Gemini 3.8 Flash is live in EBITDAI on the Pro plan

What Google shipped

Google calls 3.8 Flash its most intelligent Flash model, built for long-horizon software engineering, autonomous agents and complex multi-step reasoning. That description happens to match what an Excel modeling agent does all day: dozens of dependent steps against a workbook that changes underneath it.

List price is $0.75 per 1M input tokens, $0.075 per 1M cached input tokens and $3.75 per 1M output tokens. That is introductory pricing and it holds through December 31, 2026. On January 1, 2027 it doubles, to $1.50 and $7.50.

The benchmarks Google published

Google's announcement leads with DeepSWE v1.1, a long-horizon software engineering test. The one we watch more closely is Vals Finance Agent v2, which scores financial analyst tasks. Both charts below use Google's own figures, rebuilt from the model evaluation it published with the launch.

On DeepSWE v1.1, Gemini 3.8 Flash scores 73.7%, three tenths of a point behind Claude Opus 5 at 74.0% and ahead of every other model in the table. On Vals Finance Agent v2 it finishes first at 61.4%, against 58.6% for Opus 5. Opus 5 lists at $5.00 per 1M input tokens. Gemini 3.8 Flash lists at $0.75.

Pro: one plan, every model

Pro is $15/month, or $144/year. It runs Gemini 3.8 Flash, Kimi k3 and DeepSeek V4 on one shared pool of included monthly usage, 7.5x what Lite carries, and you can switch models mid-session without losing your place in a build.

Lite is $5/month with 100x more usage than free, on Meta Muse Spark 1.3 Contributor. Pro takes your own API keys for every provider and adds the QuickBooks, Campfire and Stripe connectors. The free plan gives you 150 requests a week on Meta Muse Spark 1.3 with no card, and Meta trains on everything you send with that model.

Speed, measured on our own relay

We ran all three models through EBITDAI's relay on September 2, 2026. Short requests, the kind that make up most of a modeling session, round-tripped in 1.8 to 2.7 seconds on Gemini 3.8 Flash. Kimi k3 through Moonshot took 8.7 to 22 seconds per step, with Moonshot overloaded for much of the day, so read the top of that range as a bad-day number rather than a steady state. DeepSeek V4.1 Flash landed between 2.5 and 12 seconds.

Those are our numbers, from our relay, on one day. They are not a universal claim about the models, and they will move with provider load. What they do explain is why Gemini is now the model we reach for when a build has thirty steps in it.

Where your workbook data goes

Of the three vendors EBITDAI offers, Google is the US-based one, and its paid API terms state plainly that your prompts are not used to train its models. The Gemini API terms say that for paid services, "Google doesn't use your prompts (including associated system instructions, cached content, and files such as images, videos, or documents) or responses to improve our products," and that prompts and responses are logged "for a limited period of time, solely for detecting and preventing violations of the Prohibited Use Policy." A US-based provider, on enterprise-grade infrastructure, with that written commitment.

For comparison: DeepSeek's models are served by Hangzhou DeepSeek Artificial Intelligence Co. in China, and Kimi k3 by Moonshot AI in Singapore. All three are good models. But if your firm has a view about where client numbers travel and who is allowed to learn from them, Gemini 3.8 Flash is the most conservative of the three choices, and it is now one click away inside the same plan.

Since September 6, 2026 the included set also offers Meta Muse Spark 1.3, served on Meta's Contributor tier. Every prompt, reply and workbook value you send with that model goes to Meta, and Meta uses it to train its models. It is an extra option on the paid plans, never the default, and the free plan now runs on it alone. Pick one of the other models if that is a problem for a given workbook.

Since September 7, 2026 Pro also carries the same model on Meta's standard tier, which the picker labels "not used for training". Same Muse Spark 1.3, different terms: Meta's published pricing terms say content submitted on the standard tier is not used to train its models, while the Contributor tier (the free plan's model, and the "trains on your data" option on the paid plans) is. Lite carries the Contributor tier only.

EBITDAI itself never stores prompts or workbook content on any paid plan (free-plan conversations are logged for product review, as the security page explains). On Pro in own-key mode you deal with the provider directly and we add no markup. Gemini has no browser-callable endpoint, so those calls pass through our stateless relay, which forwards the request and stores no prompt or workbook content; own-key DeepSeek and pay-as-you-go Kimi calls go straight from the add-in to the provider.

What the tokens cost

List prices per 1M tokens, pay-as-you-go, the rates you pay in own-key mode:

Gemini pricing is introductory through December 31, 2026; from January 1, 2027 it is $1.50 input, $0.15 cached and $7.50 output. DeepSeek rates shown are off-peak and double at peak, which is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday; DeepSeek V4 Pro is exactly 3x Flash. Kimi Code, the flat subscription you buy from Moonshot, carries no per-token charge at all.

A real job from today makes the spread concrete. Relinking an 80-row historical income statement to its source tab took 7 model calls, about 223K prompt tokens of which about 176K were cache hits, and about 2.4K output tokens. On Kimi k3 that job costs $0.23. On Gemini 3.8 Flash it costs $0.057. On DeepSeek V4 Flash, off-peak, it costs $0.013. On a Kimi Code subscription it costs nothing per token, which is still the cheapest way to run heavy days if you already have one.

Bring your own Gemini key

On Pro you can paste your own Gemini API key from aistudio.google.com/apikey and run at Google's prices with zero markup from us. Gemini cannot be called from a browser, so those calls pass through our stateless relay, which stores no prompt or workbook content. The setup is four fields and a test call, walked through step by step in the Gemini key tutorial.

If you would rather not manage a key, Pro at $15/month includes usage on all of them and you can start on Gemini 3.8 Flash today.