Supabase's October 2 announcement pairs $150M in new funding with plans to acquire Turso and a revealing usage metric: the company says agents and AI tools create 70% of its four million new databases each month. That counts creation, not successful production deployments. Halluminate's $30M Series A, announced the day before, backs evaluations for professional work: a different way to ask what the output is worth.
Proximal proposes turning workflow traces into evaluations and targeted training data; Cloudflare and Strands released models for bounded decisions; OpenAI introduced Sol at lower standard token rates than Astra. Together, these give founders more specific inputs for testing a workflow. The useful comparison includes the work a recipient accepts and the review required to get there. Define that job before choosing what to pay for the model.
Supabase announced $150M in new funding led by GIC on October 2 and said it is acquiring Turso, while continuing its Postgres and Turso's SQLite products. It reports four million new databases each month, 70% created by agents or AI-driven tools. That is database creation, not retained customers, revenue or production success; the acquisition's price and closing date are undisclosed. For builders, the useful next measurement is what those databases enable a user to complete.
OpenAI introduced Dots on September 29 with a cloud computer and connected-app access. Proactive research is read-only; assigned work follows allow, ask or block rules, with automated action review. Enterprise beta access is admin-enabled and off by default: permission to act and ability to finish correctly remain separate questions.
Apple's October 2 developer notice says it will introduce additional Full Disk Access controls requiring very explicit user action. Apple links the change to the risks of increasingly capable autonomous agents; it gives no delivery date or macOS version. A desktop agent's workable scope depends on what a customer knowingly authorizes.
Halluminate announced a $30M Series A led by Oak HC/FT on October 1, bringing funding to $38.5M. It builds evaluations and reinforcement-learning environments for knowledge work, initially finance. The investment backs a way to define better performance; it does not establish measured returns for customers.
Proximal introduced its approach on September 29 and disclosed a $15M seed led by General Catalyst. The company describes turning agent traces and work artifacts into evaluations, then targeted post-training data. The fresh signal is the public introduction; the announcement is not proof of lower customer costs or a fresh cash wire this week.
Rig Security emerged from stealth on September 29, disclosing $12M seed co-led by Ten Eleven Ventures and Brightmind Partners. The announcement concerns identity protection as agents use employee and service-account credentials; reporting places the financing close in late 2025. Attribution helps a team determine which agent performed an action and investigate what went wrong.
Atomic announced a $12.5M Series A co-led by Klass Capital and Madrona on September 29. The company says it automates 90% of DashMart purchasing, a named-customer claim rather than an independently measured industry rate. It puts the discussion at the level of a business workflow, where exceptions and review effort matter alongside throughput.
Meta expanded Muse for Small Business on September 29, adding skills and connectors for US and Canadian businesses. Meta says the product does not publish, send or spend without approval. For a small operator, the relevant result is the work that clears that approval queue, rather than the volume of drafts generated.
Salesforce announced a definitive agreement to acquire Listen Labs on September 29, with closing expected in its fourth fiscal quarter of 2027, subject to conditions including regulatory approval. The primary announcement gives no price. Listen's interviews and customer simulations offer research inputs; simulated responses do not establish actual buying behavior.
Tiny Health announced a $33M Series B led by B Capital on September 29, bringing funding to $46M, alongside a $5M research program. Its microbiome-testing business makes evidence quality a separate undertaking from financing. A larger dataset or round does not by itself establish a new clinical claim.
Cloudflare released Clef and Clef-flash on October 1, with Apache 2.0 weights and Workers AI access. Strands separately released Decider2B with weights, training data and scripts that day. These models target bounded decisions rather than arbitrary text generation; vendor benchmark results do not establish accuracy or latency for your workflow.
OpenAI introduced GPT-6.1 Sol at DevDay on September 29. Standard API rates are $2 per million input tokens and $10 per million output tokens, versus Astra's $10 and $50; cached-input pricing has a different ratio. This is a new model's rate comparison, not a cut to every existing model or proof that a customer's task costs fell fivefold.
California signed SB1246 on September 30. The senator's announcement specifies requirements by July 1, 2028, including US-based licensed remote drivers, local incident contacts and notifications during system-wide failures. Those are future obligations: operating an autonomous fleet includes the people and processes needed when normal operation fails.
ElevenLabs announced a completed $300M employee tender on September 30 at a $22B valuation. This is secondary liquidity, not $300M of new company financing; the company also says enterprise customers account for 55% of revenue. The transaction price and the operating evidence answer different questions about the business.
Anthropic launched Claude Frontier Academy on October 2 with a $100M commitment and a goal of training 10,000 Frontier Deployed Engineers by end-2027. Nomination-based cohorts face a graded simulation followed by a 12-week real-project residency; first final badges are expected in early 2027. The goal is not completed training, and a credential still needs to be read alongside delivered work.
The interesting use for cheaper intelligence is a bigger promise to the customer. Name what that customer must receive. A database created, a campaign drafted and a purchase recommendation prepared are intermediate steps; the promise is the usable result someone trusts enough to act on.
In #020 ↗, I asked readers to test an open-weight model on their main workload by October 1. The date has passed and the test remains unanswered. I don't have a measured result to report a win or concede a failure. This week's token prices cannot settle it. The quality standard and the actual bill still have to meet.
Halluminate and Proximal put evaluation at the center of their proposals. Sol gives you a lower standard token rate than Astra. Read them together as an invitation to separate the job into parts: the decisions a smaller model can attempt, the outputs a person must review, and the exceptions that need another route.
Choose one recurring job, agree acceptance criteria with its recipient, and set a decision date. Compare your current workflow with a shadow trial, including failures, retries and reviewer time in cost per accepted job. Switch only if the trial meets the same bar at lower total delivery cost; otherwise keep the current route.
When the test earns a handover, use the freed capacity to pilot a larger part of your customer's workflow. Include any cleanup you leave them with. Keep learning after the switch; the test earns your next bet, not permanent trust.