Google made Gemini 4 Carbon generally available on 10 Nov at $2 / $10 per million tokens, a price that holds only until 30 Nov. Meta shipped a $0.10 worker model a day earlier, and on 11 Nov two labs opened world models for robots under Apache 2.0. I read the announcements, docs, model cards and benchmark tables on 12 Nov. I did not run any of these models, so every number below belongs to the source named next to it.
The thread this week: the big model and the cheap model are becoming one stack. One plans, the others grind.
💸 Is Gemini 4 Carbon actually cheap?
Gemini 4 Carbon went GA on 10 Nov, about two weeks after Google opened it only to cyber defenders in the Fairwind Program, which our 29 Oct digest covered.
Per Google’s announcement by Koray Kavukcuoglu, Carbon costs $2 input and $10 output per 1M tokens through 30 Nov 2026, the same as Gemini 4 Argon, then $4 / $20 from 1 Dec. Cached input is 95% off. LiteLLM’s pricing notes add batch and flex at $1 / $5, priority at $3.60 / $18, and one line that matters: thinking tokens bill as output.
Google says Argon stays the default for most software and knowledge work. Carbon is for the hardest multi-step coding, debugging and research, and Google says it “uses more tokens than Argon”. So in November both cost the same per token, yet Carbon can cost more per task because it thinks longer. Cheaper per token does not mean cheaper per job.
Switching is a model string change, per the developer guide: gemini-4-carbon, default thinking medium, with Argon’s 1M-token context, thinking_level controls and tools.
The benchmark claims are Google’s own. Google says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench “at more than 2X lower cost per task compared to the next best external model”. On the new Artificial Analysis Multimodal Knowledge Agent (MKA) benchmark, built with Google, Carbon is second at more than 6X lower average cost per task than the top model.
For context, VentureBeat’s price table lists Claude Opus 5.5 at $4 / $20, Claude Fable 5.2 and Mythos 5.2 at $8 / $40, and GPT-6 Astra at $10 / $50. The Fable side is in this week’s Fable 5.2 and Mythos 5.2 deep dive. After 1 Dec, Carbon’s list price matches Opus 5.5.
Google also opened a /swarm preview in Antigravity: one prompt becomes a team of agents, with Carbon as planner and auditor, lighter models doing parallel work, and adversarial verifiers checking it. Access covers the Gemini API paid tier, Gemini for business, Google AI Ultra, AI Mode and Antigravity. The cyber variant stays in Fairwind.
Takeaway: run your own eval before 30 Nov while testing is cheap, and measure cost per finished task, not per token.
🔥 What is a ten cent model for?
Meta Superintelligence Labs released Muse Ember on 9 Nov, its smallest and most efficient Muse model. Meta’s post aims it at “the work agents repeat all day”: tool calls, coding subtasks, routing, extraction and long document parsing. Per the docs, muse-ember-1.0 runs on the Meta Model API and Muse Code with a 1,048,576-token context.
VentureBeat, reading Meta’s pricing page, reports $0.10 input, $0.004 cached input and $0.50 output per 1M tokens up to 200K input, then $0.30 / $0.012 / $1.50 above it. The same piece cites Artificial Analysis: an Intelligence Index of 36 at max effort, the same rounded score as Claude Haiku 5.5, at about $0.14 per task against $0.41 for Haiku. Meta also claims more intelligence per dollar than Haiku 5.5 or GPT-6 Luna at any reasoning setting. That claim is Meta’s own.
Ember is the first model trained using Muse Wonder Full, and Meta’s framing is blunt: “Muse Wonder plans, Muse Ember handles the volume.” It is closed, with no weights. Meta says it tested Ember under its Superintelligence Scaling Framework and found it below the high-risk thresholds.
This is what makes the Carbon story affordable. The planner is expensive, so it should make few calls while a cheap worker loops through the small steps. Google’s /swarm and Meta’s Wonder plus Ember pairing describe the same pattern.
Takeaway: watch the 200K line. If your agent stuffs long histories into every call, the input price triples.
🖥️ Why would your dot want a second computer?
A dot can now connect to several of your own computers, and you pick which one runs a task, per OpenAI’s community forum post of 9 Nov, 22:28 UTC. Each thread uses one computer at a time, so this is parallel independent tasks, not one task split across machines.
OpenAI’s examples include testing an iOS change in Xcode on a second Mac, or clicking through a UI flow on another computer so your own cursor stays free.
Per the dots docs, you choose “Allow access” in the ChatGPT app on each computer, which must stay online with the app open. Only “Revoke access” removes it.
According to ITmedia, dots launched at DevDay on 29 Sept on GPT-6 Astra and GPT-6.1 Terra. I am in Japan this month, so one line caught my eye: per OpenAI, Japan is second by number of dots users.
Takeaway: a spare machine as a test rig is the useful part. Each connected computer is one more box with your files on it, so connect a clean test machine, not the laptop that holds your keys.
🤖 Can a robot imagine before it moves?
Two labs released open world models on 11 Nov, both under Apache 2.0. The idea is to let an agent see what an action does before a robot does it.
PAN 2 from IFM at MBZUAI. Per MBZUAI’s announcement and the IFM blog, PAN 2 takes an optional starting image plus actions in text and simulates the outcome as video up to 20 seconds, at 704×1248 and 24 fps. An LLM “conceptual reasoner” writes a timed description of the next world state, then a 29B diffusion model trained from scratch fills in the visual physics. Weights and code are on Hugging Face and GitHub.
The scores are MBZUAI’s own. On its physical eval, PAN 2 sits just behind MiniMax-H3 and completed more multi-step plans than any model tested. It is third overall on Physics-IQ Verified and first among open-weight models, though the Hugging Face card notes that run used Claude Opus 5.5 as the reasoner. The stated limit is drift: each clip starts from one frame plus text, so long interactions wander.
Praxis-1 from Runway. Runway introduced Praxis-1 in September and has now opened it under Apache 2.0, commercial use included, with weights plus inference and fine-tuning code. It is a 27B mixture-of-experts diffusion transformer that predicts video and actions together.
Runway’s own numbers on NVIDIA RoboLab-120 show 46.2% success over 1,200 episodes, against 42.9% for FLUX 3 Action, 36.8% for NVIDIA’s Cosmos 3 Nano (Runway’s table spells it “Cosmos3 Nano”) and 28.0% for π0.5. Cosmos 3 Nano came up in Monday’s digest. It supports LeRobot and runs on anything from H100 to Jetson Thor.
Takeaway: “imagine first, act second” is becoming open infrastructure. Both labs graded their own work, so wait for someone else to reproduce it on different robots.
📦 What else shipped?
Baidu released ERNIE 5.5 Flash Preview on 11 Nov. Per Baidu’s release page, it has 280B total and 14B active parameters and a 1M context, with an API on Qianfan now and weights that “will be open-sourced under Apache 2.0 soon”. Baidu calls it an intermediate checkpoint, and I have no verified benchmark numbers to quote.
Google DeepMind released Gemini Robotics 2.5 and AlphaGenome 2 on 11 Nov, SiliconANGLE reports. Per SiliconANGLE, Robotics 2.5 was tested on more than 200 tasks and hit 96% success packing automotive parts in a Mercedes-Benz proof of concept, DeepMind’s own test. AlphaGenome 2 predicts how DNA variants affect specific cell types and is not for clinical use.
📊 How do this week’s releases compare?
Prices are in US dollars per 1M tokens, written as input / output.
| What | Who | Date | Open weights? | Price or access | Key number (source) |
|---|---|---|---|---|---|
| Gemini 4 Carbon | 10 Nov | No | $2 / $10 to 30 Nov, then $4 / $20 | #2 on MKA at 6X lower cost per task (Google) | |
| Muse Ember | Meta | 9 Nov | No | $0.10 / $0.50 | AA Index 36, about $0.14 per task (Artificial Analysis via VentureBeat) |
| dots multi computer | OpenAI | 9 Nov | No | Pro, Business Premium, Enterprise | One computer per thread |
| PAN 2 | IFM, MBZUAI | 11 Nov | Yes, Apache 2.0 | Hugging Face, GitHub | 29B diffusion model, 20 s clips |
| Praxis-1 | Runway | 11 Nov | Yes, Apache 2.0 | Hugging Face, GitHub | 46.2% on RoboLab-120 (Runway) |
👀 What should you watch next?
Carbon’s price doubles on 1 Dec. After that, watch whether /swarm leaves preview, when Baidu posts the ERNIE 5.5 Flash weights, and whether Meta opens Ember beyond its API and Muse Code. For PAN 2 and Praxis-1, the number I want is the first independent replication.
Sources
- Google, Gemini 4 Carbon now available
- Google AI Studio, Gemini 4 Carbon developer guide
- LiteLLM, Gemini 4 Carbon pricing notes
- Meta, Muse Ember
- Meta, Muse Ember docs
- VentureBeat, Gemini 4 Carbon price table
- VentureBeat, Muse Ember and Artificial Analysis
- OpenAI community, multiple local computers per dot
- OpenAI, dots computers and apps docs
- ITmedia, dots coverage
- MBZUAI, PAN 2 announcement
- IFM, PAN 2 blog
- PAN 2 on Hugging Face
- PAN 2 on GitHub
- Runway, Praxis-1 open weights
- Baidu, ERNIE 5.5 Flash Preview
- SiliconANGLE, Gemini Robotics 2.5 and AlphaGenome 2
