Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

We turned model latency into a fighting game. Eight models, single elimination, every frame driven by measured response times. Celeris-1-Magnus beat GPT-5.6 Sol, Terra and Luna, plus three Gemini Flash models, to take the cup. The only model that pushed it to the final was Celeris-1, at 188ms.

Celeris

2,163 subscribers

10,372 views • 1 month ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

2 Comments

A!1 month ago

now add a sanity check dumb policy that just randomly select a action as fast as possible maybe it only rewards raw speed rather than intelligence

david caminos15 days ago

Why does Jev still win on knowledge and reasoning? is that a model size thing?

Related Videos

OpenAI vs. Anthropic, a mysterious 300B model, and the definitive AI model tier list (drinking game). What could go wrong? 0:00 Intro 3:30 OpenAI vs. Anthropic 14:44 Anthropic’s delayed models 20:53 Kimi K3 and open weights 30:08 The Alpha stealth model 48:26 Model tier list begins 1:09:53 GPT-5.6 Luna 1:13:09 Opus and Sonnet 1:20:25 Gemini models 1:31:15 DeepSeek and local models 1:59:58 Fable vs. Sol 2:28:46 Final rankings

OpenAI vs. Anthropic, a mysterious 300B model, and the definitive AI model tier list (drinking game). What could go wrong? 0:00 Intro 3:30 OpenAI vs. Anthropic 14:44 Anthropic’s delayed models 20:53 Kimi K3 and open weights 30:08 The Alpha stealth model 48:26 Model tier list begins 1:09:53 GPT-5.6 Luna 1:13:09 Opus and Sonnet 1:20:25 Gemini models 1:31:15 DeepSeek and local models 1:59:58 Fable vs. Sol 2:28:46 Final rankings

Nerd Snipe

117,810 views • 1 month ago

Sol, Terra, and Luna, our GPT‑5.6 family of models, are starting to roll out now in ChatGPT, Codex, and the API.

Sol, Terra, and Luna, our GPT‑5.6 family of models, are starting to roll out now in ChatGPT, Codex, and the API.

OpenAI

4,546,984 views • 3 months ago

We ran a test between GPT-5.6 Sol, Kimi K3, and Gemini 3.6 Flash. One prompt. Across all three models. Using our /design command. Results (gameplay, UX, and UI): 🔹 Kimi K3 performed well across all three. 🔹 GPT-5.6 Sol is good, but it gets annoying when the ball hits the wall. 🔹 Gemini 3.6 Flash is bad (took 5–6 attempts), but it has features the others don’t, like dropping coins for power-up balls and increasing the bar length. Ranking (DX, features, and cost): 🔹 Kimi K3: 10/10 · $0.10 🔹 GPT-5.6 Sol: 9/10 · $0.52 🔹 Gemini 3.6 Flash: 6/10 · $0.45

We ran a test between GPT-5.6 Sol, Kimi K3, and Gemini 3.6 Flash. One prompt. Across all three models. Using our /design command. Results (gameplay, UX, and UI): 🔹 Kimi K3 performed well across all three. 🔹 GPT-5.6 Sol is good, but it gets annoying when the ball hits the wall. 🔹 Gemini 3.6 Flash is bad (took 5–6 attempts), but it has features the others don’t, like dropping coins for power-up balls and increasing the bar length. Ranking (DX, features, and cost): 🔹 Kimi K3: 10/10 · $0.10 🔹 GPT-5.6 Sol: 9/10 · $0.52 🔹 Gemini 3.6 Flash: 6/10 · $0.45

Command Code

61,412 views • 3 months ago

Codex users: here's an amazing way to take advantage of the efficiency and capability of GPT-5.6 Luna Max. It's called 'sol-advisor'. 1. GPT-5.6 Sol High as orchestrator 2. GPT-5.6 Luna Max as implementer for routine tasks 3. GPT-5.6 Terra Max as implementer for complex tasks 4. Fresh GPT-5.6 Sol instance as reviewer A free and open-source plugin for Codex. Installation instructions below 👇 I dare you to install it, and tell me if you hit your weekly usage limits using 'sol-advisor'. I doubt you can.

Codex users: here's an amazing way to take advantage of the efficiency and capability of GPT-5.6 Luna Max. It's called 'sol-advisor'. 1. GPT-5.6 Sol High as orchestrator 2. GPT-5.6 Luna Max as implementer for routine tasks 3. GPT-5.6 Terra Max as implementer for complex tasks 4. Fresh GPT-5.6 Sol instance as reviewer A free and open-source plugin for Codex. Installation instructions below 👇 I dare you to install it, and tell me if you hit your weekly usage limits using 'sol-advisor'. I doubt you can.

Dan McAteer

698,045 views • 2 months ago

Can you take the cheapest open model from a frontier lab & beat closed models on the hardest long range tasks? Yep Our open adaptive intelligence system Zenith ramps DeepSeek v4.1 Flash past GPT 5.6 Sol to the frontier 🚀 h/t Opus 5.5 for making the video in like 2 mins 😮‍💨

Can you take the cheapest open model from a frontier lab & beat closed models on the hardest long range tasks? Yep Our open adaptive intelligence system Zenith ramps DeepSeek v4.1 Flash past GPT 5.6 Sol to the frontier 🚀 h/t Opus 5.5 for making the video in like 2 mins 😮‍💨

Emad

32,447 views • 1 month ago

OpenAI just killed the idea of one best model. GPT-5.6 is three different jobs: - Sol for hard reasoning and long agent runs - Terra for everyday work - Luna for cheap high-volume tasks Most people will send every prompt to Sol, then complain about the cost. The edge is not having the flagship. It is knowing when you do not need it.

OpenAI just killed the idea of one best model. GPT-5.6 is three different jobs: - Sol for hard reasoning and long agent runs - Terra for everyday work - Luna for cheap high-volume tasks Most people will send every prompt to Sol, then complain about the cost. The edge is not having the flagship. It is knowing when you do not need it.

Fluixo

12,641 views • 3 months ago

GPT-5.6 is rolling out in the API. Sol is our flagship model, leading in coding, knowledge work, cybersecurity, and science. Terra delivers performance competitive with GPT-5.5 at lower cost. Luna is our fastest, most affordable model for high-volume tasks.

GPT-5.6 is rolling out in the API. Sol is our flagship model, leading in coding, knowledge work, cybersecurity, and science. Terra delivers performance competitive with GPT-5.5 at lower cost. Luna is our fastest, most affordable model for high-volume tasks.

OpenAI Developers

234,469 views • 3 months ago

Previewing Ultrafast mode for OpenAI's GPT 5.6 Sol, powered by Cerebras. GPT-5.6 Sol Ultrafast generates responses at up to 750 tokens per second. That's the full, GPT-5.6-Sol model -- up to 14× faster than the same model on Standard processing. It speedran Humanity’s Last Exam in 11h 11m, nearly 7× faster than Claude Fable 5 with comparable accuracy.

Previewing Ultrafast mode for OpenAI's GPT 5.6 Sol, powered by Cerebras. GPT-5.6 Sol Ultrafast generates responses at up to 750 tokens per second. That's the full, GPT-5.6-Sol model -- up to 14× faster than the same model on Standard processing. It speedran Humanity’s Last Exam in 11h 11m, nearly 7× faster than Claude Fable 5 with comparable accuracy.

Cerebras

1,293,179 views • 2 months ago

Qwen3.8-27B is now live at Cerebras speed. The dense, open-weight model from Qwen scores 34 on the Artificial Analysis Intelligence Index—making it comparable to models such as GPT-5.6 Luna, DeepseekV4 Pro, and Claude Sonnet 4.6.

Qwen3.8-27B is now live at Cerebras speed. The dense, open-weight model from Qwen scores 34 on the Artificial Analysis Intelligence Index—making it comparable to models such as GPT-5.6 Luna, DeepseekV4 Pro, and Claude Sonnet 4.6.

Cerebras

294,575 views • 1 month ago

We’re introducing intrinsic discovery, a new learning method that allows models to explore and learn from open-ended environments without human supervision. Using intrinsic discovery, we trained a world model that surpasses GPT-5.6 Sol at simulating terminal environments.

We’re introducing intrinsic discovery, a new learning method that allows models to explore and learn from open-ended environments without human supervision. Using intrinsic discovery, we trained a world model that surpasses GPT-5.6 Sol at simulating terminal environments.

Induction Labs

69,224 views • 1 month ago

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...

kwindla

334,795 views • 1 month ago

Okay… DeepSeek V4 Flash actually COOKED 😭 With OpenCode, I’d place it around GPT-5.6 Terra Max—and definitely above Luna Max in this test. Almost unlimited usage for around €9, by the way. This model is a MASSIVE success. Just don’t use the wrong harness.

Okay… DeepSeek V4 Flash actually COOKED 😭 With OpenCode, I’d place it around GPT-5.6 Terra Max—and definitely above Luna Max in this test. Almost unlimited usage for around €9, by the way. This model is a MASSIVE success. Just don’t use the wrong harness.

OmedTheVibeCoder

35,417 views • 2 months ago

GPT 5.6 Sol & Terra just made Fable 5's pricing look like a joke! We gave OpenAI's top models (Sol, Terra) and Anthropic's top models (Fable 5, Opus 4.8) the same 3 prompts: • Supernova boom • Meteor hitting a city • Solar system model The bill: Sol: $4.77 Terra: $1.24 Fable 5: $9.94 Opus 4.8: $2.46 Outputs came out surprisingly close. The prices didn't. Fable 5 cost unreasonably more than everything else, with Sol not far behind. Considering latest Fable 5 nerf, it's hard to see what you're paying for.

GPT 5.6 Sol & Terra just made Fable 5's pricing look like a joke! We gave OpenAI's top models (Sol, Terra) and Anthropic's top models (Fable 5, Opus 4.8) the same 3 prompts: • Supernova boom • Meteor hitting a city • Solar system model The bill: Sol: $4.77 Terra: $1.24 Fable 5: $9.94 Opus 4.8: $2.46 Outputs came out surprisingly close. The prices didn't. Fable 5 cost unreasonably more than everything else, with Sol not far behind. Considering latest Fable 5 nerf, it's hard to see what you're paying for.

AI/ML API

149,978 views • 3 months ago

Our research team just released Flex-Forcing: a video generation method that lets a single model switch between generation methods at inference time. Right now there are two main approaches to video generation. Bidirectional diffusion models attend to every frame at once, holding structure well at the cost of speed. Autoregressive models generate frame by frame, so they stream fast and scale to long clips, but accumulate error and drift over time. Flex-Forcing trains a single model to do both, letting you choose from the range at inference based on your compute budget.

Our research team just released Flex-Forcing: a video generation method that lets a single model switch between generation methods at inference time. Right now there are two main approaches to video generation. Bidirectional diffusion models attend to every frame at once, holding structure well at the cost of speed. Autoregressive models generate frame by frame, so they stream fast and scale to long clips, but accumulate error and drift over time. Flex-Forcing trains a single model to do both, letting you choose from the range at inference based on your compute budget.

NVIDIA AI

31,412 views • 3 months ago

you can now use Ox Alpha, Kimi K3, gemini 3.7 flash and 47 other models for FREEE with no credit card 😳 aihubmix just opened a 50-model free catalog, every model $0/M in and out. the models: -Ox Aplha -gemini-3.7-flash -glm-5.2 -kimi-k3 -minimax-m3 how to set it up (2 min): step 1: sign up at email or oauth, no card step 2: create an api key one key, all models, free and paid alike step 3: call any -free model base url: model: coding-glm-5.2-free, gpt-5.5-free, whatever that's it the free catalog is live in 2 minutes works in any openai-compatible tool. save this if you want a simple multi model setup later

you can now use Ox Alpha, Kimi K3, gemini 3.7 flash and 47 other models for FREEE with no credit card 😳 aihubmix just opened a 50-model free catalog, every model $0/M in and out. the models: -Ox Aplha -gemini-3.7-flash -glm-5.2 -kimi-k3 -minimax-m3 how to set it up (2 min): step 1: sign up at email or oauth, no card step 2: create an api key one key, all models, free and paid alike step 3: call any -free model base url: model: coding-glm-5.2-free, gpt-5.5-free, whatever that's it the free catalog is live in 2 minutes works in any openai-compatible tool. save this if you want a simple multi model setup later

painn

81,659 views • 2 months ago

This test shows clearly why GPT 5.6 Sol wins I asked the simplest prompt ever : > Create a gmail clone, the nearest possible to the real gmail ui You can see the result, GPT 5.6 Sol and Fable make a 100% perfect But here's the price difference: → GPT 5.6 Terra: $1.04 - 8 minutes → GPT 5.6 Luna: $0.44 - 10 minutes → GPT 5.6 Sol: $6.72 - 20 minutes → Claude Fable 5: $25.09 - 40 minutes Fable is 3.7x more expensive than Sol for "the same result"

This test shows clearly why GPT 5.6 Sol wins I asked the simplest prompt ever : > Create a gmail clone, the nearest possible to the real gmail ui You can see the result, GPT 5.6 Sol and Fable make a 100% perfect But here's the price difference: → GPT 5.6 Terra: $1.04 - 8 minutes → GPT 5.6 Luna: $0.44 - 10 minutes → GPT 5.6 Sol: $6.72 - 20 minutes → Claude Fable 5: $25.09 - 40 minutes Fable is 3.7x more expensive than Sol for "the same result"

Melvyn • Builder

22,279 views • 3 months ago

Model router in Foundry Models now supports the GPT-5.6 family and Claude Opus 4.8, with expanded regional availability. Read the latest blog to learn why model selection is becoming a continuous optimization loop:

Model router in Foundry Models now supports the GPT-5.6 family and Claude Opus 4.8, with expanded regional availability. Read the latest blog to learn why model selection is becoming a continuous optimization loop:

Microsoft Azure

37,804 views • 2 months ago

Someone vibe coded a fully interactive 3D human anatomy app using GPT-5.6 Sol and Codex, and it started from a single design image. GPT Image 2.0 generated the design. Every organ image was generated from it and converted to 3D models through Tripo AI. Codex took those models, a master prompt and the design image and built the first version. The models were nearly 150MB each and running at 16fps, so Codex optimized everything down to 2 to 5.5MB per model and reduced total assets from 900MB to 28.6MB while keeping the visual quality intact. Codex also generated the anatomical illustrations showing where each organ sits, and created the interactive hotspot markers explaining different parts of every organ. Live and fully open source.

Someone vibe coded a fully interactive 3D human anatomy app using GPT-5.6 Sol and Codex, and it started from a single design image. GPT Image 2.0 generated the design. Every organ image was generated from it and converted to 3D models through Tripo AI. Codex took those models, a master prompt and the design image and built the first version. The models were nearly 150MB each and running at 16fps, so Codex optimized everything down to 2 to 5.5MB per model and reduced total assets from 900MB to 28.6MB while keeping the visual quality intact. Codex also generated the anatomical illustrations showing where each organ sits, and created the interactive hotspot markers explaining different parts of every organ. Live and fully open source.

0xMarioNawfal

57,012 views • 2 months ago

Ran a real 3-turn LinkedIn research workflow through GMI Cloud Router in Balanced mode. Here’s what it actually picked, turn by turn : Turn 1 — “Research 2026 AI/marketing trends and summarize 5 insights” → DeepSeek-V4-Flash-0731 → ~$0.047 cheaper than Claude Opus 4.8 Turn 2 — “Analyze those insights for content angles + pain points” → GPT-5.6-Luna → ~$0.026 saved vs Opus 4.8 Turn 3 — “Turn the top angle into 5 LinkedIn post concepts” → GPT-5.6-Luna → ~$0.040 saved vs Opus 4.8 Total estimated savings: ~$0.11 across 3 turns. The interesting part was Turns 2 and 3. The task changed from analysis to writing, but Router stayed on GPT-5.6-Luna. That makes sense because GMI Router is cache-aware: when the same model can still handle the next task well, staying on that model can preserve cache reuse and avoid the cost of re-serving the context to another model. And all 3 tasks were relatively lightweight. They didn't suddenly require a more expensive frontier model just because the task type changed. So the routing wasn't about constantly switching models. It was about finding a model that was good enough for the task while keeping the cost down, and staying there when that still made sense. That’s a much more interesting way to think about model routing. GMI Cloud #AI #LLM #AIAgents #GenerativeAI

Ran a real 3-turn LinkedIn research workflow through GMI Cloud Router in Balanced mode. Here’s what it actually picked, turn by turn : Turn 1 — “Research 2026 AI/marketing trends and summarize 5 insights” → DeepSeek-V4-Flash-0731 → ~$0.047 cheaper than Claude Opus 4.8 Turn 2 — “Analyze those insights for content angles + pain points” → GPT-5.6-Luna → ~$0.026 saved vs Opus 4.8 Turn 3 — “Turn the top angle into 5 LinkedIn post concepts” → GPT-5.6-Luna → ~$0.040 saved vs Opus 4.8 Total estimated savings: ~$0.11 across 3 turns. The interesting part was Turns 2 and 3. The task changed from analysis to writing, but Router stayed on GPT-5.6-Luna. That makes sense because GMI Router is cache-aware: when the same model can still handle the next task well, staying on that model can preserve cache reuse and avoid the cost of re-serving the context to another model. And all 3 tasks were relatively lightweight. They didn't suddenly require a more expensive frontier model just because the task type changed. So the routing wasn't about constantly switching models. It was about finding a model that was good enough for the task while keeping the cost down, and staying there when that still made sense. That’s a much more interesting way to think about model routing. GMI Cloud #AI #LLM #AIAgents #GenerativeAI

Tom

29,592 views • 1 month ago

Been giving my M3 Ultra Studio some love and tinkering with MLX. Started the afternoon at 42 tok/s. Now: 73 tok/s, and there’s still more on the table. Running Qwen3.8-Flash-Next @ 4-bit locally. Now for the crazy bit: Artificial Analysis Intelligence Index v4.3: GPT-5.6 Sol 39 (Medium) GPT-5.6 Sol 42 (High) Qwen3.8-Flash-Next 42 GPT-5.6 Sol 44 (XHigh) GPT-5.6 Sol 47 (Max) Grok 4.5 39 (High) Claude Sonnet 5 38 (Max) *Qwen score currently estimated by Artificial Analysis. A 4-bit open-weights model running at 73 tok/s on a Mac under my desk... with an estimated intelligence score matching GPT-5.6 Sol at High reasoning. Local AI is getting ridiculous. 🔥 Plus, this model does not say no to a bit of red teaming. Once I am happy with it, I will do a pr for Jun Kim's oMLX :)

Been giving my M3 Ultra Studio some love and tinkering with MLX. Started the afternoon at 42 tok/s. Now: 73 tok/s, and there’s still more on the table. Running Qwen3.8-Flash-Next @ 4-bit locally. Now for the crazy bit: Artificial Analysis Intelligence Index v4.3: GPT-5.6 Sol 39 (Medium) GPT-5.6 Sol 42 (High) Qwen3.8-Flash-Next 42 GPT-5.6 Sol 44 (XHigh) GPT-5.6 Sol 47 (Max) Grok 4.5 39 (High) Claude Sonnet 5 38 (Max) *Qwen score currently estimated by Artificial Analysis. A 4-bit open-weights model running at 73 tok/s on a Mac under my desk... with an estimated intelligence score matching GPT-5.6 Sol at High reasoning. Local AI is getting ridiculous. 🔥 Plus, this model does not say no to a bit of red teaming. Once I am happy with it, I will do a pr for Jun Kim's oMLX :)

Ash Hart

18,827 views • 1 month ago