Black Forest Labs' open robot model: buy the cost curve, not the benchmark win

*A 7B open model edges Nvidia's Cosmos 3 Nano by 1.5 points and claims a 150-fold cost advantage per success over a frontier API, all on the vendor's own tests.*

Opening read

The most consequential number today is the roughly 150-fold gap in cost per successful robot task that Black Forest Labs claims between its new open 7B model and a frontier API, and I weight it above the 1.5-point benchmark win that headlines the same post. Both numbers come from the vendor's own post, and both describe models that still fail most tasks. The cost shape nonetheless settles how a robotics pilot should be structured, and that is a decision a team can make this week.

Robot models: the cheap retry beats the top score

Black Forest Labs put FLUX 3, a single model trained on images, video, and audio together, into early access in July and paired it with a robot-control variant that Audi was testing on factory tasks. This week it released FLUX 3 Action, an open-weight model (its trained parameters can be downloaded by anyone) with 7 billion parameters that predicts how a scene will unfold and picks a robot's next move. The consensus read is reasonable: the new model scores 38.3% on the RoboLab benchmark against 36.8% for Nvidia's Cosmos 3 Nano, a distilled version (a copy trained to reach its answer in fewer steps) reaches 42.2%, the backbone (the base model the robot variant is built on) is less than half the size of Nvidia's, and Cosmos is the program Nvidia treats as its world-model effort. Black Forest Labs (https://bfl.ai/blog/flux-3) Latent Space (https://www.latent.space/p/ainews-black-forest-labs-flux-3-multimodal) Black Forest Labs (https://bfl.ai/blog/flux-3-action) Nate Jones (https://natesnewsletter.substack.com/p/inside-nvidia-what-a-world-model) So a small lab taking the top open slot from Nvidia looks like a shift in who leads.

I would not buy on that gap. A 1.5-point edge on a benchmark score the vendor reported, where both models fail roughly six attempts in ten, tells a buyer little about their own robot cell, the workstation where the arm actually does the job. The figure that changes a decision is cost per success. By Black Forest Labs' own accounting, a successful task on GPT 6 Astra averages $13.47 and 16 minutes, while its own model gets there for $0.09 in under two minutes on an H200 GPU, and it runs up to 3.95 times faster than Cosmos 3 Nano in FP8, a compact number format that trades a little precision for speed, on everything from consumer cards to datacenter GPUs. Black Forest Labs (https://bfl.ai/blog/flux-3-action) I do not know how that comparison was staged, and a general frontier model was never built for this job, so I read the ratio as a ceiling. At face value the vendor's prices buy about 150 open-model attempts for the cost of one frontier success, and even at a tenth of that a team still gets roughly fifteen tries for the price of one. Either way, the economics of iteration flip.

The standing question in robotics is where the robot's brain should live, on the machine or somewhere else. Today's release does not settle it, and I would not read a 7B model in FP8 as proof it can sit beside the arm; what the post supports is a range of GPUs it runs on, and whether that includes the hardware in your own cell is a question for your own evaluation. What the release does change is the question a pilot should ask, from where the compute lives to how cheaply the control model, the part that picks the robot's next move, can be retried. Sakana AI's hiring of Jürgen Schmidhuber as chief scientific advisor points the same direction, since the company says it is building what it calls Agent-Native World Models to simulate the consequences of actions before a robot takes them, aimed at Japanese industry that wants to skip costly real-world trial and error. Sakana AI (https://sakana.ai/schmidhuber/) For an engineering lead running a manipulation pilot, the budget question becomes which model absorbs thousands of retries, and my answer is the cheap open one, with the frontier model kept as a judge or a generator of training data.

The check is one your team can run this month. Fix a set of manipulation tasks from your own cell, run each candidate model on all of them, and count successes rather than attempts. Divide total spend, GPU hours included, by that count. If your ratio lands anywhere near Black Forest Labs' figure, the frontier API belongs in evaluation and the open model belongs in the loop; if the open model's success count collapses on your tasks, the benchmark lead was the wrong signal and you learned it before you committed the budget.

On-device speed: a 9% bigger model that runs three times faster

If a local model failed your speed test this summer, that result is stale. Liquid AI has added a small helper model to its three-billion-parameter vision model, using a trick called speculative decoding: the small model guesses the next several tokens and the large model only checks them, which is faster than producing each one from scratch. The price is modest, 280M extra parameters, or 8.9% on top of the base, in exchange for decoding it says runs up to 3.13 times faster on an M5 Max and 2.66 times faster on an H100. Apple's researchers are working the same memory budget from the other side, shrinking a streaming audio encoder for on-device dictation because the always-on encoder competes for RAM with a sparsely activated model that loads only a few experts at a time, and Ollama's v0.40.0 release candidate now runs supported models on MLX, Apple's framework for its own chips, by default. Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm2-5-vl-dspark) Apple Machine Learning Research (https://machinelearning.apple.com/research/latent-space-distillation) Ollama Releases (https://github.com/ollama/ollama/releases/tag/v0.40.0-rc0) These are three pieces of one shift: the same laptop gets faster without new hardware. One caution before you cancel the bigger machine. Liquid AI's earlier 1.2B text-model DSpark result already swung by as much as 52% depending on what text the model was fed, so rerun the test on your team's real traffic rather than a demo prompt. Hugging Face Blog (https://huggingface.co/blog/LiquidAI/lfm25-dspark)

Enterprise avatars: Google ships the face before the use case

Google launched Gemini 3.8 Live on September 15 as its real-time voice model, and yesterday it added Live Avatar to Gemini Enterprise, a generated video persona that lip-syncs speech across 97 languages while calling tools in the background so the conversation never pauses. Google DeepMind Blog (https://deepmind.google/blog/introducing-gemini-3-8-live-and-3-8-live-extended-thinking/) Google DeepMind Blog (https://deepmind.google/blog/introducing-gemini-38-live-with-live-avatar/) The same day Google Research published a multi-agent director for long-form video built on Gemini and Veo, which inherits SynthID, Google's invisible watermark, from those models. Google Research Blog (https://research.google/blog/coherent-long-form-video-generation/) That paper shows Google building the video plumbing underneath features like the avatar; it is not evidence that any buyer wants a face on a support agent. My read is that the watermark solves Google's detection problem and leaves the buyer's trust problem open. The decision in front of a business leader is whether a support agent needs a face at all, and the questions to settle first are what the face adds beyond the demo and what a customer does on learning it was generated.

Buildout math: the server life assumption carries the return

Princeton's CITP blog pulls together two numbers that decide whether the buildout pays. The first, from Goldman Sachs, is that the chips and servers inside an AI data center run three to four times the price of the building that houses them, and the second is that the big cloud buyers now write those servers down over five to six years. Princeton CITP (https://blog.citp.princeton.edu/2026/09/24/does-the-current-ai-buildout-depend-on-an-automation-agenda/) Put together, they say the return rests less on construction cost than on how long the expensive part stays useful. The conventional figure for GPU life has been 3 to 5 years, and the GPUs are the costliest part inside those servers, so a five-to-six-year server life implies a longer assumed run for the expensive component than the old rule of thumb allowed. Nate Jones (https://www.youtube.com/watch?v=a-LF8VhwMeA) The two figures measure different things, a chip and the box it sits in, so I read this as a stretched assumption rather than a proven one. The Princeton post also relays Anthropic Institute scenarios for AI's share of US GDP in 2030 that run from 1.6% under modest assumptions to 32.4% under extreme ones. Princeton CITP (https://blog.citp.princeton.edu/2026/09/24/does-the-current-ai-buildout-depend-on-an-automation-agenda/) If you are the executive signing a multi-year AI commitment, you are implicitly betting on the vendor's depreciation schedule, and I would ask for that assumption in writing before discussing price.

Augur calls

• Call: Black Forest Labs will publish downloadable open weights for the FLUX 3 Dev backbone by the end of 2026.
• Criteria: A public repository (Hugging Face or bfl.ai) hosting FLUX 3 Dev weight files with a model card, referenced from a Black Forest Labs blog post, on or before the due date. No such repository by the due date resolves as a miss.
• Due: 2026-12-31
• Odds: 55%

What to watch

• Black Forest Labs has promised FLUX 3 Dev as an open-weight variant since July; a weights drop would put the base model in outside teams' hands, though FLUX 3 Action is a separate model and a Dev release alone would not let anyone rerun its RoboLab numbers. Black Forest Labs (https://bfl.ai/blog/flux-3)
• Ollama says more models will be tested and enabled during the v0.40.0 release candidate period, so the stable release will show how much of the local catalog runs on MLX by default. Ollama Releases (https://github.com/ollama/ollama/releases/tag/v0.40.0-rc0)
• Google's Live Avatar is now in Gemini Enterprise; named customer deployments, rather than launch demos, would show whether buyers want a face on their agents. Google DeepMind Blog (https://deepmind.google/blog/introducing-gemini-38-live-with-live-avatar/)
• Sakana's Recursive Self-Improvement Lab, its effort to have AI systems improve AI research, now has Schmidhuber attached, and a paper or demo of its Agent-Native World Models would turn a stated plan into something testable. Sakana AI (https://sakana.ai/schmidhuber/)
• Fireworks ships ARCv3, a tool that shrinks the model updates sent between datacenters during training, as a package for teams running their own trainers, and adoption outside its own platform would test the cross-region training claim. Fireworks AI (https://fireworks.ai/blog/arcv3-and-the-case-for-cross-region-rl)

Verification notes

Every FLUX 3 Action figure in the lead, including the RoboLab scores, the 3.95x speed claim, and the cost per success against GPT 6 Astra, comes from Black Forest Labs' own post, and the evidence available to this brief does not show how the Astra comparison was set up. Black Forest Labs (https://bfl.ai/blog/flux-3-action) The FLUX 3 Dev open-weight backbone remains a promise dating to July rather than a shipped release. Black Forest Labs (https://bfl.ai/blog/flux-3)

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-09-25?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-09-25)