Sakana's self-improving loop is an efficiency result, not a capability leap

*A mid-sized model graded its own agent swarm and raised score per token; that is the number to test before you move any budget*

Opening read

Sakana AI's new self-improvement method is the day's most consequential evidence, and my read is that it is a cost result dressed as a capability result. The gain is counted per output token (the word-sized chunks a model produces and is billed by), it comes after a training loop in which the model picked its own best workflows, and it arrives in a field that has already shown multi-agent setups can cut performance on step-by-step tasks by more than a third. The figure to test is cost per accepted answer, not the headline multiple.

Self-improvement: Sakana's loop lifts score per token, with the model as judge

Sakana AI has published a method it calls Multi-Agent Self-Supervision, or MASS. One shared model plays several temporary subagents, tries different ways of working through a task, and keeps whichever approach its own judgment rates highest. Sakana says two rounds of this self-training took a 27 billion parameter model (parameters are a model's learned settings) to a benchmark score, meaning a score on a standard test set, of 1.2 to 1.6 times the base model's level for each token it produced, measured across four research tests, with the training itself done on synthetic open-ended research tasks, which are machine-generated tasks rather than real user work Sakana AI (https://sakana.ai/mass/). That is the headline number. The consensus take is easy to state: this is the recursive self-improvement loop labs have chased, and it backs Sakana's longer bet that many small efficient models can match big ones without giant data centers Sakana AI (https://sakana.ai/japan-startup-award-2026/). I understand the pull. A model that improves itself on machine-made tasks is about the cheapest training loop you could ask for.

My view departs on what was measured and how the loop graded itself. Score per token is an efficiency figure: it tells you the model wastes fewer words per unit of quality, not that it reaches answers it could not reach before. The training tasks were synthetic, and during training the model chose its own best workflows by its own judgment. Sakana also reports that training only on the solver's task data improved both its optimizer and its evaluator, the roles that refine the work and grade it, which is the striking part and the risky part at once Sakana AI (https://sakana.ai/mass/). When judge and student share the same weights (the same learned settings), a shared blind spot scores as progress. That is not a hypothetical worry. Separate research on recursive self-evaluation concluded such loops are workable when the validity of self-produced rewards is split out and monitored on its own, and the evidence I have shows no outside check of that kind here Hugging Face Daily Papers (https://tldr.takara.ai/p/2608.24231). I do not know how the gains hold on tasks beyond those four research benchmarks, or with a person grading.

The architecture evidence tells you where this should work and where it should not. Google Research ran 180 agent configurations under controlled conditions and found a coordinated multi-agent setup improved parallelizable work (work that splits into independent pieces) such as financial reasoning by 80.9 percent, while tasks that demand strict step-by-step reasoning fell 39 to 70 percent against a single agent, and Nate Jones's reading of related Google-MIT work adds a threshold: once a lone agent clears roughly 45 percent accuracy, adding agents yields shrinking or negative returns Google Research Blog (https://research.google/blog/towards-a-science-of-scaling-agent-systems-when-and-why-agent-systems-work/) Nate Jones (https://www.youtube.com/watch?v=2EXyj_fHU48). That split matters for Sakana because of what it chose to test. Open-ended research looks closer to the splittable kind of work, which would make it friendly ground; that is my inference, not Sakana's claim. Selection is the harder problem, and MASS's core move is picking which approach to keep. A nearby version of that problem showed up in a study of 87 agent tasks: it was choosing add-on skills rather than workflows, but the same skills helped some setups and hurt others in 36.78 percent of tasks, relevance rankings failed to find the useful ones, and ranking candidates by whether they supported the operations a task required lifted first-choice pass rates 4.35 to 5.80 percentage points across three configurations Hugging Face Daily Papers (https://tldr.takara.ai/p/2610.08875). Picking well is hard, and the fix that worked there was an outside rule, not a relevance score.

For an engineer deciding whether a small self-tuned model can replace a frontier API (a paid interface for calling one of the biggest vendor models over the network), the stake is whether multi-agent setups like MASS cut cost per accepted answer on your own task mix. Try this: pull 20 recent tasks and sort them into two buckets, ones that split into independent sub-questions and ones where each step depends on the last. Run each bucket once with a single agent and once with a subagent-orchestrated pass (one where a lead agent splits the job among helpers), log tokens used, and have a person mark which outputs they would accept. Compute cost per accepted answer per bucket. If the step-by-step bucket loses acceptance under orchestration, the Google result is holding in your shop; if the orchestrated pass's per-token edge vanishes once a person grades, a per-token headline like Sakana's is not a reason to move spend. For a business leader, that reads as budget for a pilot on small models, not a cut to frontier contracts yet.

Decision models: TypeSafe's first-week $100 million meets OpenAI's fast follow

TypeSafe's Jev, a decision model that takes a chunk of text and a question with fixed answer choices, picks one, and attaches a calibrated probability (a confidence figure meant to match how often it is right), has drawn both a funding headline and a fast-following rival within three weeks of launch. Sequoia reportedly leaked that TypeSafe crossed $100 million in annual recurring revenue (the yearly value of current contracts) in its first week, the company is calling its round a "Series AI" with a headline valuation of $7.5 billion, and an earlier report had it raising over $1 billion at above $10 billion Latent Space (https://www.latent.space/p/ainews-typesafejev-at-100m-arr-75b) Latent Space (https://www.latent.space/p/ainews-opus-55-is-good-at-explainer). The numbers are still moving, and the rivals are not waiting. OpenAI's Decisions API now returns typed answers in a single pass on GPT-6 Luna, Amazon has shipped a competing decision model, and Databricks built its ai_decide function to work directly with TypeSafe's interface, which makes the category look durable even if any one vendor's numbers are not Latent Space (https://www.latent.space/p/ainews-typesafejev-at-100m-arr-75b) TechCrunch AI (https://techcrunch.com/2026/10/06/how-ai-decision-models-could-change-content-moderation/) Databricks Blog (https://www.databricks.com/blog/introducing-aidecide-make-fast-decisions-your-governed-data). An engineer building a classification or routing layer now has at least three suppliers, so put it behind a thin adapter and keep switching cheap.

AI spending: the revenue race is real, the denominators are not

The two leading labs are now described in the same breath, and the comparison is the problem. Anthropic is reported near $65 billion annualized, while OpenAI lands anywhere from $70 billion down to $50 billion depending on how partner cuts are counted; the $70 billion version was calculated by OpenAI's own investors specifically to line up against Anthropic, and OpenAI itself has told investors its figure is approaching $50 billion AI Daily Brief (https://podcasters.spotify.com/pod/show/nlw/episodes/The-5-Debates-Shaping-AI-e3qav4d) TechCrunch AI (https://techcrunch.com/2026/10/08/openais-revenue-is-reportedly-20-billion-less-than-previously-projected/). Those revenue figures are the denominator the build-out has to earn back. On the spending side, the four biggest cloud owners, Alphabet, Amazon, Meta, and Microsoft, have guided to roughly $730 billion of 2026 capital spending between them, Oracle pushes the group to about $825 billion, and Moody's sees the biggest spenders closing in on a trillion dollars in 2027, so a $20 billion swing in one lab's revenue moves the case for the spending itself AI Daily Brief (https://podcasters.spotify.com/pod/show/nlw/episodes/The-5-Debates-Shaping-AI-e3qav4d). The argument over whether that outlay is justified has been heating for several days now. My own August call that OpenAI would post per-seat pricing for ChatGPT Business Premium by September 30 is sitting as a provisional miss, and the same opacity that hides a seat price hides how revenue gets counted, which is how a $20 billion gap stays open. A business leader signing a multi-year vendor commitment should treat any lab revenue figure as marketing until the method behind it is stated.

Voice agents: the executives say the moment has not come, so believe them

Voice AI vendors are telling you their own product is not there yet, and I would take that at face value. One executive, Shawn Wen, says the next challenge is speed: reasoning has to fetch answers fast enough that the conversation feels natural, and he predicts that once voice quality earns a first engagement, customers will over time grow confident enough to deal with an agent instead of a person. Trust is the other gap. Otter, for its part, plans to tell everyone in a chat that a meeting is being recorded, even when its bot is absent TechCrunch AI (https://techcrunch.com/2026/10/11/these-execs-think-voice-ai-hasnt-reached-its-chatgpt-moment-yet/). One voice-agent builder puts the retention target at about 800 milliseconds from a customer finishing a sentence to the agent replying, and voice has lagged chat because accents, noise, and interruptions break it AI Engineer (https://www.youtube.com/watch?v=IA4lZjh9sTs) Databricks Blog (https://www.databricks.com/blog/ai-customer-service). Those are hard limits, not polish problems. For a business leader eyeing a support line, pilot on low-stakes calls and measure latency and drop-off before any staffing change.

Augur calls

• Call: Sakana AI will publish MASS training code or trained model weights in a public repository by 2026-12-31.
• Criteria: Hit if a public GitHub or Hugging Face repository under a Sakana AI account containing MASS code or weights exists on or before the due date; miss otherwise.
• Due: 2026-12-31
• Odds: 55%
• Call: TypeSafe will publicly confirm a funding round with a stated post-money valuation of at least $7 billion by 2026-11-30.
• Criteria: Hit if a TypeSafe blog post, press release, or named-investor announcement states a post-money valuation at or above $7 billion on or before the due date; miss if no such statement appears or the stated figure is lower.
• Due: 2026-11-30
• Odds: 65%

What to watch

• Sakana AI's MASS page is where any externally graded benchmark or released weights would appear; until then the per-token gain is a company-reported claim from a self-judged training loop. Sakana AI (https://sakana.ai/mass/)
• TypeSafe's "Series AI" round needs a confirmed valuation and a revenue figure that is not a leak before the first-week $100 million means anything. Latent Space (https://www.latent.space/p/ainews-typesafejev-at-100m-arr-75b)
• OpenAI's Decisions API is now reachable from third-party tools, so developer cost comparisons against Jev should surface soon. Simon Willison's Weblog (https://simonwillison.net/2026/Oct/6/llm-openai-decisions/)
• Moody's trillion-dollar 2027 projection is the bar the next round of spending guidance from the four big cloud owners will be read against. AI Daily Brief (https://podcasters.spotify.com/pod/show/nlw/episodes/The-5-Debates-Shaping-AI-e3qav4d)
• Otter's recording disclosure, if rivals copy or ignore it, will show where trust norms for voice agents settle. TechCrunch AI (https://techcrunch.com/2026/10/11/these-execs-think-voice-ai-hasnt-reached-its-chatgpt-moment-yet/)

Verification notes

Sakana's 1.2 to 1.6 times figure is per output token, reported on four research benchmarks after training on synthetic tasks in which the model selected its own workflows, and nothing in the supplied evidence shows an outside evaluation Sakana AI (https://sakana.ai/mass/). TypeSafe's revenue and valuation numbers come from a leak and a headline rather than a filing Latent Space (https://www.latent.space/p/ainews-typesafejev-at-100m-arr-75b), and OpenAI's revenue varies by $20 billion depending on who is counting TechCrunch AI (https://techcrunch.com/2026/10/08/openais-revenue-is-reportedly-20-billion-less-than-previously-projected/).

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-10-12?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-10-12)