Hacker News Reader

Hacker News

Comment by simonw | Hacker News Reader

Full thread

simonw

They don't mention it in the post, but it looks like this includes a price increase for the Gemini 2.5 Flash model.

For 2.5 Flash Preview https://web.archive.org/web/20250616024644/https://ai.google...

$0.15/million input text / image / video

$1.00/million audio

Output: $0.60/million non-thinking, $3.50/million thinking

The new prices for Gemini 2.5 Flash ditch the difference between thinking and non-thinking and are now: https://ai.google.dev/gemini-api/docs/pricing

$0.30/million input text / image / video (2x more)

$1.00/million audio (same)

$2.50/million output - significantly more than the old non-thinking price, less than the old thinking price.

View on HN

The blog post has more info about the pricing changes

https://developers.googleblog.com/en/gemini-2-5-thinking-mod...

The real news is that non-thinking output is now 4x more expensive, which they of course carefully avoid mentioning in the blog, only comparing the thinking prices.

How cute they are with their phrasing:

> $2.50 / 1M output tokens (*down from $3.50 output)

Which should be "up from $0.60 (non-thinking)/down from $3.50 (thinking)"

Is it possible to get non-thinking only now, though? If not, why would that matter, since it's irrelevant?

Yes, by setting the thinking budget to 0. Which is very common when a task doesn't need thinking.

In addition, it's also relevant because for the last 3 months people have built things on top of this.

interesting - why wouldn't you use dynamic thinking? and yeah, sucks when the price changes.

It makes responses much slower with zero benefit for many tasks. Flash with thinking off is very fast.

To be fair, the point of preview models and stable releases is so you know what is stable to build on.

The moment you start charging for preview stuff I think you give a tacit agreement that you can expect the price to not increase by a factor of 4.

that’s a somewhat naïve viewpoint.

I think the fact that everyone is like ‘wtf’ now kind of reinforces my viewpoint?

Doesn’t mean you can’t do it, but people won’t be happy.

Who cares about happy (in the short term), as long as they continue to pay.

Gmail was in beta for what, 2 decades? Did you never use it during that time? They've been using these "Preview" models on their non-technical user facing Gemini app and product for months now. Like, Google themselves has been using them in production, on their main apps. And gemini-1.5-pro is 2 months from depreciation and there was no production alternative.

They told everyone to build their stuff on top of it, and then jacked up the price by 4x. Just pointing to some fine print doesn't change that.

I'd be more worried about Google just discontinuing another product. For example Stadia was similarly high profile, but it's gone now.

More examples here: https://killedbygoogle.com/

one example where non-thinking matters would be latency-sensitive workflows, for example voice AI.

Correct, though pretty much anything end-user facing is latency-sensitive, voice is a tiny percentage. No one likes waiting, the involvement of an LLM doesn't change this from a user PoV.

I wonder if you can hide the latency, especially for voice?

What I have in mind is to start the voice response with a non-thinking model, say a sentence or two in a fraction of a second. That will take the voice model a few seconds to read out. In that time, you use a thinking model to start working on the next part of the response?

In a sense, very similar to how everyone knows to stall in an interview by starting with 'this is a very good question...', and using that time to think some more.

They seem just rebrand the non-thinking model to flash-lite, so it’s less expensive than before

Not at all. Non-thinking flash is... flash with the thinking budget set to 0 (which you can still run that way, just at 2x input 4x output pricing). Flash-lite is far weaker, unusable for the overwhelming majority of usecases of flash. A quick glance at the benchmark reveals this.

Yeah, so basically their announcement is "good news, we tripled the price, and will deprecate Gemini Flash 2.0 asap"

The OP says Flash-Lite has thinking and non-thinking, so it’s not that simple.

I have LLM fatigue, so I'm not paying attention to headlines... but LLMs are thinking now? That used to be a goal post. "AI can't do {x} because it's not thinking." Now it's part of a pricing chart?

How did I miss this?

"Thinking" means spamming a bunch of stream-of-consciousness bs before it actually generates the final answer. It's kind of like the old trick of prompting to "think step by step". Seeding the context full of relevant questions and concepts improves the quality of the final generation, even though it's rarely a direct conclusion of the so-called thinking before it.

"Thinking" really just means "write on some scratch paper" for llms.

Do you work for google?

“While we strive to maintain consistent pricing between preview and stable releases to minimize disruption, this is a specific adjustment reflecting Flash’s exceptional value, still offering the best cost-per-intelligence available.”

Anthropic did the same thing with their Haiku model when they released version 3.5. I hate it.

Pricing is a hard problem. Theoretically, if companies occasionally raise prices dramatically once something is useful, they sometimes can create early demand and more testers for future product releases. Ofc they have to be careful to avoid annoying regular users too much. When you sell the harm is limited to late users, but when you rent it is harder to figure out the optimal strategy.

> Today we are excited to share updates …

They are obviously excited about their price increase

"Soon, AI too cheap to meter" "Meantime, price go up".

"will be too cheap to meter" means we're definitely metering it now.

There are a lot more price drops, though.

From prices that were already losing services money.

If you aren't making a profit, lowering prices is only about trying to capture market share before you're forced to increase prices to remain solvent.

We have likely seen the cheapest prices already. Once we can’t function without them anymore - go as high as you can!

Hopefully we get more competition and someone willing to undercut the more expensive options

Entering the market and being competitive gets more difficult all the time. People want the best and fastest models - can you compete with trillion dollar datacenters?

You might be right, but there's plenty of deep pocketed companies who are still very excited to compete in this market.

It's more likely the shareholder zeitgeist will soon shift to demanding returns on the ungodly amounts already invested into AI.

By then comparable or even better models will easily run on edge.

So if they crank up the prices we could just switch to local and not get lured by bigger and bigger models, rag, Agentic, MCP driven tech as if all of that couldn't run locally either.

I am not as optimistic that locally run models will be able to compete anytime soon. And even if, the price to run them means you have to buy the compute/gear for a price that is likely equivalent to a lot of 'remote' tokens

The most meaningful models will run in the future on those trillion dollar data centers that are currently being build.

> By then comparable or even better models will easily run on edge.

What are you basing that on?

Presumably your goal is to extract some practical value from this and not just higher benchmark numbers. If you can get the functionality you need from last-gen, there's no point in paying for next-gen. YMMV.

Indeed that was the premise that we would be a step behind when running on edge. That's already the case.

You know that competition is a thing, do you?

Not too long ago Google was a bit of a joke in AI and their offerings were uncompetitive. For a while a lot of their preview/beta models had a price of 0.00. They were literally giving it away for free to try to get people to consider their offerings when building solutions.

As they've become legitimately competitive they have moved towards the pricing of their competitors.

No way. AI pricing is going up because people are willing to pay for it.

Just google. They were behind. So they just dumped their prices to get a foot in the door. Now they are popular and can raise it to market prices.

I still don’t think there’s any real stickiness to using a Google model over any other model, with things like openrouter. So maybe for brand recognition alone.

Yea, but brand have some stickiness. Maybe not for the absolute nerds but lots of people just stick to what they are already using. Look at all the people just using ChatGPT because that is what they tried first.

True, but that’s a website, of which the set of them come with bespoke features. I’d assume someone who adopted the API knows it’s easier to switch because it’s an open standard.

I know they undercutting the price a lot, because at first launch gemini price is not make sense seeing it cheaper than competition (like a lot cheaper)

finally we starting to see the real price

A cool 2x+ price increase.

And Gemini 2.0 Flash was $0.10/$0.40.

4x price increase over preview output for non-thinking.

1.5 -> 2.0 was a price increase as well (double, I think, and something like 4x for image input)

Now 2.0 -> 2.5 is another hefty price increase.

FWIW: On OpenRouter, the non `:thinking` 2.5 flash endpoint seems to be returning reasoning tokens now.

Good catch, that's a pretty notable change considering this was about to be the GOAT of audio-to-audio

You can also see this difference in open router.

But why is there only thinking flash now?

Apparently, you can make a request to 2.5 flash to not use thinking, but it will still sometimes do it anyways, this has been an issue for months, and hasn't been fixed by model updates: https://github.com/google-gemini/cookbook/issues/722

It might be a bit confusing, but there's no "only thinking flash" - it's a single model, and you can turn off thinking if you set thinking budget to 0 in the API request. Previously 2.5 Flash Preview was much cheaper with the thinking budget set to 0, now the price is the same. Of course, with thinking enabled the model will still use far more output tokens than the non-thinking mode.

Interesting design choice, and makes me think of "Thinking, Fast and Slow" by Kahneman.

(I thought of it quickly, not slowly, so the comparison may only be surface deep.)

Reply on news.ycombinator.com