# blount-googlecloudtech-fable-gemini-flash-same-team-2026-09-14

## Veille

Long-form article published on **September 14, 2026** on the X account **@GoogleCloudTech** (~2,000 words, four steps, six blocks of commands and configuration), signed by **Alan Blount**, Technical Solutions Consultant at Google. Starting point: a trivial question, handed to a research agent, billed at **$38**. The article proposes running **Claude Fable 5.1** and **Gemini 3.8 Flash** side by side behind the same **Gemini Enterprise Agent Platform** API (formerly Vertex AI). **(1) Configuration**: access to Fable is subject to Google Cloud's *Advanced AI Safety Addendum*, which requires a `setPublisherModelConfig` call with prompt-response sharing enabled for Anthropic, on pain of a 403; Flash works as soon as it is enabled. **(2) In-house benchmark** with promptfoo driving opencode on two tasks: on a git branch state, Flash answers in **1.1 s** for under a tenth of a cent while Fable costs **23 times more**; on a Postgres-to-Spanner migration, Fable produces in **12 s** a formal DAG with clock-drift risks and a rollback gate, whereas Flash delivers a linear sequence in 3 s. **(3) Usage**: *« asymmetric coordination »* — the fast model as first-line coordinator, the deep planner at checkpoints, an `ask_for_help` tool for escalation, and a warning against dynamic routers: *« Keep it boring »*. **(4) Token ROI**: the value of human time, shipped features, and incidents avoided, minus the cost of tokens and the team's ramp-up. The same task-boundary routing move as [[mazmanov-portal-spotify-shunt-claude-code-tokens-2026-09-03]], within the FinOps framework set out by [[rafal-wenvision-tokenomics-foundation-finops-ia-2026-06-04]].

## Titre Article

Put Claude Fable 5.1 and Gemini 3.8 Flash on the same team

## Date

2026-09-14

## URL

https://x.com/GoogleCloudTech/status/2099507349828350285

## Keywords

model routing, asymmetric coordination, tokenomics, token ROI, Claude Fable 5.1, Gemini 3.8 Flash, Gemini Enterprise Agent Platform, Vertex AI, Model Garden, Advanced AI Safety Addendum, setPublisherModelConfig, prompt-response sharing, opencode, promptfoo, in-house benchmark, evaluations, specialized subagents, deep-thinker, worker, ask_for_help escalation, dynamic routers, multi-armed bandits, task boundary, Postgres Spanner migration, DAG, Google ADK, LangGraph, thinking level, quota, provisioned throughput

## Authors

Alan Blount (Technical Solutions Consultant, Google Cloud, @zeroasterisk), publié sur le compte X @GoogleCloudTech.

## Ton

**Profile**: engineering tutorial published on a Google Cloud editorial account, written in the first person, that promotes the company's platform while describing its friction points. Audience: developers building coding agents or custom agent services, and those responsible for inference costs.

**Style**: a four-step numbered progression, each closed out with executable artifacts — environment variables, curl calls, expected JSON responses, 409 and 403 error codes with their exact message, `opencode.json` and `promptfooconfig.yaml` files. Self-deprecating humor (*« impossible things are easy, but easy things are hard »*, *« whew. We did it 🎉 »*) acknowledges the clunkiness of the activation journey. Short formulas serve as rules: *« That is a mistake »*, *« Keep it boring »*.

**Epistemic stance**: the author upfront rejects public leaderboards (*« Twitter vibes and synthetic leaderboards »*) and frames his own comparison as *« one illustrative example »* to be redone on one's own tasks. The figures given (1.1 s, 23 times, 12 s) cover two tasks in a clean container, with no repetition or detailed protocol. The escalation rates (85 to 90% handled by the worker) and the 3-to-5-tool threshold are stated as heuristics, without a source. The ROI equation is explicitly presented as impossible to generalize. The text is platform content: the models cited are the ones Google Cloud distributes, and the closing reference points to other posts from the same account.

## Pense-betes

- **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]].
- **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.
- **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.
- **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.
- **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.
- **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]].
- **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.
- **'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.
- **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.
- ⚠️ **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.
- **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]].

## RésuméDe400mots

Alan Blount, Technical Solutions Consultant at Google Cloud, publishes on September 14, 2026 on the X account @GoogleCloudTech a four-step tutorial on making Claude Fable 5.1 and Gemini 3.8 Flash work together. Starting point: a trivial question handed to an agent that searched through its conversations and documents cost 38 dollars for a correct but low-stakes answer.

Both models are accessible behind the same Gemini Enterprise Agent Platform API, the new name for Vertex AI, via Model Garden. Fable 5.1 brings Mythos-class autonomous planning, a one-million-token window, and multi-step diligence; Gemini 3.8 Flash brings near-frontier reasoning at a price of 0.75 dollars per million input tokens and 3.75 for output, with an adjustable thinking level. Routing everything to a single model is presented as a mistake.

The first step details activation. Fable falls under Google Cloud's Advanced AI Safety Addendum: prompt-response sharing for Anthropic must be enabled at the project level via a call to the setPublisherModelConfig API, otherwise every request returns a 403 error. Flash works as soon as the model is enabled.

The second step recommends not relying on public leaderboards and instead comparing models on one's own tasks, here with promptfoo driving the opencode agent. On a git branch state summary, Flash answers in 1.1 seconds for under a tenth of a cent, while Fable multiplies reasoning turns and costs 23 times more. On a Postgres-to-Spanner migration, Flash delivers a linear sequence in 3 seconds, Fable a formal dependency graph in 12 seconds, with clock-drift risks and a rollback gate.

The third step describes asymmetric coordination: the fast model on the front line, the deep planner at architecture checkpoints. In a coding harness, this translates into subagents mapped to distinct models; in a custom service, into a harness tailored to each model, automated evaluations, and an escalation tool that the worker calls on ambiguity, irreversible action, or a double failure. Dynamic routers are ruled out in favor of explicit routing by task boundary.

The fourth step lays out a token ROI equation grounded in value produced rather than pricing tables.

## GrapheDeConnaissance

- Alan Blount —travaille_chez→ Google Cloud (ORGANISATION, 0.95)
- Google Cloud —publie→ Gemini Enterprise Agent Platform (TECHNOLOGIE, 0.95)
- Gemini Enterprise Agent Platform —remplace→ Vertex AI (TECHNOLOGIE, 0.93)
- Gemini Enterprise Agent Platform —utilise→ Model Garden (TECHNOLOGIE, 0.93)
- Model Garden —permet→ Claude Fable 5.1 (TECHNOLOGIE, 0.94)
- Model Garden —permet→ Gemini 3.8 Flash (TECHNOLOGIE, 0.94)
- Claude Fable 5.1 —fait_partie_de→ Advanced AI Safety Addendum (DOCUMENT, 0.92)
- Advanced AI Safety Addendum —utilise→ setPublisherModelConfig (TECHNOLOGIE, 0.92)
- Alan Blount —mesure→ une question anodine traitée par un agent de recherche a coûté 38 $ (MESURE, 0.93)
- Gemini 3.8 Flash —mesure→ 0,75 $ par million de tokens en entrée, 3,75 $ en sortie (MESURE, 0.93)
- Claude Fable 5.1 —mesure→ fenêtre de contexte d'un million de tokens (MESURE, 0.92)
- Alan Blount —utilise→ promptfoo (TECHNOLOGIE, 0.93)
- promptfoo —utilise→ opencode (TECHNOLOGIE, 0.93)
- Gemini 3.8 Flash —surpasse→ Claude Fable 5.1 (TECHNOLOGIE, 0.88)
- Gemini 3.8 Flash —mesure→ résumé d'état de branche git en 1,1 s pour moins d'un dixième de cent, Fable 5.1 près de 6 s et 23 fois plus cher (MESURE, 0.9)
- Claude Fable 5.1 —surpasse→ Gemini 3.8 Flash (TECHNOLOGIE, 0.88)
- Claude Fable 5.1 —mesure→ plan de migration Postgres vers Spanner en DAG formel en 12 s, contre une séquence linéaire en 3 s pour Flash (MESURE, 0.9)
- Alan Blount —recommande→ coordination asymétrique (CONCEPT, 0.95)
- coordination asymétrique —est_instance_de→ routage de modèles (CONCEPT, 0.93)
- coordination asymétrique —utilise→ subagents (TECHNOLOGIE, 0.92)
- coordination asymétrique —utilise→ escalade Ask for Help (METHODOLOGIE, 0.92)
- escalade Ask for Help —affirme_que→ le worker rapide traite 85 à 90 % des requêtes et n'escalade que sur ambiguïté, action irréversible ou deux échecs d'outil consécutifs (AFFIRMATION, 0.9)
- Alan Blount —recommande→ 3 à 5 outils à schémas JSON stricts pour le modèle rapide, un catalogue de 50 outils causant confusion de paramètres (AFFIRMATION, 0.9)
- Alan Blount —s_oppose_à→ routeurs de modèles dynamiques (CONCEPT, 0.92)
- routeurs de modèles dynamiques —affirme_que→ signal insuffisant au premier tour, succès non traduisible en features, mauvais choix rejoués sur l'autre chemin (AFFIRMATION, 0.9)
- Alan Blount —affirme_que→ "Keep it boring" (CITATION, 0.95)
- Alan Blount —recommande→ ROI des tokens (CONCEPT, 0.92)
- ROI des tokens —affirme_que→ valeur du temps humain, des fonctionnalités livrées et des incidents évités, moins le coût des tokens et de la montée en compétence (AFFIRMATION, 0.9)
- coordination asymétrique —s_applique_à→ LangGraph (TECHNOLOGIE, 0.85)

---
Canonical: https://www.thekb.eu/en/fiches/blount-googlecloudtech-fable-gemini-flash-same-team-2026-09-14/