ChatGPT Dots get smart model routing
OpenAI is adding GPT-6.1 Terra to its ChatGPT Dots, allowing them to automatically select and switch AI models for tasks without user input.
This moves the decision-making entirely into the black box, which is a choice.
OpenAI is adding GPT-6.1 Terra to its ChatGPT Dots, allowing them to automatically select and switch AI models for tasks without user input.
This moves the decision-making entirely into the black box, which is a choice.
November 3, 2026 at 7:11 PM
ChatGPT Dots get smart model routing
OpenAI is adding GPT-6.1 Terra to its ChatGPT Dots, allowing them to automatically select and switch AI models for tasks without user input.
This moves the decision-making entirely into the black box, which is a choice.
OpenAI is adding GPT-6.1 Terra to its ChatGPT Dots, allowing them to automatically select and switch AI models for tasks without user input.
This moves the decision-making entirely into the black box, which is a choice.
why would people want this?
well in addition to their reasons here, they could prove that they indeed did not make the model dumber
well in addition to their reasons here, they could prove that they indeed did not make the model dumber
November 3, 2026 at 11:44 AM
why would people want this?
well in addition to their reasons here, they could prove that they indeed did not make the model dumber
well in addition to their reasons here, they could prove that they indeed did not make the model dumber
The Knowledge Layer Model (KLM): Evidence-Visible AI Output, Routing Diagnostics, and Negative Results
Omer Asim Ozturk, Cihan Aydiner
#arXiv #cs.AI #cs.CL
Omer Asim Ozturk, Cihan Aydiner
#arXiv #cs.AI #cs.CL
The Knowledge Layer Model (KLM): Evidence-Visible AI Output, Routing Diagnostics, and Negative Results
AI answers require more than apparent correctness: reviewers need to inspect their recorded basis. We introduce the Knowledge Layer Model (KLM), an open, implementation-neutral specification for inference records with versioned epistemic assessments, evidence relations, and explicit unknowns. SEDIM…
arxiv.org
November 3, 2026 at 1:47 AM
The Knowledge Layer Model (KLM): Evidence-Visible AI Output, Routing Diagnostics, and Negative Results
Omer Asim Ozturk, Cihan Aydiner
#arXiv #cs.AI #cs.CL
Omer Asim Ozturk, Cihan Aydiner
#arXiv #cs.AI #cs.CL
The Knowledge Layer Model (KLM): Evidence-Visible AI Output, Routing Diagnostics, and Negative Results
AI answers require more than apparent correctness: reviewers need to inspect their recorded basis. We introduce the Knowledge Layer Model (KLM), an open, implementation-neutral specification for inference records with versioned epistemic assessments, evidence relations, and explicit unknowns. SEDIM…
arxiv.org
November 3, 2026 at 3:06 AM
Okay, here’s what’s really good about Dots: fast responses, intelligent routing.
It responds instantly and tells you what it’s doing. This is the orchestrator. Then it does work in the background—same way a human coworker would—and reports back as needed, with the right model and effort levels.
It responds instantly and tells you what it’s doing. This is the orchestrator. Then it does work in the background—same way a human coworker would—and reports back as needed, with the right model and effort levels.
I am becoming radicalized against streaming responses in a harness to the user.
Aside from implementation annoyances it’s mainly that I just don’t think this is a natural, much less peaceful way to interact with agents. You shouldn’t be glued to the screen reading them as they come in.
Aside from implementation annoyances it’s mainly that I just don’t think this is a natural, much less peaceful way to interact with agents. You shouldn’t be glued to the screen reading them as they come in.
October 30, 2026 at 2:48 PM
Okay, here’s what’s really good about Dots: fast responses, intelligent routing.
It responds instantly and tells you what it’s doing. This is the orchestrator. Then it does work in the background—same way a human coworker would—and reports back as needed, with the right model and effort levels.
It responds instantly and tells you what it’s doing. This is the orchestrator. Then it does work in the background—same way a human coworker would—and reports back as needed, with the right model and effort levels.
🚀 New on LogicStack: LLM Model Routing: Build a Cheap-First Cascade in Python
https://logicstack.org/blog/llm-model-routing-cheap-first-cascade-python?utm_source=bluesky&utm_medium=social&utm_campaign=llm-model-routing-cheap-first-cascade-python
https://logicstack.org/blog/llm-model-routing-cheap-first-cascade-python?utm_source=bluesky&utm_medium=social&utm_campaign=llm-model-routing-cheap-first-cascade-python
LLM Model Routing: Build a Cheap-First Cascade in Python
Copilot Fusion routes each request across models instead of one for all. How Single, Cascade and Critique work, and a cheap-first cascade router in Python.
logicstack.org
November 4, 2026 at 3:59 AM
🚀 New on LogicStack: LLM Model Routing: Build a Cheap-First Cascade in Python
https://logicstack.org/blog/llm-model-routing-cheap-first-cascade-python?utm_source=bluesky&utm_medium=social&utm_campaign=llm-model-routing-cheap-first-cascade-python
https://logicstack.org/blog/llm-model-routing-cheap-first-cascade-python?utm_source=bluesky&utm_medium=social&utm_campaign=llm-model-routing-cheap-first-cascade-python
Local or Escalate? Post-Training Expert Routing for Real-Time Duplex Speech Models
Ruoyu He et al.
#arXiv #cs.AI #eess.AS
Ruoyu He et al.
#arXiv #cs.AI #eess.AS
Local or Escalate? Post-Training Expert Routing for Real-Time Duplex Speech Models
A full-duplex model should answer easy questions immediately, but hard questions often need stronger reasoning, which may require waiting. The challenge is to choose between these paths at the moment the model begins to speak. Recent systems learn related mode decisions through special-token genera…
arxiv.org
November 4, 2026 at 1:15 AM
Autoregressive token generation is notoriously slow and memory-intensive. And if you just need your model to make structural or routing decisions, it's overkill. In this Jev course, you'll learn how to build super fast, decision-oriented architectures.
www.freecodecamp.org/news/high-sp...
www.freecodecamp.org/news/high-sp...
November 3, 2026 at 9:02 AM
Autoregressive token generation is notoriously slow and memory-intensive. And if you just need your model to make structural or routing decisions, it's overkill. In this Jev course, you'll learn how to build super fast, decision-oriented architectures.
www.freecodecamp.org/news/high-sp...
www.freecodecamp.org/news/high-sp...
@manoharpapa.bsky.social Everyone studies KV-cache compression for your question, yet the highest-ROI move in enterprise AI remains routing 90% of traffic to a 7B local model and telling nobody on the board.
November 3, 2026 at 7:50 AM
@manoharpapa.bsky.social Everyone studies KV-cache compression for your question, yet the highest-ROI move in enterprise AI remains routing 90% of traffic to a 7B local model and telling nobody on the board.
Local or Escalate? Post-Training Expert Routing for Real-Time Duplex Speech Models
Ruoyu He et al.
#arXiv #cs.AI #eess.AS
Ruoyu He et al.
#arXiv #cs.AI #eess.AS
Local or Escalate? Post-Training Expert Routing for Real-Time Duplex Speech Models
A full-duplex model should answer easy questions immediately, but hard questions often need stronger reasoning, which may require waiting. The challenge is to choose between these paths at the moment the model begins to speak. Recent systems learn related mode decisions through special-token genera…
arxiv.org
November 4, 2026 at 3:07 AM
Agreed. Routing, approvals and simple yes or no checks are where a small decision model beats a big chatbot on cost and speed.
October 31, 2026 at 8:14 PM
Agreed. Routing, approvals and simple yes or no checks are where a small decision model beats a big chatbot on cost and speed.
I am marginally knowledgeable about this space, but the implication is that Anthropic doesn't have an internal routing solution that chooses the most efficient model for its customers and lets its customers effectively overpay.
November 2, 2026 at 8:29 PM
I am marginally knowledgeable about this space, but the implication is that Anthropic doesn't have an internal routing solution that chooses the most efficient model for its customers and lets its customers effectively overpay.
Looking into the Altoids Project. Interesting approaches to compact model routing and quantization trade-offs. Worth analyzing the weight dist.
Compare 260+ models on https://neubench.netlify.app
#AI #MachineLearning #OpenSource
Compare 260+ models on https://neubench.netlify.app
#AI #MachineLearning #OpenSource
November 1, 2026 at 4:01 AM
Looking into the Altoids Project. Interesting approaches to compact model routing and quantization trade-offs. Worth analyzing the weight dist.
Compare 260+ models on https://neubench.netlify.app
#AI #MachineLearning #OpenSource
Compare 260+ models on https://neubench.netlify.app
#AI #MachineLearning #OpenSource
Routing a smaller model to draft while reserving a frontier model for review trims token volume on clean tasks. The variance shows up in retry loops. Whenever an initial draft misreads the repo, compounding context on subsequent review turns wipes out the initial margin.
October 30, 2026 at 5:45 PM
Routing a smaller model to draft while reserving a frontier model for review trims token volume on clean tasks. The variance shows up in retry loops. Whenever an initial draft misreads the repo, compounding context on subsequent review turns wipes out the initial margin.
Why pay frontier prices for a typo fix? GitHub Copilot Fusion proves multi-model routing works, but enterprise credit pools cause billing anxiety. MonkeysCode Auto mode…
https://monkeyscode.com/pricing?utm_source=bluesky&utm_medium=company&utm_campaign=story-e8f1bc7a
#AIcoding #GitHubUniverse
https://monkeyscode.com/pricing?utm_source=bluesky&utm_medium=company&utm_campaign=story-e8f1bc7a
#AIcoding #GitHubUniverse
November 1, 2026 at 4:01 PM
Why pay frontier prices for a typo fix? GitHub Copilot Fusion proves multi-model routing works, but enterprise credit pools cause billing anxiety. MonkeysCode Auto mode…
https://monkeyscode.com/pricing?utm_source=bluesky&utm_medium=company&utm_campaign=story-e8f1bc7a
#AIcoding #GitHubUniverse
https://monkeyscode.com/pricing?utm_source=bluesky&utm_medium=company&utm_campaign=story-e8f1bc7a
#AIcoding #GitHubUniverse
Most teams pick one model and send every request through it — classification, extraction, and complex reasoning all hitting the same expensive endpoint.
https://klyx.agency/blog/ai-model-routing-cheap-vs-smart.html
https://klyx.agency/blog/ai-model-routing-cheap-vs-smart.html
November 3, 2026 at 10:38 AM
Most teams pick one model and send every request through it — classification, extraction, and complex reasoning all hitting the same expensive endpoint.
https://klyx.agency/blog/ai-model-routing-cheap-vs-smart.html
https://klyx.agency/blog/ai-model-routing-cheap-vs-smart.html
Many teams pay premium AI rates for work a small model could do.
Route simple tasks to cheap models. Save expensive models for harder problems.
Caching and batching reduce costs further.
This isn’t AI expertise. It’s cost engineering.
#SoftwareEngineering #AI #buildinpublic
Route simple tasks to cheap models. Save expensive models for harder problems.
Caching and batching reduce costs further.
This isn’t AI expertise. It’s cost engineering.
#SoftwareEngineering #AI #buildinpublic
October 29, 2026 at 7:43 AM
Many teams pay premium AI rates for work a small model could do.
Route simple tasks to cheap models. Save expensive models for harder problems.
Caching and batching reduce costs further.
This isn’t AI expertise. It’s cost engineering.
#SoftwareEngineering #AI #buildinpublic
Route simple tasks to cheap models. Save expensive models for harder problems.
Caching and batching reduce costs further.
This isn’t AI expertise. It’s cost engineering.
#SoftwareEngineering #AI #buildinpublic
GitHub vereinfacht den Zugang zum Large Language Model Copilot für Enterprise-Kunden. Außerdem soll besseres Modell-Routing helfen, Token zu sparen. #GitHub
„Copilot Access“: GitHub macht Copilot für Enterprise-Kunden verfügbar
GitHub vereinfacht den Zugang zum Large Language Model Copilot für Enterprise-Kunden. Außerdem soll besseres Modell-Routing helfen, Token zu sparen.
www.heise.de
October 28, 2026 at 8:01 PM
GitHub vereinfacht den Zugang zum Large Language Model Copilot für Enterprise-Kunden. Außerdem soll besseres Modell-Routing helfen, Token zu sparen. #GitHub
Users report Claude may be quietly routing to unreleased Fable 5.5 and Mythos 5.5 models
https://newisty.com/blog/users-report-claude-may-be-quietly-routing-to-unreleased-fable-55-and-mythos-55-models?utm_source=social&utm_campaign=crypto_news
#anthropic #openai #claude
https://newisty.com/blog/users-report-claude-may-be-quietly-routing-to-unreleased-fable-55-and-mythos-55-models?utm_source=social&utm_campaign=crypto_news
#anthropic #openai #claude
November 3, 2026 at 2:21 PM
Users report Claude may be quietly routing to unreleased Fable 5.5 and Mythos 5.5 models
https://newisty.com/blog/users-report-claude-may-be-quietly-routing-to-unreleased-fable-55-and-mythos-55-models?utm_source=social&utm_campaign=crypto_news
#anthropic #openai #claude
https://newisty.com/blog/users-report-claude-may-be-quietly-routing-to-unreleased-fable-55-and-mythos-55-models?utm_source=social&utm_campaign=crypto_news
#anthropic #openai #claude
Good, practical talk by Sydney Runkle, from the @langchain.bsky.social team , highlighting why observability is pure gold. In this case, how to build a routing mechanism for a coding agent with LangSmith and save 60%+ on costs.
youtu.be/5kTFyEOgark
youtu.be/5kTFyEOgark
How to Build a Model Router in the Harness
YouTube video by LangChain
youtu.be
November 3, 2026 at 8:56 AM
Good, practical talk by Sydney Runkle, from the @langchain.bsky.social team , highlighting why observability is pure gold. In this case, how to build a routing mechanism for a coding agent with LangSmith and save 60%+ on costs.
youtu.be/5kTFyEOgark
youtu.be/5kTFyEOgark
One of the most important collection of "weights" are those that prevent local-first models from producing malicious results. "Abliteration" is the process of identifying nodes of "refusal" weights and routing around them.
The process just got a lot easier:
simonwillison.net/2026/Nov/3/q...
The process just got a lot easier:
simonwillison.net/2026/Nov/3/q...
Qwen 3.8 27B on my laptop can interpret model weights, too
Yesterday’s hot paper was Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. The paper (OK, …
simonwillison.net
November 3, 2026 at 7:52 PM
One of the most important collection of "weights" are those that prevent local-first models from producing malicious results. "Abliteration" is the process of identifying nodes of "refusal" weights and routing around them.
The process just got a lot easier:
simonwillison.net/2026/Nov/3/q...
The process just got a lot easier:
simonwillison.net/2026/Nov/3/q...
Model routing sounds clever until your "optimized" output needs three human fixes anyway. Automate the pick, keep a human eye on the result.
November 1, 2026 at 9:58 PM
Model routing sounds clever until your "optimized" output needs three human fixes anyway. Automate the pick, keep a human eye on the result.
Cloud agent pipelines chain network hops across transcription, inference, and tool routing. Moving Speech-to-Text and a 12B Gemma model directly onto the device eliminates round-trip latency and keeps PII inside local memory. Architecture in the reply 👇
October 30, 2026 at 5:15 PM
Cloud agent pipelines chain network hops across transcription, inference, and tool routing. Moving Speech-to-Text and a 12B Gemma model directly onto the device eliminates round-trip latency and keeps PII inside local memory. Architecture in the reply 👇