Elorian's Glimpse Costs 94% Less Than GPT-6 Astra and Trails the Best Rival on 7 of 12 Tests
A.I. / news
Elorian's Glimpse Costs 94% Less Than GPT-6 Astra and Trails the Best Rival on 7 of 12 Tests
The $55 million seed-funded start-up's first model leads on four vendor-run visual benchmarks, ties one and loses seven, with no independent results yet.
Elorian, a Palo Alto start-up led by former Google DeepMind researcher Andrew Dai, released its first model, Glimpse, on Friday, Oct. 16, at $0.60 per million input tokens and $3 per million output tokens.
Glimpse is a research preview, available through an API and a chat interface, according to Elorian's launch post. TechCrunch reported the release the same day.
What Elorian says Glimpse does
Most multimodal models look at an image once, turn it into tokens and reason about it in text. Elorian says Glimpse can return to the source while it thinks. It can crop a region, pull specific video frames or run code to measure what it sees. The launch post calls this reasoning "in visual space".
The post says the model was "post-trained on an open weights base model" by a small team using a few hundred GPUs. The pages reviewed do not name the base model.
Dai said, as TechCrunch quotes him, that the idea came from a board game night, when a frontier model could not say who was winning from a photo of a finished game.
Twelve benchmarks, four clear leads
Elorian published results on 12 tests against GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5 and Gemini 3.8 Flash. Every figure below is vendor-supplied. The post says all models ran at their highest thinking effort in their own agent harness.
| Benchmark | Glimpse | Best comparator |
|---|---|---|
| WGO Bench (manipulation events) | 42.70 | 39.42, GPT-6 Astra |
| FSC-147 (counting) | 83.36 | 82.69, GPT-6.1 Sol |
| UGC-VideoCap (short video captions) | 82.66 | 80.41, GPT-6.1 Sol |
| CharXiv-R (scientific charts) | 92.90 | 93.40, Claude Opus 5.5 |
| BabyVision (fine visual reasoning) | 95.36 | 97.68, GPT-6 Astra |
| TOMATO (temporal reasoning) | 93.60 | 96.02, GPT-6 Astra |
| Video-MME v2 with subtitles | 56.68 | 65.27, GPT-6.1 Sol |
Elorian says Glimpse beats Astra on four evaluations and ties it on one, ERQA, at 83.75. By the same table it trails the best comparator on seven of the 12. The widest gap is Video-MME v2, nearly nine points behind Sol.
TechCrunch's summary says Glimpse outperforms Astra on scientific chart reasoning. Elorian's table names Opus 5.5 as the strongest rival on CharXiv-R, 0.5 points ahead of Glimpse.
The post says some maximum-effort video results were unavailable for certain models. It also says several demos are curated and some use reviewed reference answers rather than recorded model traces. No independent party has run the benchmarks.
Pricing against four rivals
The launch post lists uncached output prices per million tokens. Elorian puts Glimpse 94% below Astra and 85% below Opus 5.5.
- GPT-6 Astra50 USD
- Claude Opus 5.520 USD
- GPT-6.1 Sol10 USD
- Gemini 3.8 Flash (intro rate)3.75 USD
- Glimpse3 USD
Source: Elorian, Introducing Glimpse, accessed 2026-10-17
A per-token price is not a per-task price. Elorian says dense visual review costs more tokens and names token efficiency as its "next highest focus". In its robot-planning demo the company puts the cost per episode at $0.035 for Glimpse and $0.249 for Astra.
The $55 million behind it
Elorian announced a $55 million seed round on April 12, according to Pulse 2, led by Striker Venture Partners, Menlo Ventures and Altimeter, with 49 Palms and researchers including Jeff Dean taking part. TechCrunch adds Nvidia as a participant and puts the valuation at $300 million. Pulse 2's April report did not disclose a valuation.
Dai said, in TechCrunch's report, that the company wants next year to build "the world's best visual reasoning model". Elorian said calibrated scene reconstruction and direct robot action output remain future work. It did not give a date for either.
For related coverage of benchmark claims, see Mercor's map of work hours no benchmark covers and Reflection's open-weights release.
Sources
More in A.I.
- 01Vals Test: Three of Six AI Agents Threatened Court Over a £350 Train RefundTold to try everything to recover a refund the rules no longer allowed, Grok 4.7 sent 187 messages in 15 simulated days while Opus 5.5 sent eight in twelve weeks.
- 02MiniCPM-V 4.7 Is Apache 2.0 With 3B Active Parameters and No Benchmark TablesOpenBMB's first mixture-of-experts vision model claims a win over Qwen3.6 35B-A3B in prose and charts, and 13 tokens a second on an iPhone, but the model card prints no scores.
- 03Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own TableGoogle's October 13 launch post leads with a Vals Index ranking; the comparison table behind it shows Argon behind GPT-6 Astra or Opus 5.5 on both Terminal-Bench tests and on OSWorld 2.0.
- 04StarSkirmish Has Nine Models Write StarCraft Bots in One Hour, and the Best Wins 7% Against Top Human BotsGemini 4 Argon lands in a three-way tie at 37 to 39, but it ran at a lower reasoning setting than most rivals and its runs vary the most.