The Wren Journal Wren AI's default model is now GPT-5.6 Luna — 16× cheaper and 37% faster per task, with accuracy holding at 95% of evaluation assertions passed. Here's what we measured before flipping the switch.
Wren AI Product Team
Updated: Sep 01, 2026
Published: Sep 01, 2026
Same trusted answers, a fraction of the cost and the wait. Here's what we measured before flipping the switch.
What changed #
We moved Wren's default model from GPT-5.4 to GPT-5.6 Luna. Nothing changes in how you use Wren — same questions, same connectors, same governed SQL behind every answer. The difference shows up on your bill and your clock. Every account is on Luna now, with no action needed on your side.
Before switching, we replayed our full agent evaluation: 22 real ecommerce tasks, from single-fact lookups to multi-turn drill-downs and full GenBI dashboards. Luna handled all of them at a fraction of the cost, and reproduced our previous model's answers to the cent on the core data.
Averaged across all 22 evaluation cases, cost per task dropped from $0.58 on GPT-5.4 to $0.036 on Luna — 16× cheaper per task. Every one of the 22 cases came out cheaper — no exceptions.
The numbers #
Two things compound. Luna's tokens list at roughly one-twelfth the price to begin with, and Luna reaches the same answer while generating about half as many of the expensive output tokens. The list price does most of the work; trimming output tokens does the rest — together they turn a lower sticker price into a far wider real-world gap.
| Metric | Previous (GPT-5.4) | Luna (GPT-5.6) | Change |
|---|---|---|---|
| Spend per task | $0.58 | $0.036 | −94% | | List price, per million output tokens | $15.00 | $1.20 | −92% (input drops the same 12.5×) | | Model time per task | 161.9s | 102.5s | −37% | | Tokens generated per task | 14,119 | 6,902 | −51% |
Across the whole run, the 22 tasks that cost $12.86 on the old default came in at $0.80 on Luna — a saving of $12.06 on the same workload.
Accuracy held #
Cheaper and faster only matters if the answers stay right, with the SQL to prove them. On our evaluation, Luna passed 204 of 215 assertions (95%) and reproduced the previous model's figures to the cent on the core data tasks — total revenue, GMV series, Pareto splits, forecasts and multi-turn reconciliations all landed on the reference values.
95% of evaluation assertions passed on Luna22 / 22 tasks came out cheaper — every case, no exceptionTo the cent— core data answers matched the previous model exactly
You don't need to do anything. Luna is live as the default across all Wren AI Cloud accounts today. Ask your usual questions — you'll just spend less and wait less to get governed answers back.
Same questions. Answered for less. #
Give every team trusted answers from the data they already own — now on GPT-5.6 Luna by default. Try Wren AI Cloud free or request a demo.
See governed GenBI in action
Watch short demos of Wren AI turning business questions into governed SQL, charts, and reusable GenBI apps.
Related Posts #
The Genie Meter Is On: What Databricks' New Pricing Says About the ROI of GenBI
Databricks turned the Genie meter on, and agents get no free tier. Why full autonomy makes an agnostic GenBI layer, any model on any platform, the whole decision.
You Can't Trust an AI Agent You Can't Debug.
An AI agent that answers business questions has to be debuggable and measurable, or 'earned trust' is just a slogan. How thread tracing, benchmarks, and the AI Advisor close the loop.
Reducing Hallucinations in Text-to-SQL Building Trust and Accuracy in Data Access
How Schema Grounding, Semantic Layers, and Iterative Validation Can Enhance Text-to-SQL Reliability
Get the next deep dive in your inbox #
Practical GenBI guides, product updates, and customer lessons from the Wren AI team. A couple of emails a month — no noise.
Keep reading
The Genie Meter Is On: What Databricks' New Pricing Says About the ROI of GenBI Databricks turned the Genie meter on, and agents get no free tier. Why full autonomy makes an agnostic GenBI layer, any model on any platform, the whole decision.
You Can't Trust an AI Agent You Can't Debug. An AI agent that answers business questions has to be debuggable and measurable, or 'earned trust' is just a slogan. How thread tracing, benchmarks, and the AI Advisor close the loop.
Reducing Hallucinations in Text-to-SQL Building Trust and Accuracy in Data Access How Schema Grounding, Semantic Layers, and Iterative Validation Can Enhance Text-to-SQL Reliability