Models move fast — this page is dated
Prices and benchmarks were verified as of August 2026. Opus 5 became the default model on Claude Max on 2026-07-24, three days after Fable 5 was briefly its own flagship. Re-check before you wire either model into a pipeline.
See the full Fable 5 vs Opus 5 comparisonDeveloper Decision Guide
Fable 5 or Opus 5 for coding?
Both models are excellent engineers. The difference is where each one earns its keep — and where the other is simply the better tool. Here's how to pick, using the benchmarks and the real cost math.
TL;DR
Default to Opus 5 at high effort for everyday coding. Escalate to Fable 5 only for: long-horizon multi-file agentic work, security-critical/Mythos-grade code review and hardening, and 1M-context full-repo analysis.
Specs at a glance
Prices from Anthropic's published rates (USD per million tokens), as of August 2026. The token-cost difference is real; the benchmark difference is the question.
| Dimension | Fable 5 | Opus 5 |
|---|---|---|
| Released | 2026-06-09 | 2026-07-24 |
| Input price ($/M) | $10.00 | $5.00 |
| Output price ($/M) | $50.00 | $25.00 |
| Context window | 1M | 1M |
| Max output | 128K | 128K |
| Data retention | 30 days | None |
Coding benchmarks, honestly sourced
Where Anthropic has published a number we quote it; where it hasn't, we say "not confirmed" rather than guessing. Third-party numbers are labelled as such.
| Dimension | Fable 5 | Opus 5 |
|---|---|---|
| SWE-bench Pro | 80.3% | Not confirmed |
| FrontierCode | 29.3% | Not confirmed |
| Frontier-Bench v0.1 | 33.7% | 43.3% |
| CursorBench 3.2 | Peak | Within 0.5% of Fable's peak (max effort) |
| Cost per solved task | ~$2.75 | ~$2.03 |
All figures verified as of August 2026. Fable 5's SWE-bench Pro 80.3%, FrontierCode 29.3% and Frontier-Bench v0.1 33.7% are Anthropic-published. Opus 5's SWE-bench Pro and FrontierCode results are not confirmed — Anthropic has not published them. On Frontier-Bench v0.1 Anthropic lists Opus 5 at 43.3% (its best published coding score). CursorBench 3.2 (third party) puts Fable 5 at peak with Opus 5 within 0.5% at max effort. Cost per solved task is Artificial Analysis, not Anthropic.
The decision matrix
Match the work to the model. These four lanes cover the coding work that actually shows up.
Default: Opus 5 at high effort
For most day-to-day coding — feature work, bug fixes, reviews, small refactors — Opus 5 at high effort is the right call. It is roughly half the token price, near-Fable performance on the benchmarks that matter for this work, and ships with no retention requirement. Cheap, fast, good enough.
Agent: long-horizon multi-file refactors
When a task spans many files, many steps, and a long reasoning chain — a cross-cutting refactor, a migration, a subsystem rewrite — Fable 5's edge compounds. Its best published SWE-bench Pro score (80.3%) reflects the kind of sustained, multi-step work where it holds a thread Opus 5 can drop. Escalate here.
Security: Mythos-grade review & hardening
For security-critical code review, threat modelling, and hardening where a missed edge case is a live vulnerability, Fable 5 is built for the job — it pairs with the Mythos 5 safety tier and its 30-day retention preserves the audit trail. If the code is going to be attacked, pay for the extra margin.
Repo: 1M-context full-repo analysis
Both models carry a 1M context window, but for a full-repository analysis where you dump the entire codebase and ask questions, Fable 5's long-context strength and its published coding scores make it the safer bet. Use it for architecture reviews, large-scale on-boarding questions, and codebase-wide audits.
These are defaults, not rules. If you're on a budget, Opus 5 at high covers far more of the day than Fable 5 does. If you've hit a wall where Opus 5 keeps dropping the thread, that is exactly the signal to escalate.
The cost math that actually matters
Pass rate is not the whole story — retries are.
The spec sheet says Fable 5 costs 2x per token, but real spend is driven by retries. If a model solves a task in one attempt, a 2x token price is fine. If it takes three attempts to land the same result, you've paid 3x for the same output. Artificial Analysis' cost-per-solved-task estimate captures this: Fable 5 at ~$2.75 per solved task vs Opus 5 at ~$2.03. That's a 1.35x gap, not 2x — because on the harder tasks Opus 5's lower pass rate eats into its price advantage. For long agentic runs, the retry savings can flip the winner entirely.
Cost per solved task is an Artificial Analysis estimate that blends price, latency and pass rate, not an Anthropic figure. Your real number depends on your workload mix, cache hit rate, and how often you retry.
Run the numbers for your workload
Your workload is not the benchmark. Use the cost calculator to price Fable 5 vs Opus 5 on your actual request volume, then check the benchmarks page for the full picture.