52 comments

[ 4.1 ms ] story [ 72.3 ms ] thread
At the moment, Deep Think is only available with the ULTRA subscription ($250 per month).
Approach is analogous to Grok 4 Heavy: use multiple "reasoning" agents in parallel and then compare answers before coming back with a single response, taking ~30 minutes. Great results, though it would be more fair for the benchmark comparisons to be against Grok 4 Heavy rather than Grok 4 (the fast, single-agent model).
Surprised no one has released an app yet that pits all the major models against each other for a final answer.
139.99€/month for something you can't even test first, lol
Wait...

So if someone cool enough, they could actually give us a DeepThought model?

Please, let that happen.

Vendor-DeepThought-42B maybe?

Grok 4 heavy, o3 pro and Gemini Deep Think all are equivalent. I wonder how they compare?
[flagged]
You can't go anywhere without having Gemini shoved in your face. I had an immediate visceral reaction to this.
> If you’re a Google AI Ultra subscriber, you can use Deep Think in the Gemini app today with a fixed set of prompts a day by toggling “Deep Think” in the prompt bar when selecting 2.5 Pro in the model drop down.

If fixed set means fixed number it would be nice to know how many.

Otherwise i would like to know what fixed set means here.

Available via API?
I find it interesting, how OpenAI came out with a $200 plan, Anthropic did $100 and $200, then Gemini ups it to $250, and now Grok is at $300.

OpenAI is the only one that says "practically unlimited" and I have never hit any limit on my ChatGPT Pro plan. I hit limits on Claude Max (both plans) several times.

Why are these companies not upfront about what the limits are?

Per-usage pricing discourages use which limits how critical a service can be to your life or workflow. These companies want you to rely on the service such that you’ll pay the price. One customer might use it once a day and find the price reasonable; another may use it 10 times a day and still find the price reasonable. This kind of broad pricing allows for this variation.
I can’t even convince Gemini CLI while planning things to not go off and make a bunch of random changes on its own, even after being very clear not to do so, intercepting to tell it to stop doing that, then it just continues on fucking everything up.
I started doing some experimentation with this new Deep Think agent, and after five prompts I reached my daily usage limit. For $250 USD/mo that’s what you’ll be getting folks.

It’s just bizarrely uncompetitive with o3-pro and Grok 4 Heavy. Anecdotally (from my experience) this was the one feature that enthusiasts in the AI community were interested in to justify the exorbitant price of Google’s Ultra subscription. I find it astonishing that the same company providing free usage of their top models to everybody via AI Studio is nickel-and-diming their actual customers like that.

Performance-wise. So far, I couldn’t even tell. I provided it with a challenging organizational problem that my business was facing, with the relevant context, and it proposed a lucid and well-thought-out solution that was consistent with our internal discussions on the matter. But o3 came to an equally effective conclusion for a fraction of the cost, even if it was less “cohesive” of a report. I guess I’ll have to wait until tomorrow to learn more.

The part I cannot understand is why for many AI offerings, I cannot make out what each pricing tier does with a quick glance.

What happened to the simplicity of Steve Jobs' 2x2 (consumer vs.pro, laptop vs. desktop)?

The rate limits are not because of compute performance or the lack of. It's to stop people from training their own models on the very cutting edge.
I have ultra. Will not be renewing it. Useless, at least have global limits and let people decide how they want to use it. If I have tokens left, why can't I use it for code assist?
I'd be interested in tests involving tasks with large amounts of context. Parallel thinking could conceivably useful for a variety of specific problem types. Having more context than any specific chain of thought can reasonably attend to might be one of them.
Uncompetitive how, what task and Eval?

Gemini is consistently the only model that can reason over long context in dynamic domains for me. Deep Think just did that reviewing an insane amount of Claude Code logs - for a meta analysis task of the underlying implementation. Laughable to think Grok could do that.

I'm not in the AI sceptic camp (LLMs can be useful for some tasks, and I use them often), but this is the big issue at the moment.

In order for agentic AI to replace (for example) a software engineer, we need a big step up in capability, around an order of magnitude. These chain of thought models do get a bit closer to that, although in my opinion we're still a way away.

However, at the same time we need about an order of magnitude decrease in price. These models are expensive even at the current price tokens are sold at which seems to be below the actual cost. And these massive CoT models are taking us in completely the wrong direction in terms of cost

I would be interested in reading about how people who are paying for access to Google's top AI plan are intending to use this. Do you have any examples of immediate use-cases that might benefit?

Is Google using this tool internally? One would expect them to give some examples of how it's helping internal teams accelerate or solve more challenging problems, if they were eating their own dogfood.

You can spin up a version of this at home using simonw's LLM cli with the llm-consortium plugin.

Bonus 1: Use any combination of models. Mix n match models from any lab.

Bonus 2: Serve your custom consortium on a local API from a single command using the llm-model-gateway plugin and use it in your apps and coding assistants.

https://x.com/karpathy/status/1870692546969735361

  > uv tool install llm
  llm install llm-consortium
  llm consortium save gthink-n5       -m gemini-pro -n 5 --arbiter gemini-flash --confidence-threshold 99 --max-iterations 4

  llm serve --host 0.0.0.0

  curl http://0.0.0.0:8000/v1/chat/completions \
  -X POST \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gthink-n5",
    "messages": [{"role": "user", "content": "find a polynomial algorithm for graph-isomorphism"}]
  }'

You can also build a consortium of consortiums like so:

  llm consortium save gem-squared -m gthink-n5 -n 2 --arbiter gem-flash
Or even make the arbiter a consortium:

  llm consortium save gem-cubed -m gthink-n5 -n 2 --arbiter gthink-n5 --max-iteration 2
or go openweights only:

  llm consortium save open-council -m qwen3:2 -m kimi-k2:2 -m glm-4.5:2 -m mistral:2 --arbiter minimax-m1 --min-iterations 2 --confidence-threshold 95
https://GitHub.com/irthomasthomas/llm-consortium
Ladies and Gentlemen,

Here's Gemini Deep Think when prompted with:

"Create a svg of a pelican riding on a bicycle"

https://www.svgviewer.dev/s/5R5iTexQ

Beat Simon Willison to it :)

(comment deleted)
I don't have access but I wonder whether a dog on a jetski would be nearly as good
[flagged]
Your comments were being flagged by users, for what should be obvious reasons.
I'm wondering if 'slow AI' like this is a temporary bridge, or a whole new category we need to get used to. Is the future really about having these specialized 'deep thinkers' alongside our fast, everyday models? Or is this just a clunky V1 until the main models get this powerful on their own in seconds?
Been using Gemini for a few months, somehow it's gotten much, much worse in that time. Hallucinations are very common, and it will argue with you when you point it out. So, don't have much confidence.
Is the problem mainly with tool use ? and are you using it through AI studio or through the API ?.

I've found that it hallucinates tool use for tools that aren't available and then gets very confident about the results.

Via the chat prompt mostly, and sometimes via Copilot. It was quoting me sources and links that didn't exist, and when I told it the links were wrong it doubled down forever, no matter how hard I tried to tell it otherwise. Even sent screenshots, etc.

Kinda just got stuck in a self-confident loop that time. Other times the output is just far worse than Claude for similar use cases, where a couple months back it was stronger, at least in my subjective experience.

Same here. I stopped using Gemini Pro because on top of it's hard to follow verbosity it was giving contradicting answers. Things that Claude Sonnet 4 could answer.

Speaking of Sonnet, I feel like it's closing the gap to Opus. After the new quotas I started to try it before Opus and now it gets complex things right more often than not. This wasn't my experience just a couple of months ago.

my recent experience with flash and using it to prototype a c++ header i was developing:

- it was great to brainstorm with but it routinely introduced edits and dramatic code changes, often unnecessary and many times causing regressions to existing, tested code. - numerous times recursion got introduced to revisions without being prompted or without any justified or good reason - hallucinated a few times regarding c++ type deduction semantics

i eventually had to explicitly tell it to not introduce edits in any working code being iterated on without first discussing the changes, and then being prompted by me to introduce the edits.

all in all i found base chatgpt a lot more productive and accurate and ergonomic for iterating (on the same problem just working it in parallel with gemini).

- code changes were not always arbitrarily introduced or dramatic - it attempted to always work with the given code rather than extrapolate and mind read - hallucinated on some things but quickly corrected and moved forward - was a lot more interactive and documenting - almost always prompted me first before introducing a change (after providing annotated snippets and documentation as the basis for a proposed change or fix)

however, both were great tools to work with when it came to cleaning up or debugging existing code, especially unit testing or anything related to TDD

Upgraded and quickly hit my limit. And find that they have limits, I just wish that they were more transparent. Even if it's just a vague statement about limited usage. I assumed it would be similar to regular Gemini 2.5 on the pro plan but it's not
Interesting they compared Gemini 2.5 Deep Think on code to all the top models EXCEPT Claude, the best top model at code.
This comes at a time where my experience with Gemini is lacking, it seems to get worse. It's not picking up on my intention, sometimes replies in the wrong language, etc. Either that or I am just transparent that it's a tool and its feelings are hurt. I've had to call it a moron several times, and it was funny when it started reprimanding me for my foul language once. But it was wrong. This behavior seems new. I could never trust it to not do random edits everywhere in a document, so nowadays I use it to check Claude, which can be trusted with a document.