Fable 5 versus Opus 4.8 with /goal: pricing and persistence compared

Anthropic put Fable 5 back on general release this week, July 2 by 9to5Google’s timeline (“available again globally tomorrow,” said the July 1 statement), after a reported safeguard bypass took it offline and Anthropic retrained the safety classifier. It came back at $10 per million input tokens and $50 per million output, exactly double Opus 4.8’s $5/$25 on the official price list; on a days-long unattended run that doubling is the difference between one token budget and two. The product page, however, does not lead with capability numbers; it leads with persistence, describing a model that is “able to run agents for days unattended.” In this note, I examine whether that advantage justifies the doubled price, given that Claude Code already provides a persistence mechanism that works with any model in the lineup.

The benchmark numbers

The numbers below are collected results rather than my own runs. llm-stats’ comparison (Anthropic’s own announcement table ships as an image, so I cite the aggregator) puts Fable 5 at 95.0% on SWE-bench Verified against 88.6% for Opus 4.8, at 80.0% against 69.2% on SWE-bench Pro, and at 29.3% against 13.4% on FrontierCode Diamond. The last gap is the serious one; a doubling, on the benchmark closest to novel engineering work.

The benchmark numbers confirm a real capability gap. The price difference pays for one of two axes: smarter single turns, or more of them per task.

The /goal command as a persistence mechanism

/goal (Claude Code 2.1.139 and later) sets a completion condition, and a separate evaluator checks the condition after every turn, sending the model back to work until the condition holds. Under the hood it is a session-scoped Stop hook; the evaluator is the configured small fast model, Haiku by default, and the documentation prices its overhead as “typically negligible compared to main-turn spend.”

This is the same observe, act, and check loop that the product page describes, implemented outside the model rather than inside it. /goal all tests pass and git status is clean on Opus 4.8 costs half of Fable per token and keeps running just as long. This matches a pattern I keep returning to in agentic coding best practices: an external check is more reliable than the model’s own sense of done. Differences of this kind are also why I maintain a running Codex CLI versus Claude Code comparison; the loop mechanisms differ more between the tools than the models do.

The safety fallback and its cost

Fable 5 ships with classifiers that reroute cybersecurity, biology and chemistry, and distillation-flagged requests to Opus 4.8. Anthropic reports that “more than 95% of Fable sessions involve no fallback at all,” and it should be noted that rerouted requests are not billed at Fable prices, as the product page confirms. Still, llm-stats notes that 20.9% of Terminal-Bench trials hit a safety refusal, and that on cyber evaluations the deployed Fable 5 lands much closer to Opus 4.8 than the published Mythos 5 numbers. My research work touches security-adjacent topics often enough that, given the 20.9% Terminal-Bench refusal rate, a meaningful share of a Fable subscription would consist of Opus answers with additional routing latency.

Limitations of the comparison

Two honest caveats. First, /goal buys turns, not first-shot capability. No Stop hook makes Opus 4.8 solve a FrontierCode Diamond problem it cannot solve; when every step sits at the edge of the model’s ability, however, retries compound errors rather than fix them. Second, my half-price framing holds at standard speed only: Opus 4.8’s fast mode is priced at $10/$50, the exact Fable rate.

My working rule for now is therefore: pay the doubled price when single-turn quality is the bottleneck, novel architecture and debugging where a wrong first step propagates through the rest of the run. Stay on Opus 4.8 with /goal when persistence is the bottleneck, migrations, backlogs, and test-until-green loops, which covers most of my agent hours. TODO, dated July 4: run the same backlog-clearing goal on both models once the 50% Fable usage cap lifts on July 7, and count the tokens.