Talk Around Town: A Shanghai Startup Claims Its 9B Robot Brain Whipped GPT-6 Astra — On Its Own Scorecard

What the Chatter Says

Word coming out of Shanghai is that a fresh-off-the-farm startup called Daxiao Robotics is hollering from the barn roof that its new model just out-pointed OpenAI's heavyweight. According to Daxiao Robotics' own announcement — picked up by Chinese tech outlets IT Home, Jiemian, iFanr, Sohu, and CSDN, and then by English-language trade sites Gasgoo and Niux.si — the company's ACE-Brain-0.6 system bested GPT-6 Astra on six out of eight narrow visual-spatial grounding evaluations. Every one of those numbers traces back to Daxiao's own disclosure; not a single external evaluator has touched this thing with a ten-foot pole.

Daxiao Robotics, which the company describes as a Shanghai-based embodied-AI startup founded in 2025, says ACE-Brain-0.6 is what it calls a visual-spatial anchoring embodied foundation model — fancy barn-door language meaning it is built to help robots understand where stuff is in three-dimensional space. According to the company, the model runs on nine billion parameters and sits on top of Qwen3.5-9B as its base, with the visual encoder left frozen while supervised fine-tuning was applied to the language model parameters. That is the company's own description, not a finding from any neutral party.

What Is Actually Confirmed

Let us separate the dry corn from the wet corn here. A handful of things hold up just fine. GPT-6 Astra is a genuine, publicly available OpenAI model — confirmed by Wikipedia, llm-stats.com, emergent.sh, and codersera.com. It dropped on September 3, 2026, carries a price tag of $10 per million input tokens and $50 per million output tokens, and offers a 1.05-million-token context window. That part is solid ground.

Daxiao Robotics is also a real company, confirmed through Preqin and the China Embodied AI Industry Guide — not some ghost corporation somebody conjured up behind a kudzu patch. And that angel funding round is documented: Preqin confirms that in February 2026, Ant Group led the round, with Qiming Venture Partners, Golden Vision Capital, Hony Capital, and Lenovo all joining in. Serious backers, no question.

What Nobody Has Been Able to Verify

Here is where the mud gets thick. Every single benchmark figure in this story comes from Daxiao Robotics' own published evaluation, as relayed by Gasgoo AutoNews, which itself is essentially a rewrite of the company's release. Niux.si picked up the claim by way of IT Home, adding another layer of telephone-game distance without adding any corroboration. None of these outlets ran the evaluations themselves. No external lab, no academic paper, no third-party scorer has independently reproduced what Daxiao is claiming.

According to Daxiao Robotics, ACE-Brain-0.6 landed at an average of 74.53 across seven publicly recognized benchmarks, compared with GPT-6 Astra's 73.68 — a gap roughly as wide as a coat of paint on a barn door. The company has not published statistical confidence intervals, which means there is no way to tell whether that difference would hold up across repeated runs or simply dissolve like morning fog.

The most attention-grabbing number, according to the company's own results, is a score of 94.13 versus GPT-6 Astra's 91.10 on something Daxiao calls RobotGrounding — a 2,591-question proprietary evaluation covering 14 task categories. The problem, as Daxiao Robotics itself is the only party with access to that test set, is that outside researchers flat-out cannot audit or re-run it. Grading your own homework and then announcing you got an A is a time-honored tradition, but it ain't peer review.

The company also says it trained ACE-Brain-0.6 on roughly one million spatial-grounding examples gathered from more than 30 open-source robotics datasets plus proprietary data, and claims that every doubling of training-data volume pushed the average benchmark score up by about 2.4 points. That scaling observation, too, comes solely from Daxiao's own account and has not been verified externally.

Why the Optics Are Murkier Than a Catfish Pond

Even setting aside the proprietary test, that razor-thin public-benchmark margin — 74.53 against 73.68 — sits well within the kind of run-to-run variance that benchmark experts routinely warn about. Independent trackers such as Artificial Analysis and LLM Stats measure GPT-6 Astra across broad general-intelligence evaluations where it performs strongly; ACE-Brain-0.6's claimed edge covers only a narrow slice of specialized spatial-grounding tasks and has not been mapped to any of those general-purpose indexes.

The cluster of Chinese media coverage may look like a crowd of witnesses, but every one of those outlets is drinking from the same Daxiao-issued bucket. IT Home, Jiemian, iFanr, Sohu, and CSDN are all recycling the company's numbers without independent testing, which makes the apparent corroboration about as genuine as a rubber worm on a fishhook. The overall source assessment shows a single originating channel with only one independent conduit — that is not a deep well of evidence.

The Publication's Analysis

This is analysis, not reporting: The broader pattern here may be more interesting than the headline score. If a nine-billion-parameter open-source base model can — even by its own reckoning — hang competitively with a top-tier closed model on domain-specific tasks after targeted fine-tuning, that suggests competition in embodied AI may increasingly hinge on data curation and task focus rather than raw model scale. That is a meaningful strategic observation regardless of whether Daxiao's exact numbers hold up under outside scrutiny.

That said, the structure of this announcement checks basically every box on the 'handle with care' list: self-reported scores, a proprietary benchmark nobody else can run, a margin well inside statistical noise on the public tests, and a media echo chamber dressed up as independent confirmation. Until an external lab or academic group gets its hands on ACE-Brain-0.6 and runs these evaluations without Daxiao holding the stopwatch, the claim is best understood as an interesting data point from a well-funded young company playing on its home field — not as settled evidence that a nine-billion-parameter fine-tune has genuinely overtaken OpenAI's flagship at anything beyond a very specific barnyard chore.