Elorian's Glimpse Tells AI: Look Twice — at 94% Off Astra

By TechPulse USA Newsroom — October 18, 2026

Every other AI lab is teaching its models to read images faster. Elorian, a Palo Alto startup that raised $55 million before it had a product to sell, wants them to slow down — to look twice. Its first model, Glimpse, launched on Friday as a research preview, and the pitch is almost embarrassingly simple: humans glance, then look again. Machines should too.

The idea has a birthday and a board game attached. Last year Andrew Dai — Elorian's chief executive and a former Google DeepMind researcher — photographed a board game he had just finished with friends and asked a frontier model who was winning. It couldn't say. "Board games 'each have their own visual language, each have their own iconography, symbols,' he told TechCrunch. 'To an adult, it's quite easy to understand what's going on with minimal text, but these models aren't able to intuit what the symbols mean.'"

That is a lovely origin story, and it contains the entire thesis. Today's vision models mostly tokenise an image once — a single frozen glance — and then do all of their reasoning from that memory. Glimpse instead gets what Elorian calls a "visual scratch space": mid-reasoning, it can crop into a region of an image, pull denser frames from a slice of video, or run code to measure and compare what it is seeing. The launch post's own slogan: "A glimpse is where looking starts. Understanding begins when a model knows what deserves a second look."

There is also a stranger capability that most of Friday's coverage filed away as a curiosity: Glimpse can "spawn and coordinate subagents that exchange text, visual artifacts, and internal representations to solve a shared problem." Don't skip past that sentence. It is the clearest tell in the whole launch about where Elorian is actually heading — from seeing, to doing. Dai has said the longer-term goal is models that build a model of a scene and then choose actions in it, including directly controlling robots. The target customers say the same thing: companies sitting on mountains of video — robotics firms, scientific labs, media and creative companies.

The idea, to be fair, is not entirely new, and nobody at Elorian is pretending otherwise. OpenAI added crop-and-zoom during reasoning to its o-series models last year. Elorian's argument is that nobody has built the entire model around the idea, and nobody has done it on the cheap. Which brings us to the line from the TechCrunch exclusive that is going to travel the farthest: "Humans can look at Ikea diagrams, with no text, and assemble Ikea furniture, right? Why can't AI models?"

A cosmic magnifying glass: Hubble's image of galaxy cluster MACS J0454.1-0300, whose gravity bends the light of distant galaxies. Elorian's thesis is that AI needs to learn the same trick — to look twice. Editorial illustration only; this is not Elorian or Glimpse. (Photo: ESA/Hubble & NASA, CC BY 4.0)

The 60-cent knife fight

The benchmarks are fine. The pricing is the story.

Glimpse costs $0.60 per million input tokens and $3.00 per million output tokens. In Elorian's own comparison table — the company's arithmetic, not an independent audit — that works out to 94% lower than GPT-6 Astra, 85% lower than Claude Opus 5.5, 70% lower than GPT-6.1 Sol and 20% lower than Gemini 3.8 Flash's introductory rates. TechCrunch's independent framing is more sober: Glimpse is cheaper than the mid-tier models like Sonnet and Sol, but still pricier than the lightweight ones like Haiku and Luna. Both things can be true, and both are true.

Here is Elorian's own price table, so you can see the arithmetic for yourself:

Model	Input / 1M tokens	Output / 1M tokens
Glimpse	$0.60	$3.00
GPT-6 Astra	$10.00	$50.00
Claude Opus 5.5	$4.00	$20.00
GPT-6.1 Sol	$2.00	$10.00
Gemini 3.8 Flash	$0.75	$3.75

Notice where the knife is aimed. Elorian is not pricing against the cheap models; it is pricing against the frontier flagships and telling enterprise buyers the expensive era is optional. And here is the part of the story that genuinely deserves a second look: Elorian says Glimpse was built by a small team on "only a few hundred GPUs," post-trained on an open-weights base model, and is "substantially smaller than the frontier systems it's benchmarked against."

That sentence is, to my mind, the real news of the launch — and it is the bit the launch-day coverage mostly walked past. If a startup with a few hundred GPUs can ship a visual-reasoning model that trades blows with systems trained on frontier-scale compute, then the economics of this industry are not what the capex headlines say they are. The whole point of the frontier-lab spending race — the dynamic we covered when Anthropic's 45% growth still wasn't enough — is that you cannot win without a war chest. Elorian is arguing, on the record, that in visual reasoning you nearly can.

It also reframes the $55 million seed round at a $300 million valuation — raised in April, before the company had launched anything, co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean. In this market, valuations like that are the bubble debate in miniature. Elorian's answer to the sceptics is Glimpse itself: a shipped product, at a price that embarrasses the flagships, built on a fraction of their budget.

One honest admission in the launch post deserves quoting, because it tells you where the bodies are buried: "Looking that closely takes time, and tokens. Token efficiency is our next highest focus." Re-examining an image mid-reasoning is not free — every second look is more tokens, more latency, more cost. And "direct action generation remains future work," which is the company's own way of saying the robot-control pitch is still a promise, not a product.

The benchmark table nobody else printed

Most of Friday's coverage said the same thing: Glimpse beats GPT-6 Astra on four benchmarks, including chart reasoning. True, as far as it goes. But nobody printed the whole table — Elorian's own Table 1 — so here it is, all twelve rows. Fine print: Elorian's evaluations, Elorian's harness, "highest thinking effort." Company-reported, not independent; read every claim below with that caveat.

Area	Eval	Glimpse	Astra	Sol	Opus	Gem 3.8
Robotics	WGO Bench	42.70	39.42	38.91	12.34	33.99
Robotics	ERQA	83.75	83.75	82.45	82.20	81.20
Vision	BabyVision	95.36	97.68	97.16	—	83.25
Vision	HSS	46.36	53.6	46.6	49.81	41.6
Science	CharXiv-R	92.90	92.10	92.60	93.40	91.80
Video caps	UGC-VideoCap	82.66	80.32	80.41	76.35	78.33
Video caps	DREAM-1K	76.04	79.05	77.93	72.05	70.00
Temporal	TOMATO	93.60	96.02	89.02	72.44	83.02
Temporal	Video-MME v2	56.68	64.06	65.27	—	59.58
Temporal	E.T. Bench	63.55	—	59.79	—	63.45
Counting	FSC-147	83.36	82.02	82.69	73.78	67.06
Counting	CountQA	82.66	86.42	86.42	82.20	81.09

Now look at what the coverage left out. First, the chart-reasoning win: Glimpse scores 92.90 against Astra's 92.10 — a margin of 0.8 points, which is thin — but Claude Opus 5.5 scores 93.40, ahead of both. "Beats GPT-6 Astra on charts" is true. "Best chart reader" is not. We re-read Elorian's own Table 1 row by row against the launch coverage and found that lead sitting there, unmentioned, in Elorian's own numbers.

Second, the table is not all wins. Astra dominates scene understanding (HSS: 53.6 to Glimpse's 46.36 — the largest gap in the table), leads on temporal reasoning (TOMATO, Video-MME v2), and ties on embodied reasoning. And one of the "four wins" comes with a dash: Astra wasn't even evaluated on E.T. Bench, so there is no comparison to beat.

Third — the discrepancy nobody has cleared up. TechCrunch says the chat is waitlist-only; Elorian's launch post says the model is "available through API and our chat UI today." FourWeekMBA caught it first; as of writing neither side has walked it back. Treat the chat as waitlist-only until Elorian says otherwise.

And the judgment call: the launch-day coverage was shoddy on the numbers. Cherry-picking two wins from a company-supplied table and filing them as "beats GPT-6" is stenography, not journalism — especially when the same table hands you the losses. Elorian deserves credit for publishing the full table at all; most labs would have cropped it. Glimpse is genuinely interesting — the scratch space is a real idea, the price a real threat — and none of that needed the cherry-picking. The cherry-picking is what makes me cautious.

Dai wants "the world's best visual reasoning model" next year — a promise the current table shows is some distance away. The robotics play is still, by the company's own admission, future work. And the price war has started whether or not the benchmarks hold: at $0.60 in and $3.00 out, Glimpse doesn't need to beat Astra to hurt it. It just needs to be good enough. OpenAI keeps expanding the battlefield elsewhere — but in visual reasoning, the battlefield just got cheaper.

Sources: Aditya Mehta's TechCrunch exclusive (Oct 16, 2026); Elorian's launch post "Introducing Glimpse" (elorian.ai) with the company's full benchmark and price tables; FourWeekMBA; Dealroom; Startup Fortune; Pondero.ai; RuntimeWire; NewsCord; Pulse 2 (April seed round).

TechPulse USA's AI desk covers the labs, models and money shaping artificial intelligence.