Gemini 3.8 Flash Pricing: Same Cliff, a Hidden Tax

Google shipped Gemini 3.8 Flash on 2 September at $0.75 per million input tokens and $3.75 per million output. That is the figure in every one of the fifteen results I pulled this morning for Gemini 3.8 Flash pricing, and every one of them is correct. It is also the same figure Gemini 3.7 Flash launched at in August, and the same figure 3.6 Flash was cut to retroactively on the same day, and all three of them expire on the same date: 31 December 2026. On 1 January the price doubles.

We covered the cliff when it first appeared under Gemini 3.7 Flash. What is new is that Google has now shipped a third model onto it, and the pricing table for that model carries four words that no result on the page-one SERP quotes: the output price is listed as “including thinking tokens”. I wanted to know what that costs on a real WordPress chatbot rather than a benchmark, so I ran one through the API from this site’s own server. The short version: on a typical support reply, 71% of the output tokens you are billed for are ones nobody ever reads.

What Google’s pricing page actually says

Everything in this table is read from ai.google.dev/gemini-api/docs/pricing on 11 September 2026, not from a tracker. Prices are per million tokens, paid tier.

Three things fall out of that table that the launch coverage did not say.

First, three generations share one price and one date. 3.6, 3.7 and 3.8 Flash are identical on every meter, including context caching ($0.075 now, $0.15 in January) and cache storage ($0.50 per million tokens per hour now, $1.00 in January). There is no pricing reason to stay on an older Flash model, and Google’s own launch post now positions 3.7 Flash as the model that “remains fully supported for efficiency-first workloads” — the polite way of saying the new one is for harder jobs.

Second, the January price is still cheaper than 3.5 Flash. This is the one people miss. The doubled 3.8 Flash rate of $1.50 / $7.50 undercuts the standing, non-promotional 3.5 Flash rate of $1.50 / $9.00 on output. So “the price doubles in January” is true, and “after January you might as well use the old model” is false. The old model was never the cheap one.

Third, every output price on this page carries the same qualifier: including thinking tokens. Gemini 3.x models reason before they answer, and the reasoning is billed at the output rate. How much reasoning a support reply triggers is not on the pricing page. It is not on any of the fifteen results either. It is measurable, though, so I measured it.

The tokens you never see, measured on a live chatbot

The test is the shape of a real reply on this site. The system prompt is our support assistant’s instructions plus the retrieved text of the MxChat documentation page — about 5,720 input tokens, which is what a retrieval-backed WordPress chatbot sends on every turn whether the visitor typed three words or thirty. Then one visitor question. I ran five questions against Gemini 3.8 Flash at its default settings, five more with thinking switched off, and single samples of 3.7, 3.6, 3.5 Flash and 3.5 Flash-Lite for comparison. Every call went through Google’s OpenAI-compatible endpoint, which reports prompt_tokens, completion_tokens and total_tokens; the difference between the total and the other two is the thinking the model did and you paid for.

The replies themselves were, to my eye, the same answer. Every 3.8 Flash run — thinking or not — told the visitor to open the MxChat Chatbot tab, find the chat-model setting, choose a Gemini model and paste a key, in the same order, with the same caveats. The visible reply length barely moved: 166 tokens with thinking, 178 without. What moved was the bill. At default settings the model spent, on average, 412 tokens deciding what to say and 166 saying it. That is 71% of the billed output going to reasoning that a question about which dropdown to click did not need.

A word on what “off” means here, because Google’s documentation says reasoning cannot be turned off for Gemini 3 models. What I can see is the bill. With reasoning_effort set to none — or to low, which is the value Google’s own 3.8 Flash example uses — the usage block reported zero hidden tokens on every run, and the native endpoint did the same with thinkingLevel: "low" or thinkingBudget: 0. Whether the model still reasons somewhere I cannot see is not my problem; whether I am charged for it is, and I was not.

Two smaller findings from the same runs. The spread on thinking is wide — 183 tokens on the easiest question, 498 on the hardest of the five — so a single measurement would have misled either way. And the older the Flash model, the more it thinks about the same question: 3.5 Flash spent 551 tokens on it, 3.8 Flash 412. That cuts against Google’s own launch post, which warns that 3.8 Flash “works harder” and “might use more tokens to maximize performance, especially at higher effort levels”. Both can be true: the warning is about complex agentic tasks, and a question about which dropdown to click is not one. For a support chatbot, the newer model is not just cheaper per token; it used fewer of the tokens you cannot see.

The 22-token question that billed 315 tokens

If the retrieval prompt makes that look abstract, here is the smallest possible version. I sent 3.8 Flash a 22-token prompt — a one-line system instruction and “In one sentence, what is a WordPress plugin?” — with streaming on, the way a live chat widget calls it. The reply was 25 tokens. The usage block at the end of the stream said total_tokens: 315. Two hundred and sixty-eight tokens of thinking for a one-sentence definition, billed at $3.75 per million. The dollar amount is nothing. The ratio is the point: on a short conversational turn, thinking was ten times the size of the answer, and the streaming widget gave the visitor no sign it had happened.

What it costs per 1,000 replies, now and in January

Same 5,720-token prompt, measured output from the table above, priced from Google’s page. The “thinking off” rows use the reasoning_effort: none figure, which reported no hidden tokens on all five runs; low behaved the same on the bill.

Read the first two rows against each other. Switching thinking off on 3.8 Flash saves $1.50 per 1,000 replies today — 23% of the bill — and $3.00 per 1,000 after the cliff, for replies I could not tell apart. A site doing 100,000 conversations a month is looking at roughly $646 versus $496 now, and $1,292 versus $992 from January.

Read the 3.5 Flash row against the 3.8 rows. A site still on 3.5 Flash at default settings pays 2.4 times what 3.8 Flash costs today for the same reply, and 1.2 times what 3.8 Flash will cost after its price doubles. That row matters more than it looks, for a reason I will come to.

And notice the last column. Input is the majority of every bill, exactly as we found when we priced GPT-5.6 Luna per 1,000 replies on a similar prompt shape — because a retrieval chatbot re-sends thousands of tokens of context on every turn. The thinking tax is a real 23%. The context you send is still the bigger lever, which is why the tokenizer differences between vendors end up mattering more than the headline output rate.

Running 3.8 Flash from a WordPress chatbot today

Here is the honest part, because this is our plugin. The Google Gemini dropdown in MxChat 3.2.20 tops out at Gemini 3.5 Flash, which it labels the recommended default. Google has shipped three Flash generations since that list was written, all of them cheaper on every meter than 3.5 Flash, and the fastest of them doubles in price in sixteen weeks. That is a catalog update we owe you, and I have filed it. You do not have to wait for it, though. Two routes work today, and I tested the first one from this server this morning.

Route 1 — the custom provider. Google runs an OpenAI-compatible endpoint, and the plugin’s “Custom (OpenAI-compatible)” provider is built for exactly that shape: it appends /chat/completions to whatever base URL you give it and sends your key as a bearer token, which is what Google expects. In the API Keys tab:

• Base URL: https://generativelanguage.googleapis.com/v1beta/openai
• API key: your Gemini API key from AI Studio (the same key the Gemini dropdown uses)
• Model name: gemini-3.8-flash

Then pick Custom Provider as the chat model. Streaming works — I confirmed the endpoint returns the standard data: chunk stream, and the plugin’s custom-provider path is a streaming client. Function calling, which MxChat uses for its tools, is supported on the compatible endpoint as well.

Route 2 — OpenRouter. OpenRouter lists google/gemini-3.8-flash at Google’s own rate, and the plugin’s OpenRouter provider loads the model list from your OpenRouter key. Slightly more moving parts, one more account, but no base URL to type.

The caveat either way is the thinking tax. The plugin sends a reasoning_effort parameter only for OpenAI’s GPT-5 family; on a Gemini model — native or through the custom provider — it sends nothing, so the model thinks at its default level and you pay the 412 tokens. Until the catalog update lands with a thinking control, a 3.8 Flash reply through MxChat costs the $6.46 row, not the $4.96 one. That is still 2.4 times cheaper than the current recommended default, so the switch is worth making today; it is just not the whole saving. If you call the API from your own code instead, reasoning_effort: "low" on the compatible endpoint or thinkingConfig: {thinkingLevel: "low"} on the native one both returned zero hidden tokens in my runs — and note that 3.5 Flash-Lite rejects the parameter with a 400, because it does not think in the first place.

Batch, Flex and Priority: three of the four tiers are not for chatbots

The pricing page lists four service tiers for 3.8 Flash, and the launch coverage tends to quote whichever is cheapest. For a chatbot only one applies.

Batch and Flex are half price, and neither can answer a visitor who is waiting. That is the same shape we found with the OpenAI Batch API: the discount reaches the parts of an AI feature that run overnight — embedding your knowledge base, summarising transcripts — and none of the reply bill. Priority is the reverse: it buys throughput guarantees at 1.8 times Standard, which a site doing a few thousand conversations a month does not need. Grounding with Google Search is metered separately at $14 per 1,000 requests after 5,000 free a month, shared across all Gemini 3 models, so a chatbot that searches on every turn is buying a second product.

One more model appeared on 2 September that you will see in headlines and cannot buy: Gemini 3.8 Flash Cyber, tuned for vulnerability detection and patching, is available only through Google’s Fairwind Program to government bodies, critical-infrastructure operators and software maintainers. It has no price on the public page because it is not on the public API.

What to do before 31 December

The cliff is sixteen weeks out and Google has published both sides of it, which is more than most vendors do — Anthropic cancelled its own scheduled increase in August, and OpenAI published GPT-5.6 Sol’s rate as holding “at least through 21 November” without saying what follows. Given a known date and a known number, the plan writes itself.

• If you are on Gemini 3.5 Flash, move. Every Flash model above it is cheaper today and after January. There is no scenario in which staying is the lower bill.
• Measure your own prompt shape. Ours is roughly 5,700 input tokens per turn; the earlier 3.7 Flash piece measured a multi-turn figure of about 7,200. Yours depends on how much retrieved context you attach. Input is the majority of the bill on every row above, so this number matters more than the model choice.
• Budget the January rate, not the September one. A site whose Gemini line is $650 a month in October will see $1,300 in January with no change in traffic. Put it in the spreadsheet now.
• Turn thinking off for support replies where you can. The replies did not get worse in my five-question test. They got cheaper by 23% and, on the median run, faster.
• Read the vendor’s page, not this one, before you change anything. Every figure here was current on 11 September 2026. Google has changed its Flash line-up twice in four weeks. The full cost breakdown for a WordPress chatbot covers the parts of the bill that are not tokens, and the MxChat documentation covers the custom-provider fields.

Frequently asked questions

What does Gemini 3.8 Flash cost per million tokens?

$0.75 input and $3.75 output on the paid Standard tier through 31 December 2026, then $1.50 and $7.50 from 1 January 2027. Context caching is $0.075 per million now and $0.15 in January. The output price includes thinking tokens.

Is Gemini 3.8 Flash cheaper than 3.7 Flash?

Per token, no — they are priced identically on every meter, and so is 3.6 Flash. In practice 3.8 Flash was slightly cheaper in my runs because it produced fewer hidden thinking tokens on the same question: 412 against 3.7 Flash’s 512.

Do thinking tokens really get billed?

Yes. Google’s pricing page lists every output price as “including thinking tokens”, and the API’s usage block reports them: on a 5,720-token support prompt, 3.8 Flash returned a 166-token reply and billed 578 output tokens by default. Setting reasoning_effort to low or none on the OpenAI-compatible endpoint brought the reported thinking to zero.

Can I use Gemini 3.8 Flash with a WordPress chatbot plugin now?

Yes. In MxChat, use the Custom (OpenAI-compatible) provider with base URL https://generativelanguage.googleapis.com/v1beta/openai, your Gemini API key, and model gemini-3.8-flash, or select google/gemini-3.8-flash through OpenRouter. The native Gemini dropdown will list it once the catalog is updated.