Routing Claude Fable 5.1 and Gemini 3.8 Flash Together

Key takeaways

1. The founding case is a poor ROI, not an outage. a low-stakes question, an agent trawling chats and documents, the right answer, $38. The task did not call for costly planning. This is the marginal value of the token, discussed in [[gupta-token-budget-wars-marginal-token-utility-2026-05-28]].
2. Two model shapes, one API surface. Fable 5.1 for Mythos-class autonomous planning, the one-million-token window, and multi-step diligence; Gemini 3.8 Flash for speed at $0.75 / $3.75 per million input/output tokens, with an adjustable thinking level. Routing everything to either one is named as the mistake to avoid in both directions: paying frontier rates to read diffs, or firing off a fast model on an irreversible migration with no plan.
3. Activating Fable on Google Cloud is a multi-step journey. enable the API, enable the model, fill out a usage form, then accept the Advanced AI Safety Addendum by calling setPublisherModelConfig with dataSharingEnabledProvider: ANTHROPIC at the project level. A 409 means it's already done; an explicit 403 flags the missing step. The model name in the URL uses hyphens, claude-fable-5-1. A global endpoint is recommended, with multi-regional and regional variants.
4. The in-house benchmark fits in two files. opencode configured with both Vertex providers, promptfoo launching opencode run --auto with each model on the same prompts, 60-second timeout. The rule: do not modify the task files between passes.
5. The two results cross over. on a PR branch state, Fable runs through meta-reasoning turns re-inspecting unchanged tree hashes (nearly 6 s, 23 times the cost); Flash triggers the git tools and answers in 1.1 s. On the Postgres-to-Spanner migration, Flash renders a clean linear sequence in 3 s; Fable, in 12 s, a full DAG with clock-drift risks on dual writes, idempotency keys, and a reversible rollback gate.
6. Pattern adopted, asymmetric coordination. cheap speed on the front line, deliberate depth at architecture checkpoints. In a coding harness, named subagents mapped to distinct models (worker on Flash, deep-thinker on Fable); the planner identifies the cause and breaks down the work, the workers execute and report, the planner verifies and confirms or reassigns. Two model families checking each other, a possible parallel with [[willison-fable-judgement-delegation-subagents-2026-07-03]].
7. Reshape the harness to fit the model. 3 to 5 targeted tools with strict JSON schemas for the fast model, versus a 50-tool catalog creating parameter confusion and wasted context; architecture documents, schemas, and guidelines for Fable, but without direct write access to files.
8. 'Ask for Help' escalation. every request starts on the fast worker, equipped with an ask_for_help(reason, failed_attempts, context) tool; it handles 85 to 90% of requests on its own and only escalates on ambiguity, an irreversible action, or two consecutive tool failures.
9. Against automatic 'smart' routers. mature in predictive ML (bandits, ad tech, fraud) because inputs are bounded and feedback is immediate; in multi-turn agents, the first turn lacks signal, success doesn't translate into learnable features, and bad choices get replayed on the other path, eating into the savings and adding latency. Verdict: compose explicitly, route by task boundary, let code assertions or human intent govern handoffs.
10. ⚠️ The comparative figures rest on two tasks run once each, with no variance or published protocol; the author says so himself and refers everyone to their own evaluations.
11. The ROI equation. adds up human time saved, features shipped faster, and errors and outages avoided, minus the cost of tokens and the team's ramp-up. The author notes that this calculation is moved more by doing more work, faster, than by saving tokens. A per-team comparison worth reading alongside [[patel-block-buzz-teams-tokens-benchmarks-2026-08-06]].