The Vibe Father
Model Review
Claude Fable 5 is Anthropic's flagship coding model, released on June 9, 2026, and it arrived with a genuine milestone attached: it is the first model ever to crack 95 on SWE-bench Verified. We have been running it on real projects and on our live leaderboard at /benchmarks since launch day, and the short version is this — it is the best coding model money can buy, and it charges accordingly.
Fable 5 ships with a 1M-token context window, streams at 67 tokens per second, and costs $10 per million input tokens and $50 per million output tokens. That output price is the number to sit with. It is double Opus 4.8 and more than three times Sonnet 5. Anthropic is not pretending this is a workhorse model; it is priced like a specialist, and in our experience it earns that positioning on exactly the problems where cheaper models fall over.
The numbers
Here is how Fable 5 stacks up against the models it actually competes with on our board. Our Vibe Coding Index weights SWE-bench Verified at 40% and Terminal-Bench and LiveCodeBench at 30% each — you can see the live rankings at /benchmarks.
| Model | SWE-bench Verified | Terminal-Bench | LiveCodeBench | Price (in/out per M) |
|---|---|---|---|---|
 Claude Fable 5 | 95.0 | 83.1 | 89.8 | $10 / $50 |
| Claude Opus 4.8 | 88.6 | 78.9 | 87.8 | $5 / $25 |
 GPT-5.5 | 80.6 | 83.4 | 85.3 | varies by tier |
Two things stand out. First, that 95.0 on SWE-bench Verified is a 6.4-point gap over Opus 4.8, its own stablemate — on the benchmark that best predicts real repository work, nothing else is close. Second, Fable 5 is not unbeatable everywhere: GPT-5.5's 83.4 on Terminal-Bench edges out Fable's 83.1, making GPT-5.5 the current Terminal-Bench leader on our board. Fable 5 wins the index; it does not win every event.
Benchmarks only ever tell part of the story, so here is the part they miss. In our hands-on use, Fable 5's signature trait is restraint. Ask lesser models a hard question and they start typing; Fable 5 reads first. On a thorny debugging session it will trace the actual call path through four files before proposing a change, and its first proposal is usually the one that ships. It is also the least likely model we have used to quietly break something adjacent to what it touched — the failure mode that makes cheaper models expensive. The 67 tok/s streaming speed feels unhurried next to Sonnet 5's 89, but for the deliberate, plan-heavy work we give it, we have never once wished it were faster and dumber.
The price math
Benchmarks are free to read. Tokens are not. Take a realistic heavy month for someone running agents daily: 50M input tokens and 10M output tokens.
Fable 5: 50 × $10 + 10 × $50 = $500 + $500 = $1,000 per month. For comparison, the same workload on Opus 4.8 costs $500, on Sonnet 5 it costs $300, and on GPT-5.3 Codex it costs $227.50. You are paying roughly double Opus and more than four times Codex for the privilege of the top score.
Is that worth it? Sometimes, genuinely, yes. When Fable 5 solves a gnarly migration in one pass that Sonnet 5 would have churned through three failed attempts on, the arithmetic flips — three failed Sonnet runs plus your time can easily cost more than one clean Fable run. But if your work is mostly CRUD endpoints, UI wiring, and test scaffolding, you are lighting money on fire. We break down the budget angle in our cheapest coding models roundup.
Its best seat on the team
We think about models in terms of roles — planner, builder, scout, reviewer — and Fable 5 is a planner and a hard-problem closer, full stop. In a multi-agent setup, this is the model you hand the architecture decisions, the ambiguous debugging sessions, and the tasks that have already defeated a cheaper model. Its 1M context means it can hold an entire mid-sized codebase in view while it plans, which materially changes the quality of its decomposition.
What it should not be is your default builder. Routing every routine implementation task through a $50-per-M-output model is how invoices get embarrassing. The pattern that works: Fable 5 plans and reviews, Sonnet 5 or GPT-5.3 Codex builds, and Fable 5 gets tagged back in only when something is stuck. We map this out role by role in the best model for each agent role.
Verdict
Claude Fable 5 is the real thing. The first 95 on SWE-bench Verified is not a rounding-error headline; on hard, multi-file, real-repository work it is visibly better than everything else we have tested, and it currently tops our best coding model rankings for 2026.
Who should pay for it: teams shipping complex production systems, anyone whose time is worth more than their token bill, and multi-agent operators who need one genuinely elite planner at the top of the stack.
Who should not: hobbyists, solo devs on straightforward projects, and anyone using it as an everything model. If your tasks would score fine on Opus 4.8 — and most tasks would — you are paying a 100% premium for headroom you never touch. And if your work lives in the terminal, note that GPT-5.5 actually beats it there; see our full cross-lab shootout before you commit. Buy Fable 5 like you would hire a principal engineer: for the problems that deserve one.