Every score traces back to a scorer receipt in a hash-chained record, and any result that fails a check is refused rather than reported. On ASI-Bench's 60 research tasks, the Swedish company's engine finished ahead of all 18 state-of-the-art AI model and agent combinations in the benchmark paper at every guidance level, on an ordinary laptop's CPU, for an estimated $0.20 and with no model training. Unizen AI will next adapt the framework to client cases in financial risk, insurance and energy.
Key facts about Unizen AI's ASI-Bench result
- Verification: every score traces back through per-task scorer receipts; scoreboard versions are chained with SHA-256 hashes, the checkers' rules are machine-checked, any result that fails a check is refused rather than reported, and separate re-runs reproduced the results. A public index of receipt hashes is at [verification index link].
- Score (local, non-official): 79.86 out of 100 across all 60 ASI-Bench research tasks, and between 79.77 and 79.92 at each of the four guidance levels.
- Comparison (local scoring): above the best score in Table 2 of the benchmark paper at every guidance level. At B3 it scores 79.92, 28.32 points above the best published result (51.60), from a frontier model at its "ultra" reasoning setting, and 3.00 times the paper's average (26.62); at B1, 79.87 against a best of 72.29.
- Guidance-invariant by design: from the full method (B1) to only the objective and data (B3), the engine's score moves +0.04, while the average across the paper's 18 configurations falls 24.29 points. The engine does not read the guidance prompts, so the flat profile is a property of its architecture.
- Cost: an estimated $0.20 for the engine to run the entire benchmark locally on an ordinary laptop CPU, covering electricity and hardware wear, against about $1,550 per run for the top-scoring combination in the benchmark paper.
- No model training: the engine required none. Epoch AI estimates that training a frontier model released in 2025 cost about $490 million in compute, excluding staff costs, and projects that, if current trends continue, the largest training runs will cost more than $1 billion by 2027.
- Hardware: the engine ran on the CPU of an ordinary laptop, with no data centre or cloud computing.
- Method: no language model at solve time; each task is solved with programmatically tuned analytic and first-principles models built for that specific task.
- Benchmark: ASI-Bench (arXiv:2608.17271), designed to test the research capabilities its authors say artificial superintelligence requires: 60 project-level research tasks across 11 scientific fields.
- Next step: client business cases in financial risk, insurance and energy, each tied to a measurable outcome.
STOCKHOLM, Sweden - 4 November 2026 -Unizen AI, the Swedish company that builds AI for measurable business performance, today published an ASI-Bench result in which every score traces back to a scorer receipt, every version of the scoreboard is hash-chained, and any result that fails a check is refused rather than reported. Its engine scored 79.86 out of 100 on ASI-Bench, ahead of all 18 state-of-the-art AI model and agent combinations tested in the benchmark paper, including a frontier model at its "ultra" reasoning setting, at every guidance level under local scoring, and ran the entire benchmark on an ordinary laptop processor for an estimated $0.20, with no model training. The engine solves each task with analytic and first-principles models built for that task and uses no language model at solve time. Researchers from Tsinghua University, MIT, Harvard, Carnegie Mellon University, the University of Michigan, Boston University and other institutions built ASI-Bench to test the research capabilities they say artificial superintelligence requires.
The result is a local, non-official measurement made with the benchmark's public scorer on its public instance set. It is not an official leaderboard entry. In plain terms, Unizen AI ran the benchmark's public scoring software itself; the benchmark's organisers have not scored, checked or ranked this result, which is why the rules require it to be labelled non-official.
How every Unizen AI result can be verified
Every reported score traces back through per-task scorer receipts to an executed computation, apart from two judge-assessed points on one task that were counted as zero during a judge outage. Successive versions of the scoreboard are chained with SHA-256 hashes, the checkers' rules are machine-checked with an automated solver, each result must agree across two separate reference checks, and any result that fails a check is refused rather than reported. Separate re-runs reproduced the results. A public index listing the SHA-256 hash of every per-task scorer receipt is at [verification index link], so anyone shown a receipt can confirm it is the one that was scored; the full receipt chain is available for audit on request.
"Anyone can publish a number; we build so that every result can be checked line by line," said Alexander Stojcevski, CEO of Unizen AI. "Every score comes with its receipt, every version of the scoreboard is hash-chained, and a result that fails a check is refused rather than reported. For banks, insurers and grid operators, that audit trail is the product. And because we build each method in advance, drawing on the full spectrum of frontier AI rather than training one large model, the whole benchmark ran on a laptop for twenty cents."
"We publish reproducible public-contract measurements and label them as such," he added. "When we cannot measure something, we say so rather than substitute a proxy number. An official score on the benchmark's private seed42 set would require submitting our engine's outputs to the benchmark operator, which we decline on IP grounds."
About ASI-Bench: "At the Dawn of Artificial Superintelligence"
ASI-Bench was introduced in August 2026 in the paper "ASI-Bench: At the Dawn of Artificial Superintelligence" (arXiv:2608.17271). Its authors describe it as "the first benchmark to jointly evaluate AI systems' capabilities of innovative exploration and autonomous scientific execution across general research domains." They write that artificial superintelligence requires AI to move beyond mastering existing knowledge toward "exploring the unknown, creating new knowledge, and turning new ideas into verifiable results." More than 40 experts built its 60 research tasks across 11 fields, from physics and chemistry to robotics, medicine and biostatistics, in more than 31,000 hours of work. Each task is posed at four levels of guidance: the full method (B1), the method's name only (B2), only the objective and data (B3), and B3 with distracting information (B4).
Unizen AI's ASI-Bench scores against the best published results
Under local scoring, Unizen AI's engine exceeds the best score in Table 2 of the benchmark paper at every guidance level. The paper's 18 state-of-the-art configurations pair frontier models with agentic frameworks; the strongest is a frontier model at its "ultra" reasoning setting:
- B1, full method given: 79.87, against a best of 72.29 (7.58 points ahead)
- B2, method name only: 79.87, against a best of 49.57 (30.30 points ahead)
- B3, objective and data only: 79.92, against a best of 51.60 (28.32 points ahead)
- B4, B3 with distracting information: 79.77, against a best of 50.41 (29.36 points ahead)
- Overall: 79.86, against a best of 55.84 (24.02 points ahead)
At B3, the engine's 79.92 is 3.00 times the paper's average of 26.62, and from B1 to B3 its score moves +0.04 while the paper's average falls 24.29 points. Instead of training a model on data, Unizen AI works out each task's method in advance and writes it as compact, programmatically tuned analytic and first-principles code, such as physics solvers, exact rational algebra and digital-twin-style pipelines, which runs directly on a laptop's CPU; a one-page technical backgrounder at [backgrounder link] walks through one task. Because the engine does not read the guidance prompts, its flat profile across the four levels is a structural property of the architecture, not a claim that the harder levels are easy.
How to read the comparison with the ASI-Bench paper
Across its 18 state-of-the-art configurations, the paper reports an average of 33.41 overall, falling from 50.91 at B1 to 26.62 at B3, and average token use of 4.35 million to 6.91 million per task, depending on the level. The comparison is with published results, not an official ranking: the paper's figures come from the authors' evaluations of general-purpose frontier models in agentic frameworks attempting each task at test time, while Unizen AI's engine was built for each task in advance and scored locally on the public instance set, whose reference answers are published. The paper does not say which instance set its own scores use, so the instances may differ.
Cost of running ASI-Bench: an estimated $0.20 on a laptop against about $1,550 per run, with no model training
Unizen AI's engine ran the entire benchmark locally on the CPU of an ordinary laptop, with no data centre or cloud computing. The cost, an estimated $0.20 and well under $1 even on conservative assumptions, covers the laptop's electricity and hardware wear; the engine itself incurs no per-token charges. Scoring uses the benchmark's public scorer, some of whose checks call an LLM or VLM judge. The paper reports that its three top-scoring model and agent combinations cost between about $684 and $2,728 per run, with the strongest, a frontier model at its "ultra" reasoning setting, at about $1,550. Training frontier models costs far more: the research organisation Epoch AI estimates that training a frontier model released in 2025 cost about $490 million in compute, excluding staff costs, and projects that, if current trends continue, the largest training runs will cost more than $1 billion by 2027. Unizen AI's engine required no model training at all.
What the result means for banks, insurers and energy companies
The properties behind the score matter as much as the score itself. Reproducible results, a computation behind every number and a compute bill measured in cents on ordinary hardware are what model-risk, audit and finance teams look for before AI touches pricing, risk or grid decisions. Unizen AI designs its systems to be safe and controlled, with sovereign deployment that keeps data, models and infrastructure under the client's control.
Next steps: from benchmark results to client business cases
Unizen AI will now develop the framework further and adapt it to clients' business cases where precision and auditability matter most, from financial risk and insurance to energy. Each engagement starts from a measurable outcome: lower cost, stronger margins or lower risk.
"We love to push the boundaries of what AI can do today, but we are a business-value company first," said Stojcevski. "Benchmarks are where we test the science; our customers' results are where we prove the value. We will build only where we can measure the return."
Frequently asked questions about Unizen AI's ASI-Bench result
- Does Unizen AI's ASI-Bench result mean it has achieved artificial superintelligence? No. ASI-Bench tests research capabilities that its authors say artificial superintelligence requires. Unizen AI's result is a measurement on the benchmark's 60 tasks, not a claim of superintelligence.
- Is Unizen AI's ASI-Bench score an official leaderboard result? No. It is a local, non-official measurement made with ASI-Bench's public scorer on the public seed31415 instance set, under the benchmark's local-evaluation rules. An official score requires submitting the engine's outputs on the private seed42 set for the maintainers to score, which Unizen AI has chosen not to do, to protect its intellectual property. Results on the public set cannot be submitted to the ASI-Bench leaderboard.
- How does Unizen AI's result compare with state-of-the-art AI models and agentic frameworks on ASI-Bench? Under local scoring, it is above the best score in Table 2 of the benchmark paper at every guidance level; the strongest configuration there is a frontier model at its "ultra" reasoning setting running in an agentic framework. The paper's 18 combinations were evaluated by its authors at test time, while Unizen AI's engine was built for each task in advance, so the comparison is with published results, not an official ranking.
- What does "LLM-free" mean for Unizen AI's engine? The engine uses no language model at solve time. It solves each ASI-Bench task with compact, programmatically tuned analytic and first-principles models built for that specific task, drawing on physics solvers, exact rational algebra and digital-twin-style pipelines.
- What did it cost Unizen AI to run ASI-Bench? An estimated $0.20 for the engine's run of the entire benchmark on the CPU of an ordinary laptop, and well under $1 even on conservative assumptions. The estimate covers electricity and hardware wear; the engine needs no per-token charges, no cloud computing and no model training. Scoring uses the benchmark's public scorer, some of whose checks call an LLM or VLM judge. The benchmark paper reports about $684 to $2,728 per run for its three top-scoring model and agent combinations.
- Did Unizen AI train a model to run ASI-Bench? No. The engine required no model training. For comparison, Epoch AI estimates that training a frontier model released in 2025 cost about $490 million in compute, excluding staff costs, and projects that, if current trends continue, the largest training runs will cost more than $1 billion by 2027.
About the measurement
Local, non-official measurement. Scored with ASI-Bench's official public scorer on the public seed31415 instance set under the benchmark's local-evaluation rules, which require local reports to be marked non-official. Not an official leaderboard entry (IP policy; the score-submission route is closed by choice). Comparisons to published values compare our locally-scored results against the paper's own reported Table 2/abstract numbers.
The scorer labels this run official=False and score_scope=public_local, on the reproducible public seed31415 instance set.
Eight of the 60 tasks have documented limits where fresh attempts did not improve on earlier solutions; they are reported as measured. Every solution candidate is checked for ground-truth contamination before scoring; any that shows it is refused and never enters any reported figure. For one task, the benchmark's judging service was unavailable for two judge-assessed points. The scorer treats such points as not scored; Unizen AI counted them as zero instead, which can only lower the reported figures.
ASI-Bench's authors and maintainers were not involved in this evaluation and have not reviewed or endorsed the result.
References: ASI-Bench paper (arXiv:2608.17271) · ASI-Bench leaderboard · Epoch AI on frontier training costs · Epoch AI training-cost estimate, 2025
About Unizen AI
Unizen AI is a Swedish, business-oriented frontier-technology company that turns advanced AI into measurable business results. Going beyond large language models and traditional machine learning, it builds cognitive systems and is developing quantum-aware platforms, all designed to be safe and to deliver lower cost, stronger margins and better control, with sovereign deployment that gives clients full control over data, models and infrastructure. Learn more at unizenai.ai.
Connect with founder and CEO Alexander Stojcevski on LinkedIn.
