GPT-6.1 Terra beats Astra on Roboflow vision benchmarks — 47% faster and 92% cheaper

TL;DR

SkalskiP (Roboflow open-source lead) runs OpenAI’s just-released GPT-6.1 Terra against GPT-6 Astra on Roboflow’s playground vision benchmarks and reports that Terra wins across object detection (low- and high-effort), OCR (small-text, license plates, electricity meters, digital scales, handwritten cheques, posters, weather observations, pharmacy bills), and counting — while running ~47% faster and ~92% cheaper. OpenAI positioned Terra on 2026-10-26 as a cheaper, faster model targeted at computer-use workflows, not a frontier successor to Astra, which makes the vision-eval sweep across the Roboflow task mix a surprise.

Key claims

Terra beats Astra on object detection at both low and high reasoning effort on the Roboflow playground benchmark [Tweet 2/N, playground.roboflow.com/evals/object-detection].
Terra pulls furthest ahead of Astra on small-text OCR, with the gap narrowing on larger text [Tweet 8/N, playground.roboflow.com/evals/ocr].
Terra wins on JSON extraction from documents including handwritten cheques, weather observations, and pharmacy bills [Tweet 12-15/N].
Terra wins on find-text-and-read tasks (posters, license plates, electricity meters, digital scales) over Astra on the Roboflow OCR eval [Tweet 16-19/N, playground.roboflow.com/evals/ocr].
Terra runs ~47% faster and ~92% cheaper than Astra on the same evals [Tweet 21/N].

Method

Benchmark is Roboflow’s public playground eval suite (playground.roboflow.com/evals/object-detection, playground.roboflow.com/evals/ocr). Example categories span technical drawings, Lego bricks by shape, solar-panel hot spots, document layout, football players and referees, handwritten cheques, pharmacy bills, weather observations, license plates, electricity meters, digital scales, and posters. The tweet thread is image-first — the public artifact is a leaderboard screenshot plus side-by-side example outputs; no methodology paper, prompts, or raw accuracy numbers are published inline. Cost and latency numbers reference OpenAI API pricing against Astra on the same prompts.

Results

Headline: Terra beats Astra across the three advertised axes (text, object detection, counting) on Roboflow’s internal mix. The ~47% speedup and ~92% cost reduction are the quantitative numbers from the final tweet in the thread. No per-task accuracy deltas are quoted — the examples are selected wins. The announcement is from Roboflow (adjacent to OpenAI, not from OpenAI) and SkalskiP explicitly flags surprise (“wtf OpenAI? had to double-check the results”), signaling the result was unexpected given Terra’s positioning as a cheaper computer-use model rather than a flagship vision model.

Why it’s interesting

Live counter-example to the pattern VLM Perception Failures documents — small-text OCR, counting, and precise visual grounding are the exact failure modes the concept page tracks across frontier models, and here a cheaper/faster model within the same family (Terra vs. Astra) sweeps them. Mirrors the pattern Gemini 3.7 Flash saturates Robocurve physical tool-use benchmark at 92% vs 32% for 3.6 Flash (three weeks apart) documented three weeks after a flagship: a cheaper general-purpose refresh (Gemini 3.7 Flash) dominated a domain-specific flagship (Gemini Robotics-ER 2) on the same physical-task harness. Both suggest the “flagship always wins on vision” assumption is breaking in a few-weeks cadence. Note the eval is Roboflow’s own playground on vendor-selected example tasks (not an independent audit), so the headline numbers should be read as directional until reproduced with a third-party harness.