgehammer and using it like one is how you blow a budget.
Where the Number-Ones Sit
I keep saying "#1," so let me lay out exactly where that holds, as of the July 2026 Arena snapshot:
- Arena overall: #1, holding the top spot for consecutive weeks.
- Text Arena: #1.
- Code Arena: #1 (1563 Elo, ~10 points ahead of Claude Opus 4.7 Thinking).
- Agent Arena: #1 (see above).
- Artificial Analysis Intelligence Index: #1.
- Tool-hallucination: #1. It invents fewer bad tool calls than any other frontier model.
The headline isn't any single score. It's the breadth. Most models dominate one or two benchmarks and trail elsewhere. Fable 5 is the number-one model on five of the benchmarks an independent evaluator tracks (MMLU-Pro, SWE-bench Verified, HLE, Chatbot Arena, and the AA Intelligence Index) and within a point or two on the rest. That's what "flagship" actually means in 2026: not one big number, but no weak spots.
Fable 5 Against the Field
The same weight class, priced to the July snapshot:
| Model | AA Index | Pricing (in/out per MTok) | Output speed | Best at |
|-------|----------|---------------------------|--------------|---------|
| Claude Fable 5 | ~60 (#1) | $10 / $50 | ~63-65 t/s | Long reasoning, creative writing, complex code |
| GPT-5.6 Sol | ~58.9 (#2) | $5 / $30 | ~78 t/s | Math/logic, the reasoning_effort dial, tools |
| Gemini 3.5 Flash | Flash tier | $1.50 / $9 | 289 t/s | Speed, native multimodal, 1M+ context |
| Claude Opus 4.8 | just below Fable 5 | $5 / $25 | faster | The fallback Fable 5 itself routes sensitive turns to |
| Qwen3.7 Max | 56.6 | $2.50 / $7.50 | mid | Chinese content, 1M context, cost |
| DeepSeek V4 | high-80s GPQA | $0.14 / $0.28 | mid | Cost-bound volume |
The pattern: Fable 5 is the smartest and the most expensive, and also the slowest of the frontier tier. GPT-5.6 is nearly as smart at roughly a third the cost per task, with a tunable reasoning dial that's genuinely useful. Gemini 3.5 Flash is a fraction of the price and over 4x the speed, and it eats Fable 5's lunch on multimodal. For Chinese, Qwen3.7 Max is both cheaper and more natural. For pure volume, DeepSeek V4 is basically free.
There's no single king. There's a model that's best for your task, and the gap between "best overall" and "best for this specific job" is where most of the money gets wasted.
Pick It When, Skip It When
- Use it for hard reasoning that spans many steps: architectural decisions, multi-file refactors, contract analysis, deep research synthesis.
- Use it for English long-form and creative writing where voice and risk matter. It's the only frontier model whose prose doesn't read like a template.
- Use it for agentic loops where task-completion rate is the whole point; the Agent Arena numbers aren't a fluke.
- Don't use it for high-volume simple tasks; the $50/M output price will ruin you. Route those to DeepSeek V4 or Qwen3.7 Max.
- Don't use it for real-time chat where latency is king; 63 t/s loses badly to Gemini 3.5 Flash's 289.
- Don't use it as your default Chinese writer. Qwen3.7 Max is both cheaper and more natural.
The mental model: Fable 5 is the senior engineer you bring in for the hard problem, not the one you assign to every ticket.
What It Means Off the API
I evaluate models for x-rush.ai; on that platform the framing above flips. You don't buy tokens; you open a feature like AI Chat or Auto Content, describe the job, and the platform routes each task to a model that fits it. Providers stay out of the product vocabulary, so nobody on the platform promises you "Fable 5 in the pool", and this review is not such a promise. What carries over from a teardown like this is calibration: knowing a hard reasoning task is worth more than a thousand easy ones. Billing agrees; text generation is a flat 0.5 credits per unit whether the task behind it was trivial or Mythos-class.
The Verdict, For Now
At the July snapshot, Fable 5 was the smartest general-purpose model you could buy: #1 on the Arena leaderboard, the AA Intelligence Index, the coding and agent arenas, first across eight industry benchmarks. It was also the most expensive at $10/$50, among the slowest at ~63 t/s, and the wrong tool for Chinese. The September 30 v4.3 refresh has since rearranged the top of the board. The shape has not moved: a scalpel for hard, English-shaped work, and an expensive habit on high-volume tickets.
Rent the scalpel for the hard job. Or hand the whole job to a router and let the task pick the model.
AI Chat on x-rush works that way: describe the job, and the model question stays an implementation detail.