ModelsArchitectures & capability
Elorian launches Glimpse, a visual reasoning model priced far below GPT-6 Astra
Palo Alto startup Elorian released Glimpse as a research preview via API and chat. The model re-inspects images and video mid-reasoning by cropping, sampling frames or running code, and charges $0.60 per million input tokens and $3 per million output tokens.
From the TechCrunch AI article of October 16, 2026 and cited sources: Glimpse goes back to the image or video while it reasons. [1] [2] [3]
SOURCES[1] TechCrunch: Elorian's first AI model wants to reason with… techcrunch.com[2] Elorian: Introducing Glimpse elorian.ai[3] OpenAI: Thinking with images openai.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- TechCrunch AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 16, 2026 article published
- Covered
- Oct 18 edition
- Evidence
- 8 sources across 7 sites
- 1TechCrunch: Elorian's first AI model wants to reason with…techcrunch.com
- 2Elorian: Introducing Glimpseelorian.ai
- 3OpenAI: Thinking with imagesopenai.com
- 4Pulse 2.0: Elorian AI raises $55 millionpulse2.com
- 5TechCrunch: How Elorian AI pulled off a $300M pre-seed…techcrunch.com
- 6RuntimeWire: Elorian launches Glimpseruntimewire.com
QuietHow it’s measured
Hacker News · 2 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Glimpse goes back to the image or video while it reasons. It can zoom into a region, replay a clip at a higher frame rate, follow one feature over time or run code to measure something. OpenAI's o3 introduced in-reasoning crop-and-zoom in 2025. Elorian's difference is that it built a whole model around the approach, post-training an unnamed open-weights base model on a few hundred GPUs. [1] [2] [3]
Benchmark results are mixed, and the 12-benchmark comparison comes from Elorian itself. Glimpse beats or ties GPT-6 Astra on 5 tasks, such as WGO Bench and FSC-147 counting, and trails it on 6. It is weakest on long-video temporal reasoning (Video-MME v2: 56.68 vs 64.06). On chart reasoning it edges Astra, but Opus 5.5 scores higher. [2] [8]
Per-token prices are about 94% below Astra's. That is where the widely shared '17x cheaper' figure comes from, but it does not measure cost per task. Each extra look at an image adds output tokens, and Elorian has named token efficiency as its next priority. [1] [2] [7]
Elorian is 19 months old and came out of stealth in April 2026 with a $55M seed round at a $300M valuation, before it had a product. It is targeting video-heavy customers such as robotics firms, labs and media companies. Its robotics benchmarks point to a stated ambition of moving into robot control, though Glimpse currently only perceives and reasons and does not take actions. [1] [4] [5]
The evidence shows a per-token price far below Astra's and mixed benchmark results.
Read the full assessment
The untested implication is cheaper high-volume video captioning, counting and chart extraction, if accuracy holds on real workloads and repeated image inspection doesn't erase the savings.
Executive brief
Elorian is a Palo Alto startup that is 19 months old. Its first model, Glimpse, prices input and output tokens 94% below OpenAI's GPT-6 Astra. On Elorian's own benchmark table, it beats or ties Astra on 5 of 12 visual tasks (Elorian). The model was post-trained from an open-weights base model on a few hundred GPUs. While it reasons, it can go back to an image or video to crop, pull frames or run code (TechCrunch). No independent evaluation has been published. In the company's own table, Glimpse loses to Astra on 6 of the 12 tasks.
What changed and event timeline
OpenAI ships "thinking with images."
o3 and o4-mini could crop, zoom and rotate images inside their chain of thought. This set the precedent Glimpse builds on ().
Elorian exits stealth with $55M
The seed round, led by Andrew Dai (formerly of Google Brain and DeepMind), was co-led by Striker, Menlo and Altimeter, with backing from 49 Palms and Jeff Dean (). It valued the company at $300M before it had a product ().
Glimpse launches as a research preview
It is available through an API and a chat interface. Pricing is $0.60 per million input tokens and $3 per million output tokens ().
Elorian publishes a 12-benchmark comparison
The table compares Glimpse with GPT-6 Astra, GPT-6.1 Sol, Opus 5.5 and Gemini 3.8 Flash. Wins, ties and losses are mixed ().
Capabilities and access
- Model: Glimpse, Elorian's first model, released as a research preview. Parameter count is not disclosed (Elorian).
- "Visual scratch space": mid-reasoning, the model can zoom into a region, play a clip at a denser frame rate, track one feature over time or run code to measure things (TechCrunch).
- Access: the API is open. One outlet reports the chat interface is waitlisted (RuntimeWire), while Elorian says both are available "today" (Elorian).
Read the full section
- Model: Glimpse, Elorian's first model, released as a research preview. Parameter count is not disclosed (Elorian).
- "Visual scratch space": mid-reasoning, the model can zoom into a region, play a clip at a denser frame rate, track one feature over time or run code to measure things (TechCrunch).
- Access: the API is open. One outlet reports the chat interface is waitlisted (RuntimeWire), while Elorian says both are available "today" (Elorian).
- Target customers: companies with lots of video, such as robotics firms, labs, media and creative businesses.
Technical analysis for researchers and developers
- Build: Glimpse is post-trained from an unnamed open-weights base model on "a few hundred GPUs" (Elorian). It is not a pretrained frontier model.
- Harness differences: every model ran at its highest thinking effort in its own native agent harness. Comparisons across models are therefore not like-for-like.
- Reproducibility: there are no weights, eval code or published prompts, and the base model is not named.
Read the full section
- Build: Glimpse is post-trained from an unnamed open-weights base model on "a few hundred GPUs" (Elorian). It is not a pretrained frontier model. Most of the gains likely come from post-training for tool use, though this is inference.
- Harness differences: every model ran at its highest thinking effort in its own native agent harness. Astra ran with Code Interpreter, and some video runs used "high" effort because "max" results were not available. Comparisons across models are therefore not like-for-like.
- Reproducibility: there are no weights, eval code or published prompts, and the base model is not named.
- Next focus: Elorian says token efficiency is its next priority. That matters because repeated looks at an image add output tokens, which affects real cost per task.
Claims and evidence
All numbers below are Elorian's own figures (Elorian): TechCrunch's line that Glimpse "outperforms even GPT-6 Astra" on chart reasoning is accurate only against Astra. Opus 5.5 scores higher on CharXiv-R (TechCrunch). A widely shared post describes the model as "17x cheaper" (Deedy on X). That matches the per-token price cut, not cost per task.
Read the full section
All numbers below are Elorian's own figures (Elorian):
| Benchmark | Glimpse | GPT-6 Astra | Best in table |
| WGO Bench | 42.70 | 39.42 | Glimpse |
| CharXiv-R | 92.90 | 92.10 | Opus 5.5 (93.40) |
| UGC-VideoCap | 82.66 | 80.32 | Glimpse |
| FSC-147 | 83.36 | 82.02 | Glimpse |
| Video-MME v2 | 56.68 | 64.06 | Sol (65.27) |
| HSS | 46.36 | 53.6 | Astra |
TechCrunch's line that Glimpse "outperforms even GPT-6 Astra" on chart reasoning is accurate only against Astra. Opus 5.5 scores higher on CharXiv-R (TechCrunch). A widely shared post describes the model as "17x cheaper" (Deedy on X). That matches the per-token price cut, not cost per task.
Context and prior work
- OpenAI's o3 already did crop-and-zoom inside its reasoning in 2025 (OpenAI). Elorian's pitch is that it built the whole model around this approach and did it cheaply.
- Elorian and Scale AI built a benchmark called "Humanity's Sixth Sense," which appears as HSS in the table. Glimpse scores below Astra on it.
Read the full section
- OpenAI's o3 already did crop-and-zoom inside its reasoning in 2025 (OpenAI). Elorian's pitch is that it built the whole model around this approach and did it cheaply.
- Elorian and Scale AI built a benchmark called "Humanity's Sixth Sense," which appears as HSS in the table. Its failure analysis found that most frontier-model failures were basic perception errors: missed details (21%) and misidentified objects or people (20%) (Startup Fortune). Glimpse scores below Astra on it.
Limitations, safety and contested findings
- Glimpse trails Astra on BabyVision, HSS, DREAM-1K, TOMATO, Video-MME v2 and CountQA (Elorian). It is weakest on long-video temporal reasoning.
- Startup Fortune calls Glimpse a narrow specialist that is "almost certainly" not a better general model than Astra (Startup Fortune).
- The investor lists conflict. TechCrunch names Nvidia as an investor; the April funding report lists 49 Palms and does not mention Nvidia (Pulse 2.0).
Read the full section
- Glimpse trails Astra on BabyVision, HSS, DREAM-1K, TOMATO, Video-MME v2 and CountQA (Elorian). It is weakest on long-video temporal reasoning.
- Startup Fortune calls Glimpse a narrow specialist that is "almost certainly" not a better general model than Astra (Startup Fortune).
- The investor lists conflict. TechCrunch names Nvidia as an investor; the April funding report lists 49 Palms and does not mention Nvidia (Pulse 2.0).
- The reviewed sources include no safety documentation or system card.
Business and practitioner implications
- Possible cost savings: high-volume video work like captioning, counting and chart extraction could get much cheaper, if the model's accuracy holds on your own data.
- Test before switching: benchmark cost per task, including the extra tokens from repeated image inspection, before migrating.
- Robotics: the robotics benchmark results (WGO, ERQA) and Dai's goal of moving into robot control point to where Elorian is headed.
Read the full section
- Possible cost savings: high-volume video work like captioning, counting and chart extraction could get much cheaper, if the model's accuracy holds on your own data.
- Test before switching: benchmark cost per task, including the extra tokens from repeated image inspection, before migrating. The startup's real test is whether a deep discount can overcome switching costs (Startup Fortune).
- Robotics: the robotics benchmark results (WGO, ERQA) and Dai's goal of moving into robot control point to where Elorian is headed. Today Glimpse only perceives and reasons; it does not take actions.