Gemini 2.5 flash: $0.30/$2.50
Gemini 3.0 flash preview: $0.50/$3.00
Gemini 3.5 flash: $1.50/$9.00
Interesting pricing direction. I don't think we have ever seen a 3x price increase for in the immediate next same-sized model (and lol @ 3 only ever getting a preview).
3.5 flash costs similar to Gemini 2.5 pro which was $1.25/$10
Gemini 2.5 flash (27 score): $172 (1.0x)
Gemini 2.5 pro (35 score): $649 (3.8x)
Gemini 3.0 Flash (46 score): $278 (1.6x)
Gemini 3.5 Flash (55 score): $1,552 (9.0x or 2.4x compared to 2.5 pro)
This is a massive price increase... 5.6x compared to Gemini 3.0 Flash
Qwen 3.6 hit hard in the self-hosting space. It's incredibly capable for its size, really shaking up what's possible in 64GB or even 32GB of VRAM.
The Prism Bonzai ternary model crams a tremendous amount of capability into 1.75GB.
And, DeepSeek V4 is crazy good for the price. They're charging flash model prices for their top-tier Pro model, which is competitive with the frontier of a few months ago.
The winners in the AI war will be the companies that figure out how to run them efficiently, not the ones that eke out a couple percent better performance on a benchmark while spending ten times as much on inference (though the capability has to be there, I think we're seeing that capability alone isn't a strong moat...there's enough competent competition to insure there's always at least a few options even at the very frontier of capability).
You can lower that to at least 24GB. I've been running Qwen 3.5 and 3.6 with codex on a 7900 XTX and the long horizon tasks it can handle successfully has been blowing my mind. I would seriously choose running my current local setup over (the SOTA models + ecosystem) of a year ago just based on how productive I can be.
This is what you get for relying on the generosity of billionaires. Keep offshoring your thinking ability to a machine and let me know how competitive you. Hint, you wont be. There's nothing special about being able to use an LLM.
Please go run some numbers.The hardware needed to Run Deepseek v4 flash at 20 tps for a single session is nowhere close to what is required to run it at 50tps for 5,000 concurrent sessions.
Imagine what it takes to be profitible when running at 150 tps for 30cents per 1mm. You make less than 1k per month and the hardware required to run that cost 10k a month to rent with hardly any concurrent session capability.
Or maybe they think because their benchmarks are good they can ramp up the prices. Seems like they don’t have the market share to justify a move like that yet to me.
My guess: it's the price at which they make more money than if they rent the TPUs to other companies.
The Gemini team has had trouble securing enough TPUs for their user's needs. They struggle with load and their rate limits are really bad. Maybe at a higher price, they have a better chance at getting more TPUs assigned?
Just because you are vertically integrated doesn't mean you get to discount the one business units products to the other. Doing so discounts the opportunity cost you pay and is just bad accounting.
Open-source model inference providers (who do not have to bear the cost of training) seem able to do it at much lower prices.
https://www.together.ai/pricing
https://fireworks.ai/pricing#serverless-pricing (scroll down to headline models)
Of course, it's possible that they are burning through investor cash as well, and apples-to-apples comparisons are not possible because AFAIK Google does not mention the size/paramcount for 3.5 Flash.
But if the prevailing wisdom is true, I think it's actually encouraging. It suggests that OpenAI and Anthropic could perhaps, if they need to, achieve profitability if they slow down model development and focus on tooling etc. instead. If true that's probably good news for everybody w.r.t. preventing a bursting of this economic bubble.
...my opinions here are of course, conjecture built on top of conjecture....
You have free local models for most tasks, $20 subscriptions for near-frontier intelligence, and API per token costs for frontier intelligence.
Flash seems to be targeting the near-frontier category.
https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flas...
3.1 flash lite isn’t quite as good as 3 flash preview (which is the most incredible cheap model… I really love it) — but 3.1 is half the price and the insane speed opens up different use cases.
For comparison, Opus models are $5/$25
Since Gemini 3.5 Flash is raising the price to $1.50/$9.00, it's priced between Haiku and Sonnet. If it outperforms Sonnet, it remains a good value, I guess. Though DeepSeek V4 Flash is much cheaper than all of them, and seemingly competitive.
I use it _a lot_ and it’s very capable if you just plan correctly. I actually almost exclusively use 3.1 flash lite and 2.5 flash lite (even cheaper) and we have 99.5% accuracy in what we do.
That said, I think we’ll see the lite/flash models and the pro models will diverge more price wise. The pro models will become more and more expensive.
Inference alone is certainly profitable. I'm running models at home that are comparable to performance of paid models a year or so ago for free. Even for much larger models the cost around inference serving are clearly manageable.
Training is where the costs are, but I'm increasingly convinced those too could have costs dramatically reduced if necessary. Chinese companies like Moonshot.ai are doing fantastic work training frontier models for a fraction of the cost we're seeing from Anthropic/OpenAI.
This isn't like Uber or Doordash where the economics fundamentally don't make sense (referring to the early days of these services where rates were very cheap).
It's a compelling story that "current AI is unsustainable", but it doesn't pan out in practice for a multitude of reasons (not the least of which is that we can always fall back to what models did last year for basically free).
Profitable maybe, in terms of having low costs, but why pay Google or whoever when you can do it yourself for cheaper/"free"?
Ed Zitron and Gary Marcus are... confused.
Even anthropic who does not own any hardware still have a big margin providing claude models.
Fwiw it’s beating Claude Sonnet in most benchmarking (benchmaxxing?), yet they’ve priced it almost half off on a per token basis.
Question is are you going to persuade anyone with this argument?
Are there many devs at Google who legit prefer Gemini over Claude and Codex? Would love to hear about that.
A few weeks ago, Steve Yegge claimed he'd heard that Google employees are banned from using Claude & Codex.
https://x.com/Steve_Yegge/status/2046260541912707471
A number of Googlers replied to say that was totally false, including Demis Hassabis, but they were all on the DeepMind team.
https://x.com/demishassabis/status/2043867486320222333
This person here claims they left Google because of the ban, and because the ban applied outside of Google work as well:
and far cheaper than comparable models, Gemini Pro is cheaper than Claude Sonnet (Anthropic still gets to charge a brand premium)
Not the most intelligent but perfect balance of cheap, fast and not-too-dumb.
I mean, the benchmarks for Gemini 3.5 Flash are very strong, but at those prices it has to be. I guess the time of subsidized tokens from the big guys is slowly coming to an end.
> Create animated SVG of a frog on a boat rowing through jungle river. Single page self contained HTML page with SVG
https://gistpreview.github.io/?5c9858fd2057e678b55d563d9bff0...
3.5 Flash: Thinking High - 7280 tokens
https://gistpreview.github.io/?1cab3d70064349d08cf5952cdc165...
3.1 Pro - 28,258 tokens
https://gistpreview.github.io/?6bf3da2f80487608b9525bce53018...
Though 3.1 took 3 minutes of thinking to generate, but it only one that got animated movement.
https://gistpreview.github.io/?3496285c5dac5ba10ebbc0b201a1a...
Gemini 2.5 Pro - 5,325 tokens:
https://gistpreview.github.io/?cc5e0fefeaaffecd228c16c95e736...
Gemini 2.5 Flash - 7,556 tokens:
https://gistpreview.github.io/?263d6058fe526a62b8f270f0620ec...
Gemma 4 31B IT - 3,261 tokens via AI Studio:
https://gistpreview.github.io/?858a42b96af864859a3b89508619d...
Gemma 4 26B A4B IT - 4,034 tokens via AI Studio:
https://gistpreview.github.io/?4adb7703897e0c6b583f9de928e4a...
https://gistpreview.github.io/?da742884e5e830ce71ee4db877519...
OFC this is just for fun, but nevertheless gave me working code on first try.
https://gistpreview.github.io/?557f979c82701862bc26d24f10399...
> Create animated SVG of a frog on a boat rowing through jungle river. Single page self contained HTML page with SVG. Use the Brave Browser to verifty that the image is indeed animated and looks like a proper rowing frog; iterate until you are satisfied with it.
It was able to discover and fix an animation bug, but the result is still far from perfect: https://gistpreview.github.io/?029df86d03bfe8f87df1e4d9ed2f6...
8112 tokens @ 52.97 TPS, 0.85s TTFT
https://gistpreview.github.io/?7bdefff99aca89d1bc12405323bd4...
Full session: https://gist.github.com/abtinf/7bdefff99aca89d1bc12405323bd4...
Generated with LM Studio on a Macbook Pro M2 Max
https://huggingface.co/hesamation/Qwen3.6-35B-A3B-Claude-4.6...
[1] https://github.com/htdt/godogen
[2] https://drive.google.com/file/d/1ozZmWcSwieZQG0muYjbj7Xjhhlz...
Now think, plan how to tell this story in a cartoon, make scene outline and then generate SVG animation story for "Three Little Pigs" in self contained HTML page. Just single animation no control buttons.
Actual results for models, one shot:
Gemini 3.5 Flash - Three Little Pigs - 9,050 tokens:
https://gistpreview.github.io/?ed9faa53604035005cae86c63c766...
Gemini 3.1 Pro - Three Little Pigs - 24,272 tokens:
https://gistpreview.github.io/?f506bbfd9b4459c8cd55d89605af8...
Gemini 3.5 Flash - Three Little Pigs - 9,050 tokens:
https://gistpreview.github.io/?ed9faa53604035005cae86c63c766...
Gemma 4 31B IT - Three Little Pigs - 5,494 tokens:
https://gistpreview.github.io/?a3aa75abbe8fd7818b73f6fa55ee6...
Gemma 4 26B A4B IT - Three Iittle Pigs - 6,375 tokens:
https://gistpreview.github.io/?1e631caebeb54f9f0cd6d0e3d4d5e...
Latest update: May 2026
I have a very bad feeling about this lag.
With strong tool use, it maybe doesn't even matter that the models are using older data. They can search for updated information. Though most models currently don't, without a little nudge in that direction.
Also, I believe the Qwen 3 series are all based on the same base model, with just fine-tuning/post-training to improve them on various metrics. Maybe everything in the Gemini 3 series is the same, and maybe they're concurrently training the Gemini 4 base model with updated knowledge as we speak.
This actually really does matter. Otherwise, the model simply won't 'know' about your product and will always suggest only a few market leaders.
Searching for information on the internet became a jungle a decade ago, and to be visible you have to pay Google for the sunlight. Now, we risk falling into real darkness — until some paid model emerges.
Taking into account the sometimes blind belief that 'LLMs know everything', the outcome could be very costly, especially for technologies and businesses unfortunate enough to emerge after 2025.
still the cutoff is very much concerning and inconvenient
And I guess Gemini 3.5 pro will have the pricing increment, too. 12 x 5 = 60?
It seems like google does want us to use Chinese models.
6x the price of 3.1 flash lite
Cost per task is a more productive measure, but obviously a more difficult one to benchmark.
Compare to the GPT-5.5 announcement: https://openai.com/index/introducing-gpt-5-5/
$0.15 / million tokens
$1.00 / 1,000,000 tokens per hour (storage price)
I confirmed this by running a bunch of prompts through Gemini 3.5 Flash without doing anything special to configure caching and noting that it comes back with a "cachedContentTokenCount" on many of the responses.
The "storage price" quoted is for an optional Gemini feature that most people don't care about: https://ai.google.dev/gemini-api/docs/caching#explicit-cachi...
score age size name
44.2 97 large GLM-5 (Reasoning)
44.7 187 - GPT-5.1 (high)
44.9 29 - Qwen3.6 Max Preview
45 0 - Gemini 3.5 Flash
45.5 27 large MiMo-V2.5-Pro
45.6 75 - GPT-5.4 (low)
which you can invoke with
$ curl day50.dev/art-analysis.sh | bash
inspect the code. it's tiny.
I use it all the time and maintain it. Snag a copy and pull it down again if it breaks on you. I stay on top of it.
They continue to focus on smaller models while openai and anthropic are increasing compute requirements for their SOTA models.
Can you link to a source?
They are just refining their current models while they finish training the next generation.
They will all come out at about the same time. Anthropic, OpenAi, Google, xAI
Hold on, I think this claim needs some hard data. Here you go gentlemen:
https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5...
There might be a harness difference, but also, this CTF-type benchmark might not capture the capability difference fully.
For intelligence/size only OpenAI and Anthropic are the frontier. Google has more compute so it can compensate for that with size of the models...
Nobody really knows the answer to which one is more optimal
* Large model trained on a large amount of data across multiple domains, that doesn't need any extra content to answer questions.
* Smaller model that is smart enough to go fetch extra relevant content, and then operate on essentially "reformatting" the context into an answer.
Gemini Pro 3.1 for agentic coding is still clumsy. It chews a lot, has a harder time with tools and interacting with the codebase. I haven't tried any 3.5 version, yet, though. The benchmarks look promising.
I'll note I like the Google models' prose better than any others at the moment, though. Even the small open models (Gemma 4 family) have excellent prose that doesn't stink of the LLMisms that I find so annoying about OpenAI (especially) and Anthropic models. So, I'll probably start using Gemini for writing API docs, even if all code is Claude.
> Gemini 3.5 Flash’s pricing shifts the Pareto frontier in Text. 8 models from GoogleDeepMind dominate the Text Arena Pareto curve where only 4 labs are represented for top performance in their price tiers.
Artificial Analysis's "Cost to run" model (aka num_tokens_used * price_per_token) is much better, but even that is likely problematic since it's not clear whether running a bunch of benchmarks maps cleanly to real-world token use.
It’s not possible to uptrain on preview releases and it did not get that much love for a while.
https://storage.googleapis.com/gweb-uniblog-publish-prod/ori...
More often than not, people are using images in responses that go awry. Which is fair, the models are sold as multi-modal, but image analyses is still at gpt-4.0 text-analyses levels.
Also knowledge cutoff issues, where people forget the models exist months to a year or more in the past.
Two of the three strip titles are hallucinated and two of the three strips are bad examples. Haley is mute in strip 403 and does nothing. Strip 578 is the start of the arc that shows the behavior Gemini is talking about, but has things going wrong so it's not a good example either.
Claude picks a good strip but also hallucinates the strip title: https://claude.ai/share/56be379d-c3da-443e-b60f-2d33c374eba8
I was trying to understand a game I've been playing, The Last Spell. I asked it for a tier list of omens -- which ones the community considers most important. At least a few of the names it posts are hallucinated ("omen of the sun" does not exist, and the omens that give extra gold are "savings," "fortune," and "great wealth").
Obviously not a critical use case but issues like this do keep me on my toes regarding whether the thing is working at all. I should ask 3.5 flash to do the same job. (I did try and it once again hallucinated the omen names and some of the effects.)
Claude also believes it knows how AWS' KMS works, quite confidently, while getting things wrong. I have a separate "this is how KMS replication actually works" file just to deal with its misconceptions.
For gemini, I typically use it to query information from large corpuses, but it often web searches and hallucinates instead of reading the actual corpus. On a book series, it will hallucinate chapters and events which, while reasonable and plausible, do not exist. "Go look at the files and see if your reference is correct" shows that it's not correct, and it's a mandatory step. But that doesn't prevent hallucination, but makes sure you catch it after the fact, just like a method in a class that doesn't exist gets found out by the compiler. The LLM still hallucinated it.
Also, prompts that reliably produce hallucinations is kind of a hard ask. It's inconsistent. One day the LLM I work with quotes verbatim from the PCIe spec and it's super helpful. The next day it gives me wrong information and when I ask it what section of the spec that information comes from it just makes up a section number
The fix is easy enough though, a line in my global AGENTS.md instructing agents to search/ask for documentation before working on API integrations.
```
Build a Nango sync that stores Figma projects.
Integration ID: figma
Connection ID for dry run: my-figma-connection
Frequency: every hour
Metadata: team_id
Records: Project with id, name, last_modified
API reference: https://www.figma.com/developers/api#projects-endpoints
```
Note: You do need a Nango account and the Nango Skill installed before it could work.
...my chats are all pretty long and involve personal conversations, or I've deleted them. It's a lot to ask for someone to post receipts. The number of complaints is enough data.
No matter how big the model is there will be edge cases where it has no data or is out of date. In these cases it just makes stuff up. You can detect it yourself by looking for words like usually or often when it states facts, e.g. "the mall often has a Starbucks." I asked it about a Genshin Impact character released in June 2025 and it consistently interpreted the name (Aino) as my player character because Aino wasn't in its data.
Honestly I'm surprised your haven't encountered it if you're using it more than casually. Pro is much better but not perfect.
And when I say all the time, I mean it, and this is for Opus 4.7 Adaptive.
I often have to say, please do searches and cite sources, as if it doesn't it will confidently give me wrong or outdated information.
If you're often asking questions about a topic that's not in your specialist knowledge you won't notice them.
If you aren't paying attention it can spend a long time (and a lot of tokens) spinning in that loop. Sometimes there might be more than two approaches in the loop, which makes it even harder to see that it's repeating itself in a loop. It's pretty frustrating to see it working away productively (so you think) for 20 minutes or so only to finally notice what's going on
Coding, however, is solved like magic. Easier to add tests, to be fair.
AI psychosis would be the problem people talk about more, not just outright agreement but subtle ways of making you feel confident in your ideas. "yes, buy that domain name buy these other ones for defensibility"
(the domain name is dumb and completely unmarketable)
I did not expect such a huge (3x) price increase from 3.0 Flash and I bet many people will not just blindly upgrade as the value proposition is widely different.
One interesting point to note is that Google marked the model as Stable in contrast to nearly everything else being perpetually set as Preview.
[0] https://artificialanalysis.ai/models/gemini-3-5-flash [1] https://artificialanalysis.ai/models/gemini-3-1-pro-preview
3.1 has 57M output tokens from Intelligence Index, 3.5 Flash has 73M, so not a lot more, and 3.5 is a bit cheaper, I don't get how 3.5 can be 74% more expensive.
That's everything I needed to know.
That said, haste makes waste as the price point completely invalidates that
Reiner Pope gave a talk on Dwarkesh Patel about token economics. I guess faster is a lot more expensive, generally.
Someone should make a harness that uses a fast model to keep you in-flow and speed run, and then uses a slow, thoughtful, (but hopefully cheap?) model to async check the work of the faster model. Maybe even talk directly to the faster model?
Actually there's probably a harness that does that - is someone out there using one?
On my tasks it has not been as good as even Sonnet 4.6 so far.
Instruction following over long context feels worse.
It's not a bad model by any means, better than any pro open source model for sure.
"Yes, your idea is excellent."
"How this works beautifully:"
"This is a fantastic development!"
"This is an exceptionally clean and robust architecture."
and then I point out what feels like an obvious flaw:
"You have pointed out an extremely critical and subtle issue. You are absolutely 100% correct."
I'm sad that I'll probably stop using 3.5 Flash because I just hate its personality.
If not then I’m not using it.
Cancelled my account 3 months ago, only Claude code level capability would bring me back.
For reference, this is a Rust codebase, deep "systems" stuff (database, compiler, virtual machine / language runtime)
They're still months behind OpenAI and Anthropic on coding.
Mind you I also find Claude Code careless and unreliable these days, too. (But it's good at tool use at least).
I do use Gemini for "lifestyle" AI usage (web research etc) tho.
Also concerned about Gemini models being benchmaxxed generally
I would say they are the least benchmaxxed out of all the top labs, for coding. They've always been behind opus/gpt-xhigh for agentic stuff (mostly because of poor tool use), but in raw coding tasks and ability to take a paper/blog/idea and implement it, they've been punching above their benchmarks ever since 2.5. I would still take 2.5 over all the "chinese model beats opus" if I could run that locally, tbh.
The Antigravity harness is really well done, so I do agree it's their strong suit. Can't say the same about gemini-cli (though it has a really nice interface)
Would still choose Deepseek for the price
Feels like the AI pricing noose is tightening sooner rather than later.
This model isnt an advancement, its a previous model that runs more compute, which is why it costs more
GDM is making (or has been backed into a corner into making) the bet that high throughput, low latency, low capability models are the path forward.
That probably works for vibe coded apps by non-practitioners.
I suspect that practitioners/professionals will wait longer for better results.
And Google is trying to make something affordable enough for a mass market, ad-supported audience.
They aren’t hyper focused on enterprise like Anthropic is. And that’s okay. There’s room for different players in different markets.
Plus the vibe of the gemini models are so weird particularly when it comes to tool calling
At this point I kinda need them to shock me to make the switch
Gemini 3.5 Flash: $0.75 input / $4.50 output per 1M tokens, 1M context window. The output price explicitly "includes thinking tokens" — which is why it's higher than a typical flash-class model.
For comparison within the Gemini lineup: - Gemini 2.5 Flash: $0.30 / $2.50 - Gemini 3.1 Flash-Lite: $0.25 / $1.50 - Gemini 3.1 Pro Preview: $2.00 / $12.00
So 3.5 Flash is ~2.5x more expensive input vs 2.5 Flash. The pricing and "including thinking tokens" framing position it as a reasoning-capable flash model rather than just a pure speed optimization.
If this is the big model release out of google, its a disappointent.
(I suspect you're viewing the "flex" pricing).
> The pricing and "including thinking tokens" framing position it as a reasoning-capable flash model rather than just a pure speed optimization
Every Gemini model starting with 2.5 has been a reasoning model.
Not a great bicycle though, it forgot the bar between the pedals and the back wheel and weirdly tangled the other bars.
Expensive too - that pelican cost 13 cents: https://www.llm-prices.com/#it=11&ot=14403&sel=gemini-3.5-fl...
https://www.gianlucagimini.it/portfolio-item/velocipedia/
> most ended up drawing something that was pretty far off from a regular men’s bicycle
edit: fixed human hallucination
I ask because:
Insofar as the original pelican test is zero-shot, it effectively serves as a way to test for the presence of a kind of "visual imagination" component within the layers of the model, that the model would internally "paint" an SVG [or PostScript, etc] encoding of an image onto, to then extract effective features from, analyze for fitness as a solution to a stated request, etc.
But if you're trying to do a multi-shot pelican, then just feeding back in the SVG produced in the previous attempt, really doesn't correspond to any interesting human capability. Humans can't take an SVG of a pelican and iteratively improve upon it just based on our imagined version of how that SVG renders, either! Rather, a human, given the pelican, would simply load the pelican SVG in a browser; look at the browser's rendering of the pelican; note the things wrong with that rendering; and then edit the SVG to hopefully fix those flaws (and repeat.)
I imagine current (mult-modal and/or computer-use) LLMs would actually be very good at such an "iterative rendered pelican" test.
And I am saying that if you take one of these SVGs and ask an LLM to look for flaws, it rarely spots those obvious flaws and instead suggests adding a sunset and fish in the birds mouth.
wtf
`<!-- Gold Rim -->`
WTF??