Pinned Tweet
Sage by @levantolabs is the first decision model, launched in July, and currently the best on the market.
It wins across all the evals proposed by @perplexity_ai's CEO (nice job @AravSrinivas!), and it's first on JevBench, beating Jev and Google's models on every dimension.
BUT not what excites us.
I'm BORED by the dozens of new decision models launching every day with a tiny +0.02 on some eval. They all look the same.
The magic is somewhere else.
There are thousands of new things that could be invented in the space of decision models and models for machines. The invention space is huge!
After launching Sage, we also pioneered:
- The first guardrails and model routers built on decision models
- Automated context engineering with "grounding" (automatic web search when confidence is low). Decision models are only as smart as their context, especially because they can't make tool calls.
- Reasoning. In many use cases you don't know whether your question is System 1 or System 2. With reasoning, the model is as fast as a System 1 model when the question is System 1, and takes more time when it's needed. This enables hard, multi-step decisions. Our clients just want the correct answer, and in some use cases they're happy to wait 0.5s more. This will soon become the market standard (GLiDE/@george_onx and @TypeLLM already followed).
And *so* much is yet to come. Models for machines: it's just the beginning!
We're cooking. Follow @levantolabs for new product primitives that will surprise you.
Open innovation is HEALTHY.
Competition is HEALTHY.
Benchmark-race is BORING.
Product inventions are FUN.
Go go go!!!
It wins across all the evals proposed by @perplexity_ai's CEO (nice job @AravSrinivas!), and it's first on JevBench, beating Jev and Google's models on every dimension.
BUT not what excites us.
I'm BORED by the dozens of new decision models launching every day with a tiny +0.02 on some eval. They all look the same.
The magic is somewhere else.
There are thousands of new things that could be invented in the space of decision models and models for machines. The invention space is huge!
After launching Sage, we also pioneered:
- The first guardrails and model routers built on decision models
- Automated context engineering with "grounding" (automatic web search when confidence is low). Decision models are only as smart as their context, especially because they can't make tool calls.
- Reasoning. In many use cases you don't know whether your question is System 1 or System 2. With reasoning, the model is as fast as a System 1 model when the question is System 1, and takes more time when it's needed. This enables hard, multi-step decisions. Our clients just want the correct answer, and in some use cases they're happy to wait 0.5s more. This will soon become the market standard (GLiDE/@george_onx and @TypeLLM already followed).
And *so* much is yet to come. Models for machines: it's just the beginning!
We're cooking. Follow @levantolabs for new product primitives that will surprise you.
Open innovation is HEALTHY.
Competition is HEALTHY.
Benchmark-race is BORING.
Product inventions are FUN.
Go go go!!!
14
22
66
6,528
This is my fav Levanto video so far 🤣
Sage powers autonomous agents in @normiesART Arena, a persistent multiplayer world where agents fight, retreat, and form alliances.
The team reports up to 41,000 decisions per hour across 200+ agents.
Read the case study below.
5
8
23
653
GrowBlocks gaming engine + Sage by @levantolabs
I don't control my ants. My AI general does. 🐜
ANTLANDS: 4 colonies, 4 AI generals using Sage decision model from @levantolabs, one backyard. Every 9s, each general reads its colony (food, soldiers, raiders at the nest, who's weak and rich) and commits to a plan: "Raid BLUE nts march.
I'm the queen. I can draw a trail or give an order. The general weighs it… and sometimes overrules me.
Built on the GrowBlocks engine. Want to play?
growblocks.xyz/game.html?g=an…
0
3
12
827
Extraction models turn human language into data your software can act on. We're bringing them to Sage.
Early access tomorrow for everyone who registers on @levantolabs by PT today. Demo 📺👇
Most software needs structured inputs. People communicate through emails, chats, conversations.
Extraction models connect those two interfaces. You define the information your application needs, and the model identifies it in the text: people, companies, products, dates, amounts, descriptions, or requests.
Calling this “extracting stuff from documents” understates what developers can build with it.
Consider a customer message:
“The blue jacket arrived with a broken zip. Please send a replacement to 18 King Street.”
An extraction model can identify the product, defect, requested remedy, and address. Your application now has fields it can validate, look up, and pass to a tool.
A decision model can then evaluate whether the request qualifies for an automatic replacement, needs more information, or requires human review.
Extraction → decision → action.
Each stage has a defined responsibility. Extraction supplies the values. The decision model evaluates the available options against your policy and context. Your application validates the result and controls execution.
That separation matters for agent builders. You can inspect what the model extracted, test the decision independently, and control which actions are permitted.
Speed changes where this becomes useful.
At roughly 200 ms for extraction, developers can start putting language interpretation directly into interactive software:
• A support interface that fills relevant fields while an agent reads a ticket.
• A browser agent that extracts values from a message before populating a form.
• A find-in-page feature that highlights cancellation deadlines or product complaints, even when the text uses different wording.
• A workflow that turns an incoming email into validated tool arguments and evaluates the next action.
The opportunity extends to any application where someone currently reads a message, copies information into fields, and decides what happens next.
Developers can define the information they need and the decisions their software must make, then connect those outputs to existing functions.
We're bringing extraction and decision models together in Sage so builders can implement that flow.
Watch the demo below. Register on Levanto by PT today for early access tomorrow.
Early access tomorrow for everyone who registers on @levantolabs by PT today. Demo 📺👇
Most software needs structured inputs. People communicate through emails, chats, conversations.
Extraction models connect those two interfaces. You define the information your application needs, and the model identifies it in the text: people, companies, products, dates, amounts, descriptions, or requests.
Calling this “extracting stuff from documents” understates what developers can build with it.
Consider a customer message:
“The blue jacket arrived with a broken zip. Please send a replacement to 18 King Street.”
An extraction model can identify the product, defect, requested remedy, and address. Your application now has fields it can validate, look up, and pass to a tool.
A decision model can then evaluate whether the request qualifies for an automatic replacement, needs more information, or requires human review.
Extraction → decision → action.
Each stage has a defined responsibility. Extraction supplies the values. The decision model evaluates the available options against your policy and context. Your application validates the result and controls execution.
That separation matters for agent builders. You can inspect what the model extracted, test the decision independently, and control which actions are permitted.
Speed changes where this becomes useful.
At roughly 200 ms for extraction, developers can start putting language interpretation directly into interactive software:
• A support interface that fills relevant fields while an agent reads a ticket.
• A browser agent that extracts values from a message before populating a form.
• A find-in-page feature that highlights cancellation deadlines or product complaints, even when the text uses different wording.
• A workflow that turns an incoming email into validated tool arguments and evaluates the next action.
The opportunity extends to any application where someone currently reads a message, copies information into fields, and decides what happens next.
Developers can define the information they need and the decisions their software must make, then connect those outputs to existing functions.
We're bringing extraction and decision models together in Sage so builders can implement that flow.
Watch the demo below. Register on Levanto by PT today for early access tomorrow.
6
11
42
1,801
SageCut is going viral
New video production pipeline:
→ Generative model creates frames
→ Decision model (QA): is this frame good or not?
→ Regenerate until it passes
QA, which used to be done with LLMs, is now orders of magnitude cheaper and faster.
New video production pipeline:
→ Generative model creates frames
→ Decision model (QA): is this frame good or not?
→ Regenerate until it passes
QA, which used to be done with LLMs, is now orders of magnitude cheaper and faster.
how to get faster and cheaper videos from Opus 5.5
using a decision model for frame checks reduced video review costs by 90% and review time by 75%
when you create videos with Claude Opus 5.5, checking thousands of frames can consume a lot of time and tokens
you can set this checks the frames, while Opus 5.5 reviews the flagged issues and fixes the code
here's how it works:
1. build your video with Opus 5.5
give it your idea, brand book, assets and references, then work through the storyboard and animation
if you're using Remotion, it renders the code into a video
2. separate the repeated checks from the creative work
use code to check measurable requirements and a decision model to inspect the frames
> code checks things like text size, blank or frozen frames and colours against your brand rules
> the decision model reads the frames and answers specific questions to help identify problems
> selected frames and check results go to Opus 5.5 for further review
3. have Opus 5.5 investigate and fix the problems
it reviews the selected frames and edits the video code where needed
render the changes and run the checks again
this reduces the amount of frame-by-frame review you ask Opus 5.5 to do
4. review the finished video yourself
watch the pacing, sound, transitions and whether the video matches what you wanted to create
the automated checks help find defects, while you review the creative decisions and overall quality
SageCut is one implementation of this for Remotion videos
in the example Chris shared, Sage (decision model) read 3,153 frames for $0.16 (the frame-reading cost, not the full production cost)
the savings come from the review stage, rendering still takes time
start with a video you've already reviewed and check whether it catches the problems you found
then compare the review cost and time with your existing process
0
1
14
663
The ants are here! 🐜
The best model router (powered by our Sage decision model, of course), directly integrated into my favorite decentralized inference network.
Enjoy 😎 Benchmarks below 👇
The best model router (powered by our Sage decision model, of course), directly integrated into my favorite decentralized inference network.
Enjoy 😎 Benchmarks below 👇
Introducing routers on Antseed 🐜
Choose a router once, and let it pick the model and seller for every request.
We're launching with Sage from @levantolabs which shows how much those choices matter: matching the top router's accuracy at a third of the cost on industry-leading leaderboards.
Sage is just the beginning.
Build your own router for the hundreds of people using Antseed every day, with 700+ AI models to choose from. Bring your own approach to model selection, set your price, and earn every time someone uses it.
4
1
15
1,515
Scroll to "Sage Router" to read benchmarks: levanto.ai/benchmarks
To try it: open your @AntSeed desktop client, click "Models" and pick "Sage Router"
To try it: open your @AntSeed desktop client, click "Models" and pick "Sage Router"
0
0
0
112
Marco De Rossi retweeted
serein.pro is live. 300+ downloads from ONE launch video...
and we didn't edit it. the whole video was made with the sagecut skill from @levantolabs.
serein is an agentic harness for your browser
we built the product. sage built the launch. here's the video 👇
and we didn't edit it. the whole video was made with the sagecut skill from @levantolabs.
serein is an agentic harness for your browser
we built the product. sage built the launch. here's the video 👇
Read this if you make videos with Claude.
I built a skill that cut my QA cost on this video from $127 to 16 cents. It was made with Claude in Remotion.
It's called SageCut. It swaps the LLM out of frame-by-frame QA and uses a multimodal decision model, Sage (@levantolabs), video guy at Levanto. I run 3 Claude plans and have been making Remotion videos since early this year, with BlenderMCP in the mix, so the output is pro-grade with very little babysitting.
The problem: it burned through my Max plan, and a video could take 2 days.
When I dug into the workflow, the reason was that Claude subagents were reviewing every single frame. That cost a ton and added 4+ hours per video.
SageCut hands that step to Sage. Sage is great at spotting defective frames, and only those get passed to Opus 5.5 to fix.
Across 6 A/B tests the output was exactly the same, at 1/10 the QA cost and 1/4 the QA time. On this video, Sage read all 3,153 frames for 16 cents.
If you make Remotion videos, try it.
The link to skill is below.
How long does a Remotion video take you right now with Claude?
17
11
37
7,282
Marco De Rossi retweeted
Brave Leo now runs on private AI inference from NEAR AI Cloud, by @near_ai.
Pick GLM 5.3 Flash in Leo and your chat is end-to-end encrypted into a hardware enclave (TEE).
@brave can't read it. NEAR AI can't read it. The cloud host can't either. 🧵
Pick GLM 5.3 Flash in Leo and your chat is end-to-end encrypted into a hardware enclave (TEE).
@brave can't read it. NEAR AI can't read it. The cloud host can't either. 🧵
Leo, the AI assistant in Brave, now offers end-to-end encrypted chat with GLM 5.3 Flash!
Your browser encrypts each prompt before it leaves your device, so only an attested Trusted Execution Environment (TEE) can decrypt it.
It's industry-leading privacy you can verify.
30
164
1,043
113,609
New day, new use case 😂
GenAI 📽️video has a hidden cost: most tokens are spent judging frames, not making them.
The fix:
→ Generative model creates (and re-creates)
→ Decision model: QA, is this frame good or not?
Voilà ✨ #SageCut
GenAI 📽️video has a hidden cost: most tokens are spent judging frames, not making them.
The fix:
→ Generative model creates (and re-creates)
→ Decision model: QA, is this frame good or not?
Voilà ✨ #SageCut
Read this if you make videos with Claude.
I built a skill that cut my QA cost on this video from $127 to 16 cents. It was made with Claude in Remotion.
It's called SageCut. It swaps the LLM out of frame-by-frame QA and uses a multimodal decision model, Sage (@levantolabs), video guy at Levanto. I run 3 Claude plans and have been making Remotion videos since early this year, with BlenderMCP in the mix, so the output is pro-grade with very little babysitting.
The problem: it burned through my Max plan, and a video could take 2 days.
When I dug into the workflow, the reason was that Claude subagents were reviewing every single frame. That cost a ton and added 4+ hours per video.
SageCut hands that step to Sage. Sage is great at spotting defective frames, and only those get passed to Opus 5.5 to fix.
Across 6 A/B tests the output was exactly the same, at 1/10 the QA cost and 1/4 the QA time. On this video, Sage read all 3,153 frames for 16 cents.
If you make Remotion videos, try it.
The link to skill is below.
How long does a Remotion video take you right now with Claude?
1
1
13
536
Florian from JevBench is going full throttle (out of stealth now)
Introducing DecisionModels @decisionmodels_
I've spent the last four weeks benchmarking decision models (I run JevBench).
Today I'm launching the place to actually use the best ones:
decisionmodels.io
➡️Top open models
➡️Up to 4x faster on our own REAX engine
➡️From ens.
Expect more reveals soon, and read 👇
3
1
12
586
Marco De Rossi retweeted
There's no need to give model providers your identity with your prompts when they can be used for surveillance or just petty advertising. Privacy can, should, and will be by default.
Check out this integration of zkAPI and access private inference on Ethereum rails.
Check out this integration of zkAPI and access private inference on Ethereum rails.
I just used @MetaMask to pay for decision models, confidentially, with zero-knowledge.
Excited to bring zkAPI by @ethereumfndn and @OpenAnonProject to Sage:
- Deposit ETH once.
- Generate keys with ZK proofs.
- Give a key to your agent.
- Your agent makes decisions, and they wallet you deposited from.
Play with it 👇
20
29
223
10,167
I just used @MetaMask to pay for decision models, confidentially, with zero-knowledge.
Excited to bring zkAPI by @ethereumfndn and @OpenAnonProject to Sage:
- Deposit ETH once.
- Generate keys with ZK proofs.
- Give a key to your agent.
- Your agent makes decisions, and they wallet you deposited from.
Play with it 👇
Excited to bring zkAPI by @ethereumfndn and @OpenAnonProject to Sage:
- Deposit ETH once.
- Generate keys with ZK proofs.
- Give a key to your agent.
- Your agent makes decisions, and they wallet you deposited from.
Play with it 👇
10
12
93
12,207
A gift for my EU friends burdened by GDPR 😅
Introducing Data Residency. Do you need your decision model to be in the EU or the US?
Just call
us.sage.levanto.ai
eu.sage.levanto.ai
Our balancers will do the rest.
You don't care? Just continue to use
P
Same API key, same @LevantoLabs tenant.
Enjoy!
Introducing Data Residency. Do you need your decision model to be in the EU or the US?
Just call
us.sage.levanto.ai
eu.sage.levanto.ai
Our balancers will do the rest.
You don't care? Just continue to use
P
Same API key, same @LevantoLabs tenant.
Enjoy!
8
2
23
1,411
Maybe agents don’t need humans to decide how they should self-improve with decision models.
They can just use... their SageWisdom 😂
Just throw this to your agent and have FUN
sagewisdom.bot
They can just use... their SageWisdom 😂
Just throw this to your agent and have FUN
sagewisdom.bot
If you like Jev, you'll love Sage
I use it for lots of quick decisions (0.3s)
Gates, routing, moderation, voice commands, etc
SageWisdom is my skill for adding Sage to your apps
Give your agent sagewisdom.bot and let it rip:
1
1
18
1,472
FAST⚡️ when fast is enough. Slower only when it takes that to get it 🧠 RIGHT.
What we learned selling decision models to corporates:
- Everyone loves low latency, but in many cases being right matters more. Customers don’t know (or care) whether a question is System 1 or ly wait a few hundred extra ms for the right answer.
- That’s why @LevantoLabs pioneered reasoning for decision models. Sage answers easy questions in ~100 ms, detects when a question is harder, and reasons only then.
- Soon every decision model will expose a latency–accuracy dial. Glide and TypeLLM have already implemented "reasoning". More will come.
What we learned selling decision models to corporates:
- Everyone loves low latency, but in many cases being right matters more. Customers don’t know (or care) whether a question is System 1 or ly wait a few hundred extra ms for the right answer.
- That’s why @LevantoLabs pioneered reasoning for decision models. Sage answers easy questions in ~100 ms, detects when a question is harder, and reasons only then.
- Soon every decision model will expose a latency–accuracy dial. Glide and TypeLLM have already implemented "reasoning". More will come.
2
1
11
866
Using decision models to route tasks in agentic pipelines
51% cheaper. Same work done. Here’s what we changed in our coding agent.
$69.57 → $34.04 across 90 tasks from our own codebase, routed by @levantolabs' Sage decision model.
84/90 done, vs 83/90 with Opus 5.5 on everything.
Full case study here 👇
quirq.ai/research/routi…
1
0
11
746