AI builders get more token options
More tokens, more agents, more choices. Today’s useful thread: put abundant model access to work, then keep your workflows simple enough to trust.
In today’s edition
- Supabase says Astra tokens are unlimited, giving builders another place to test token-heavy workflows.
- Omarchy will route all sponsored tokens through OpenRouter as it positions itself as an agent-first operating system.
- Mistral raised a €3B Series D to expand its AI infrastructure, products, and deployments at scale.
- A practical 3D training recipe pairs deterministic geometry rewards with a vision model acting as judge.
- GPT-6 Pro showed promise as a paper reviewer that surfaces substantive issues alongside strengths.
Supabase says Astra tokens are unlimited
Token limits can turn a promising build session into a budgeting exercise. Supabase says Astra tokens are now unlimited on its platform.
- The announcement specifically promises unlimited Astra tokens at Supabase.
- That can matter for builders running iterative prompts, long agent loops, or repeated prototype tests.
- The post does not include plan details, usage rules, or technical setup information.
What it means If you already build on Supabase, test your most token-hungry prototype there and measure whether unlimited Astra access changes the workflow economics.
Source: supabase
Omarchy routes sponsored tokens through OpenRouter
Agent-first computing is moving closer to the desktop. Omarchy calls itself the first operating system built from the ground up for agents and says it will use OpenRouter for all sponsored-token routing.
- Omarchy is positioning its operating system around agents from the start.
- OpenRouter says it will route all sponsored tokens for Omarchy.
- The setup puts model-token routing into the operating-system layer rather than leaving it entirely to each app.
What it means Watch agent-native operating systems closely, because model routing at the platform layer could reshape where you build integrations, manage spend, and reach users.
Source: OpenRouter
Mistral raises €3B for AI expansion
Mistral just stocked the war chest. The company announced a €3B Series D, which it calls the largest equity round ever raised by a European tech company, three years after launch.
- Mistral says the round totals €3B.
- It describes the financing as the largest equity round ever raised by a European tech company.
- In a related post, Mistral says the money will fuel infrastructure, product capabilities, and deployment at scale.
- Mistral also frames its open-weight models, products, and infrastructure as a choice over how and where organizations run AI.
What it means Keep Mistral on your vendor shortlist if deployment choice and open-weight options matter, because this funding supports a larger infrastructure and product push.
Source: MistralAI
A 3D training recipe worth stealing
Want a smaller model to make usable 3D rooms in Blender? One builder’s experiment points to a concrete recipe: align geometry with deterministic rewards, then use a vision-language model as the judge.
- The experiment used small Qwen3.5 9B to create low-poly 3D rooms in Blender.
- The builder says geometry needs alignment with deterministic rewards.
- It also uses a VLM-as-a-judge approach.
- The challenge is available for people, or Astra, to try.
What it means For visual-agent projects, define machine-checkable geometry goals and add a vision-model judge instead of relying on vague prompts and subjective output reviews.
Source: mervenoyann
Try GPT-6 Pro as a first reviewer
The blank-page problem has a cousin: the first-review problem. Ethan Mollick says GPT-6 Pro is strong at reviewing academic papers, finding substantive issues while balancing them against strengths and suggesting additions.
- The suggested prompt is simple: upload a paper and ask, “review this for me.”
- Mollick says the model finds real issues rather than only minor complaints.
- He also says it balances critiques with a paper’s strengths and recommends additions.
- This is a practical review workflow for structured, high-stakes drafts.
What it means Use an AI review pass before sending important drafts, but ask for strengths, substantive issues, and additions so feedback becomes a prioritized revision plan.
Source: emollick