Three New Models Skip the General Race and Pick One Job
Mistral's large open-weight model targets coding and security, Canva Brush targets design edits, and Baseten's joey-1 targets fast decisions — each measured on its own task, not on everything.
- Mistral Large 4 is a 1-trillion-parameter multimodal model under the Mistral Open-Weights license, priced at $0.70 per million input tokens and $2.10 per million output tokens. Self-reported m:…9481
- Canva Brush is a vision-language model for design, reported 69% faster and 46% lower cost per edit than the previous design agent. Self-reported m:…3515
- Baseten joey-1 is a 2B-parameter decision model with 20 ms median latency and $0.01 per million tokens. Self-reported m:…7269
- Canva Brush was post-trained on real design and editing tasks and uses an agentic loop that writes and runs code to adjust design elements. Self-reported m:…3515
One model per job
Mistral AI released Mistral Large 4, a 1-trillion-parameter multimodal model available under the Mistral Open-Weights license, with three reasoning modes called Now, Medium and High. m:…9481
Mistral lists three reasoning modes for the model, called Now, Medium and High, so the same model is presented as adjustable for different kinds of coding and security work rather than running at a single fixed setting. m:…9481
The company says the model leads open-weight models on the FrontierCode 2.0 benchmark with a score of 19.7 percent, and reports a 96.6 percent average refusal rate on malicious cyber requests. m:…9481
Mistral prices the model at 0.70 dollars per million input tokens and 2.10 dollars per million output tokens, according to the announcement. m:…9481
For developers, the combination of an open-weights license and availability through Microsoft Foundry for Azure developers means the same model can be run under their own arrangements or inside a cloud platform, rather than only through a single vendor endpoint. m:…9481
Mistral also says the model is available through Microsoft Foundry for Azure developers, resists 97.1 percent of attacks on the B3 Security Benchmark, and is natively fluent in more than 160 languages, with a specialized variant for enterprise design partners optimized for cybersecurity workflows. m:…9481
Design edits as a model task
Canva launched Canva Brush on October 29, 2026, a vision-language model for design that is available to Canva Pro subscribers from that date. m:…3515
The company says Brush is 69 percent faster for general design editing tasks than the previous design agent, and 46 percent lower cost per edit task, and it uses an agentic loop that writes and runs code to adjust design elements. m:…3515
Brush was post-trained on real design and editing tasks, and Canva says it plans to add Brand Kit support, attachment references and better handling of image-heavy layouts in the future. m:…3515
Brush was developed by Canva Research, a 160-person AI lab, and trained using large-scale reinforcement learning; Canva Craft, an image-generation model, was released by Canva one week before Brush, and DesignArena measured Canva as the world's second most advanced AI lab for image generation, according to the announcement. m:…3515
Decisions at a penny scale
Baseten released joey-1, a 2-billion-parameter decision model that returns a label and a confidence score in one forward pass, post-trained from Qwen3.5-2B. m:…7269
The company reports a median end-to-end latency of 20 milliseconds and a price of 0.01 dollars per million tokens, with a 16,000-token context window on its Model APIs. m:…7269
Dedicated deployments of joey-1 support a 262,000-token context window, the model is available with open weights on Hugging Face, Baseten offers zero data retention by default for joey-1 on its Model APIs, and the company says joey-1 is 2.5 times faster than elephant-1 in median end-to-end latency. m:…7269
The contrast matters for buyers: Mistral sells capability for coding and security work, Canva sells faster and cheaper edits inside a design tool, and Baseten sells a narrow decision at very low latency and cost, and none of the three asks to be judged as a general assistant. m:…9481 m:…3515 m:…7269
A buyer comparing the three should therefore compare different numbers for each one: a benchmark score and a refusal rate for Mistral, speed and cost per edit for Canva, and latency and price per million tokens for Baseten, because each vendor is pricing a different kind of work. m:…9481 m:…3515 m:…7269
What is still unresolved
- All three performance and price claims come from the vendors, with no independent benchmark results in the cited material.
- Whether Mistral Large 4's refusal rate holds outside the tested malicious-request set is not detailed in the cited material.
- Canva's future Brush features have no dates in the cited material.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.