Decision guide
Gemini 3.6 Flash vs GPT-6 Luna: legacy pin decision
Historical comparison: use current model profiles for current rates, IDs and availability. 3.6 Flash remains in the catalog for teams already pinned to that id. For new Google Flash work, 3.8 Flash (2026-09-02) is the current coding/agent row; on the OpenAI side GPT-6 Luna ($0.10/$0.50) replaced GPT-5.6 Luna on 2026-09-22.
Switch if
- You are already pinned to `gemini-3.6-flash` and a cutover to a newer model is blocked by id, evaluation, or contract.
- Your existing 3.6 Flash batch still meets validators at the catalogued introductory $0.75/$3.75 list.
- You need this older Flash row's multimodal input and cannot move the Google route yet.
Stay if
- You can evaluate Gemini 3.8 Flash instead — that is the current Google Flash coding/agent row in this catalog.
- GPT-6 Luna already meets your pass rate. Its current $0.10/$0.50 catalog list is below Gemini 3.6 Flash's $0.75/$3.75 introductory rate.
- You are choosing a new Google Flash id rather than keeping a pin — start from 3.8 Flash.
Check before production traffic moves
- 01Run the same high-volume batch on both models with identical validators.
- 02Include multimodal inputs only where your production workload actually sends them.
- 03Compare p95 latency, rate limits, and retry behaviour — not only the rate card.
This is an older-Flash pin-in-place decision, not the current Google Flash shortlist. 3.6 Flash still brings multimodal breadth at the catalogued introductory $0.75/$3.75; GPT-6 Luna's current catalog list is $0.10/$0.50. New Google Flash work should start from Gemini 3.8 Flash.
Run a controlled trial first
A second option in a labelled pilot is cheaper than a production cutover. Score the trial, then migrate or roll back on evidence.
Pilot plan
- Select a high-volume slice with repeatable inputs and a known error budget.
- Run GPT-6 Luna and Flash on the same batch, including structured outputs and tool calls.
- Track cost per accepted task, latency p95, and reviewer minutes on a human sample.
Score the trial
- Modality fit
- Pass rate on the inputs your pipeline actually sends.
- Cost per success
- Tokens, retries, and review divided by accepted tasks.
- Throughput
- Rate limits, queue time, and p95 latency under load.
- Integration cost
- Adapter changes, auth, logging, and support path.
Migration sequence
- Add Flash behind a feature flag with Luna as the default fallback.
- Route only the pilot cohort first; cap spend and monitor validator failures.
- Expand only when cost per success and latency stay within tolerance.
Rollback plan
- Keep Luna routing and prompt templates provider-specific for instant switch-back.
- Stop Flash traffic if validation or latency breaches thresholds; replay safe failures.
- Archive batch logs so a later re-test can separate release changes from integration bugs.
Hidden costs to price
- Multimodal tiers can have separate limits or surcharges beyond the headline rate.
- A second cloud provider adds billing, policy review, and on-call complexity.
- Search grounding and tools on Gemini can add separate line items.