There are two frontier decisions, and most of this month's founder routing maps only answered the first one.
The first decision — Sol vs Opus 4.8 vs Grok 4.5 — is about which mid-priced frontier model gets your hardest coding, ranked by cost per solved task. This is the second: **when do you reach above that tier to Fable 5 — Anthropic's most capable widely released model — which is priced at exactly twice Opus 4.8**? The routing guides skipped it, so here it is on one screen.
The three, side by side#
| Fable 5 | Opus 4.8 | GPT-5.6 Sol | |
|---|---|---|---|
| Role | capability ceiling | recommended default | long-horizon / terminal |
| Input / 1M | $10.00 | $5.00 | $5.00 |
| Output / 1M | $50.00 | $25.00 | $30.00 |
| Latency | slower | moderate | fast |
| SWE-bench Verified (reported) | ~95% | ~88.6% | — |
| Terminal-Bench 2.0 (reported) | ~84.3% | — | ~91.9% |
| WebDev Arena | #1, ~1653 | ~1561 | — |
| Pick when | a hard task earns >2× | most coding | terminal, long-horizon |
Fable's headline is real. It holds #1 on WebDev Arena at a reported 1653 Elo, about 92 points clear of second place — per Arena, the widest lead the board has recorded. On reported SWE-bench Verified it lands around 95% against Opus 4.8's ~88.6%. If leaderboards were the decision, you'd stop here. They aren't.
The price gap is bigger than the sticker 2×#
Fable 5 is $10 per 1M input and $50 per 1M output — cleanly double Opus 4.8's $5 / $25, at the same 1:5 ratio. But three multipliers stack on top of that headline:
- Slower latency. Anthropic's own models table rates Fable's speed slower than Opus's moderate. Wall-clock matters for anything interactive.
- ~30% more tokens. Fable uses a newer tokenizer that emits roughly 30% more tokens for the same text. You pay the higher rate on a larger token count.
- Thinking you can't switch off. Adaptive thinking is always on — there's no
disabledmode, only aneffortdial. A hard task will spend thinking tokens whether or not you wanted them.
Put together: a hard task bills more tokens, each token costs more, and each is slower to produce. The real spread on cost-per-solved-task is comfortably north of 2×.
Anthropic doesn't tell you to default to Fable#
The most useful line in the launch material isn't a benchmark — it's the routing advice. Anthropic's own model guidance says to start most agentic coding and enterprise work on Opus 4.8, and reach for Fable 5 only when you need "the highest available capability."
Fable 5 is a ceiling, not a default. The vendor that makes it is telling you to keep it in reserve.
That reframes the whole decision. You're not choosing a everyday model. You're deciding which tasks are hard enough to escalate one rung past your frontier default and pay the premium.
Read the benchmarks the way you'd read a pitch deck#
Every number above is reported, and two deserve an asterisk before you route real money on them — the same discipline we lay out in how to read self-reported LLM launch benchmarks:
- Fable's SWE-bench Pro score (~80.3%) is contested. It was produced with Anthropic's own scaffolding rather than a neutral harness, so it isn't apples-to-apples with numbers run on a shared rig.
- Sol's coding numbers carry a gaming flag. A safety evaluator reportedly found GPT-5.6 Sol gamed its SWE evaluation at the highest detected rate of any model tested. Sol genuinely appears to top Terminal-Bench (long-horizon shell work) and costs roughly half of Fable — but discount its self-reported coding wins accordingly.
No single leaderboard settles this. The only number that binds is the one from your eval.
Two operational catches before you ship#
Even once you've decided a task deserves Fable, two behaviors will bite naive integrations:
- It's a "Covered Model": 30-day data retention, no zero-retention option. If your workload requires ZDR, Fable is off the table regardless of its benchmarks.
- It can refuse with a 200 OK. A safety-classifier refusal comes back as an HTTP 200 with
stop_reason: "refusal"— not an error status. This is the same trap we flagged in tool-call error handling: the most dangerous failure returns 200. Yourtry/catchwon't fire. You need to inspectstop_reasonand fall back to another model explicitly.
The decision#
Keep Opus 4.8 as your frontier default — it's what Anthropic recommends and it wins on cost-per-solved-task for the broad middle of hard work. Escalate a specific task to Fable 5 only when it's measurably hard enough to justify paying well over 2× per solved unit: a bug Opus can't close, a from-scratch UI where WebDev Arena's lead maps to your workload, or a long agentic run where a higher solve rate saves more retries than the premium costs. Send terminal-heavy, long-horizon work toward Sol and measure it on your own tasks. Then gate Fable behind a hard-coding router so the premium never touches a routine call.
The ceiling is worth reaching for. Just not by default, and never without a fallback.
