Elorian's Glimpse Lists at 6% of GPT-6 Astra's Price. Its Own Table Has It Losing 6 of 11

• Book: AI Agents Pocket Guide
• Also by me: Thinking in Go (2-book series) — Complete Guide to Go Programming + Hexagonal Architecture in Go
• My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
• Me: xgabriel.com | GitHub

Elorian's CEO, Andrew Dai, told TechCrunch he once photographed a board game he had just finished playing with friends and asked a frontier model who was winning. It couldn't say.

On October 16, his startup shipped its answer to that: Glimpse, a "visual thinking model" released as a research preview with an API and a chat UI. It lists at $0.60 per million input tokens and $3 per million output. Elorian's own comparison puts that at 94% below GPT-6 Astra's $10 and $50.

The same launch post carries a 12-row benchmark table. Against Astra, Glimpse wins 4 rows, ties 1, and loses 6. The twelfth row has no Astra score. If you build anything that reads screenshots, charts, product photos or video, the losing rows are the ones to read before you switch.

What shipped on October 16

Elorian is a Palo Alto startup that raised a $55 million seed round at a $300 million valuation before it had a product, with Nvidia and Google's Jeff Dean among the backers, according to TechCrunch. Dai is a former Google DeepMind researcher.

Glimpse is built on "an open weights base model" that the launch post doesn't name, post-trained by "a small team with only a few hundred GPUs." The post doesn't state a license or a context window.

You can call it two ways. Elorian's own platform offers an API and a chat UI (launch post). It is also listed on OpenRouter as elorian/glimpse, where the endpoint data shows one provider, Together, with a 262,144-token context, text, image and video input, text output, and cached input reads at $0.06 per million tokens.

On price, TechCrunch places it above lightweight models like Haiku and Luna and below mid-tier ones like Sonnet and Sol. Elorian's post also claims it comes in 20% under Gemini 3.8 Flash's introductory rates of $0.75 and $3.75.

The idea: go back to the image

Most multimodal models, as TechCrunch describes them, turn an image into tokens once and then reason about it in text. Glimpse returns to the source while it thinks. It can crop a region or pull a single frame from a video, and it can run code to measure what it sees. Elorian calls this a "visual scratch space."

The launch post says the model was trained end to end to use multimodal tools that write code, transform images and watch videos, and that it can spawn subagents that pass text, visual artifacts and internal representations between them. TechCrunch notes that OpenAI added a similar crop-and-zoom ability to its o-series models in 2025. Elorian's claim is that it built the behavior into the whole model.

That is an agent loop: look, call a tool, look again. It also means a single answer can take several internal passes over the same image, and every pass is billed.

Read the whole table

Here is Glimpse against GPT-6 Astra on every row of Elorian's table:

The same post adds caveats. Unless noted, every model ran at its highest thinking effort in its own native agent harness, and the OpenAI and Anthropic models had Code Interpreter. On the video rows, Sol and Opus 5.5 ran at high effort because Elorian had no max-effort results for them, and the HSS row includes results taken from papers rather than fresh runs. On CharXiv-R, Claude Opus 5.5 tops the row at 93.40, ahead of both. On Video-MME v2, GPT-6.1 Sol leads at 65.27.

Every number here is Elorian's. TechCrunch attributes the results to "Elorian's evaluations" and doesn't cite any independent testing. I haven't run Glimpse myself.

The fair reading: a small team's first model sits within a few points of a frontier model on most rows, and ahead on four, at a fraction of the list price. But if your workload looks like video question answering, scene understanding across images and video, or counting items in product photos, its own table says Astra still answers better. The two widest gaps, HSS and Video-MME v2, are both about 7 points.

Per-token price is not per-task price

A 94% lower token price only turns into a 94% lower bill if both models spend the same number of tokens. Glimpse's design spends tokens on purpose. Every crop, frame pull and code run is more reasoning, and Elorian says so in the launch post: careful review "takes time, and tokens," and "Token efficiency is our next highest focus."

The post does give one per-task number: on a single LIBERO robot-planning task, an episode cost $0.035 on Glimpse against $0.249 on GPT-6 Astra. The vendor picked that task, and one episode says little about a workload average. Your own images and effort setting decide the real figure.

Test it on your own images first

Because OpenRouter speaks the chat completions format, a first call is small. Image inputs go in a content array next to the text, per OpenRouter's image input docs:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "elorian/glimpse",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text",
         "text": "How many pallets are in this photo?"},
        {"type": "image_url",
         "image_url": {"url": "https://example.com/yard.jpg"}}
      ]
    }]
  }'

Before you route real traffic to it, run a small comparison:

1. Collect 30 to 50 images or clips from your own product, each with a known correct answer.
2. Run them through Glimpse and through whatever model you use today, at the reasoning effort you would actually ship. Elorian's published scores used each model's highest effort on most rows.
3. Record accuracy, the usage token counts from every response, and wall-clock time.
4. Divide total cost by correct answers. That is the price you compare, not the list price.

Treat it as what Elorian calls it: a research preview. The launch post names robot control and calibrated reconstruction as work still ahead, and it states no license. Keep a fallback model behind it until your own numbers say otherwise.

Would you move an image pipeline to a first model from a seed-stage lab if your own eval came out ahead? Tell me in the comments.

If this was useful

I build Hermes IDE, a free, source-available IDE that runs Claude Code, Codex and other coding agents side by side on macOS, Windows and Linux: download it or star it on GitHub.

Glimpse's trick is an agent loop folded into a model, billed per pass. If you build the outside version of that loop yourself, with tool calls, subagents and limits on how long it runs, the AI Agents Pocket Guide covers the patterns and the guardrails that keep the token bill readable.