RECENTLY UPDATED / (SHANGHAI TIME)

CHEER LIN / ARCHIVE

GPT-6.2 Astra Use Cases

Explore 84 GPT-6.2 Astra public cases across 3D animation, web interaction, game prototypes, tool use, reasoning and creative writing. Watch demos and check the original accounts.

84 PUBLIC CASES8 TASK CATEGORIES3 LANGUAGES

USE CASE MAP / 01

Explore Astra through public examples

Model labels, timing and comparison observations follow the creators’ accounts; this site has not independently rerun the tasks. Reposts of the same work are grouped. Question marks, routing labels and differences in tools remain in the case descriptions.

84curated real-world cases
8task categories
3English, Chinese, and Japanese

HOW TO USE / 02

From case discovery to workflow validation

Find a related task, play the preview and review the original post, then use the disclosed inputs and tools to design your own test.

  1. 01

    Choose a related task

    Find a similar goal in animation, games, tool use, writing or another category.

  2. 02

    Watch the work

    Use the video or screenshot to inspect the visible result.

  3. 03

    Review the creator’s account

    Open the creator link to check the task, model label, tools and comparison conditions.

  4. 04

    Test your own result

    Use a public brief or your own inputs and record the actual model, time and outcome.

PUBLIC CASES / 03

84 GPT-6.2 Astra public cases

Select a video cover to play it or the creator’s name to open the original post, work or evaluation. Full details retain the task context and comparison conditions.

84 cases
SORT
CASE 01Evaluation
3D & animation

3D helicopter scenes

Explore procedural helicopter modeling through two contrasting Three.js scenes. SadAlbert10 asks for a realistic Mario-themed helicopter in Three.js without using an existing model. The paired clips show two different scene treatments: an airfield view and a forest helipad. The source labels the left result Astra 6/6.2 and the right Opus 5.5. Use the pair to compare the helicopter shape, environment and presentation against the same brief, while keeping the creator’s model labels and ordering intact. This card plays the clip labeled Astra 6/6.2 by the source; the Opus 5.5 comparison clip is available in the original post.

3D helicopter scenes

Explore procedural helicopter modeling through two contrasting Three.js scenes. SadAlbert10 asks for a realistic Mario-themed helicopter in Three.js without using an existing model. The paired clips show two different scene treatments: an airfield view and a forest helipad. The source labels the left result Astra 6/6.2 and the right Opus 5.5. Use the pair to compare the helicopter shape, environment and presentation against the same brief, while keeping the creator’s model labels and ordering intact. This card plays the clip labeled Astra 6/6.2 by the source; the Opus 5.5 comparison clip is available in the original post.

SadAlbert10 ↗
SadAlbert10
CASE 02Demo
Illustration & SVG

Realistic PS5 controller SVG

Use a familiar product as a test of shape, shading and material detail. SadAlbert10 presents a PS5 controller illustration as an SVG result, reporting a single attempt without a reference image. The public brief is short: make a realistic PS5 controller. The visible image combines a white shell, dark center grips, familiar buttons and soft shading. It is a useful visual-design example for examining recognizable geometry and product detail; the original post provides the image and the author’s description of the vector output.

Realistic PS5 controller SVG

Use a familiar product as a test of shape, shading and material detail. SadAlbert10 presents a PS5 controller illustration as an SVG result, reporting a single attempt without a reference image. The public brief is short: make a realistic PS5 controller. The visible image combines a white shell, dark center grips, familiar buttons and soft shading. It is a useful visual-design example for examining recognizable geometry and product detail; the original post provides the image and the author’s description of the vector output.

SadAlbert10 ↗
SadAlbert10
CASE 03Evaluation
Front-end & motion

A robot awakens in a hand-drawn forest

Turn a detailed scene brief into a short animated sequence. Leo compares an ancient robot awakening and taking flight in a forest, with the upper sequence labeled Astra and the lower Fable 5.5. The public prompt asks for a hand-drawn 2D look and a single self-contained HTML deliverable, specifying lighting, movement and staging. The source credits Chetaslua for the Fable side and describes the Astra result as 6(6.2). Compare how each sequence interprets the same setting, story beat and visual direction.

A robot awakens in a hand-drawn forest

Turn a detailed scene brief into a short animated sequence. Leo compares an ancient robot awakening and taking flight in a forest, with the upper sequence labeled Astra and the lower Fable 5.5. The public prompt asks for a hand-drawn 2D look and a single self-contained HTML deliverable, specifying lighting, movement and staging. The source credits Chetaslua for the Fable side and describes the Astra result as 6(6.2). Compare how each sequence interprets the same setting, story beat and visual direction.

leo114119 ↗
leo114119
CASE 04Demo
3D & animation

Pelican cycling around a vertical loop

Combine character motion, a bicycle and a looping three-dimensional track. FiniYang’s first demonstration places a white pelican on a red bicycle above a coastal scene. The track rises into a vertical loop, and the sequence shows the rider upside down against an eclipsed sun before returning toward the bridge. This makes the clip useful for inspecting how character pose, bicycle orientation, track geometry and camera framing work together. Reposts and the later split-screen comparison of this same loop-track output are grouped with the original work.

Pelican cycling around a vertical loop

Combine character motion, a bicycle and a looping three-dimensional track. FiniYang’s first demonstration places a white pelican on a red bicycle above a coastal scene. The track rises into a vertical loop, and the sequence shows the rider upside down against an eclipsed sun before returning toward the bridge. This makes the clip useful for inspecting how character pose, bicycle orientation, track geometry and camera framing work together. Reposts and the later split-screen comparison of this same loop-track output are grouped with the original work.

FiniYang ↗
FiniYang
CASE 05Demo
Games & scenes

A semi-open-world 3D RPG

Inspect a small game world that combines a character, environment and interface. FiniYang’s second demonstration is a semi-open-world role-playing game, described by the creator as made in one go. The footage shows a third-person character moving through stone ruins and vegetation, with a minimap, task text and action icons framing the scene. The clip brings together world-building and game-interface design in one visible result. Follow the original for the creator’s gameplay presentation and use the scene as a reference for the components of a compact RPG prototype.

A semi-open-world 3D RPG

Inspect a small game world that combines a character, environment and interface. FiniYang’s second demonstration is a semi-open-world role-playing game, described by the creator as made in one go. The footage shows a third-person character moving through stone ruins and vegetation, with a minimap, task text and action icons framing the scene. The clip brings together world-building and game-interface design in one visible result. Follow the original for the creator’s gameplay presentation and use the scene as a reference for the components of a compact RPG prototype.

FiniYang ↗
FiniYang
CASE 06Demo
3D & animation

Coastal Cruise: a 3D pelican cyclist

Present a character animation inside a compact scene-player interface. Leo’s Coastal Cruise page shows a helmeted pelican on a green bicycle beside the sea and a lighthouse. Playback and speed controls sit below the scene, giving the result the feel of a small interactive exhibit. Leo reports roughly eleven minutes of work. The character, bicycle, coastal setting and page layout distinguish this 3D example from the separate hand-drawn cycling study. Inspect both the scene composition and the way the player presents the animation.

Coastal Cruise: a 3D pelican cyclist

Present a character animation inside a compact scene-player interface. Leo’s Coastal Cruise page shows a helmeted pelican on a green bicycle beside the sea and a lighthouse. Playback and speed controls sit below the scene, giving the result the feel of a small interactive exhibit. Leo reports roughly eleven minutes of work. The character, bicycle, coastal setting and page layout distinguish this 3D example from the separate hand-drawn cycling study. Inspect both the scene composition and the way the player presents the animation.

leo114119 ↗
leo114119
CASE 07Demo
3D & animation

The Pelican Cyclist: an illustrated seaside ride

Explore a flat illustrated treatment of the classic pelican-on-a-bicycle task. Leo’s separate Pelican Cyclist example uses a hand-drawn seaside scene, a sunset, a lighthouse and a lightly textured color palette. A pause control and playback-speed options sit beneath the illustration. The work has a different visual language and page layout from Coastal Cruise, so it is retained as its own demonstration. Use it to compare how the same familiar subject can become either a dimensional character scene or a quieter illustrated motion study.

The Pelican Cyclist: an illustrated seaside ride

Explore a flat illustrated treatment of the classic pelican-on-a-bicycle task. Leo’s separate Pelican Cyclist example uses a hand-drawn seaside scene, a sunset, a lighthouse and a lightly textured color palette. A pause control and playback-speed options sit beneath the illustration. The work has a different visual language and page layout from Coastal Cruise, so it is retained as its own demonstration. Use it to compare how the same familiar subject can become either a dimensional character scene or a quieter illustrated motion study.

leo114119 ↗
leo114119
CASE 08Evaluation
Games & scenes

A Japanese fantasy city in pixel art

Compare two interpretations of a detailed city-animation brief. notjazii shares a split GIF of two Japanese-style waterfront cities, with castles, bridges, boats and a moonlit sky. The upper panel is labeled GPT-6.2 Astra and the lower GPT-6 Astra. The creator reports the same prompt, a single attempt and max reasoning for both. The newer run was made by a friend in Codex, while the earlier run used Devin. Keep that difference in host and operator alongside the visual comparison.

A Japanese fantasy city in pixel art

Compare two interpretations of a detailed city-animation brief. notjazii shares a split GIF of two Japanese-style waterfront cities, with castles, bridges, boats and a moonlit sky. The upper panel is labeled GPT-6.2 Astra and the lower GPT-6 Astra. The creator reports the same prompt, a single attempt and max reasoning for both. The newer run was made by a friend in Codex, while the earlier run used Devin. Keep that difference in host and operator alongside the visual comparison.

notjazii ↗
notjazii
CASE 09Demo
Front-end & motion

An animated model introduction

Inspect a short motion-design piece built around abstract imagery and typography. GulaGala001 shares a 24-second introduction and attributes its creation to gpt-6.2-astra. The clip combines a reflective abstract ring with changing introduction text, including an Every question frame. It extends the collection beyond game and character scenes into a compact promotional motion piece. Look at the relationship between the central graphic, contrast and typography in the published clip; the source does not specify a particular implementation or production method.

An animated model introduction

Inspect a short motion-design piece built around abstract imagery and typography. GulaGala001 shares a 24-second introduction and attributes its creation to gpt-6.2-astra. The clip combines a reflective abstract ring with changing introduction text, including an Every question frame. It extends the collection beyond game and character scenes into a compact promotional motion piece. Look at the relationship between the central graphic, contrast and typography in the published clip; the source does not specify a particular implementation or production method.

GulaGala001 ↗
GulaGala001
CASE 10Evaluation
Games & scenes

Four models build an open-world RPG

Compare four game-scene treatments in one shared presentation. noclipepe presents four RPG outputs in a single comparison video: GPT-6.2 Astra (?), Fable 5.5 (routing), Opus 5.5 and Fable 5.1. The publisher describes a common prompt and contrasts the environments, lighting and gameplay presentation. The Astra panel shows bright ruins and a golem scene. Keep the question mark and routing label from the source when reading the comparison. The AI/ML API-branded video includes the publisher’s own ratings, rather than a controlled benchmark.

Four models build an open-world RPG

Compare four game-scene treatments in one shared presentation. noclipepe presents four RPG outputs in a single comparison video: GPT-6.2 Astra (?), Fable 5.5 (routing), Opus 5.5 and Fable 5.1. The publisher describes a common prompt and contrasts the environments, lighting and gameplay presentation. The Astra panel shows bright ruins and a golem scene. Keep the question mark and routing label from the source when reading the comparison. The AI/ML API-branded video includes the publisher’s own ratings, rather than a controlled benchmark.

noclipepe ↗
noclipepe
CASE 11Demo
3D & animation

3D pose estimation on a basketball court

skalskip92 shows multiple players mapped to a 3D court. The author reports one GPT-6.2 Astra API call at high reasoning effort. The clip lets you inspect player placement and poses. The post gives no quantitative accuracy measure.

3D pose estimation on a basketball court

skalskip92 shows multiple players mapped to a 3D court. The author reports one GPT-6.2 Astra API call at high reasoning effort. The clip lets you inspect player placement and poses. The post gives no quantitative accuracy measure.

skalskip92 ↗
skalskip92
CASE 12Demo
Games & scenes

Single-file Babylon.js dungeon RPG

intheworldofai used Astra Ultra to make a dungeon RPG in Babylon.js. The author says a short follow-up asking for improvements added a frozen forge, a blacksmith and a golem that erupts the floor beneath the player. Both dungeons reportedly run offline from one HTML file.

Single-file Babylon.js dungeon RPG

intheworldofai used Astra Ultra to make a dungeon RPG in Babylon.js. The author says a short follow-up asking for improvements added a frozen forge, a blacksmith and a golem that erupts the floor beneath the player. Both dungeons reportedly run offline from one HTML file.

intheworldofai ↗
intheworldofai
CASE 13Demo
Front-end & motion

Browser-based CD jukebox

ShadeLurk demonstrates an arm pulling discs from their cases and two decks transitioning between tracks. The author says the single HTML file bundles five tracks, using Canvas and WebGL for visuals and Web Audio for sound. The post points to the shared HTML in the replies.

Browser-based CD jukebox

ShadeLurk demonstrates an arm pulling discs from their cases and two decks transitioning between tracks. The author says the single HTML file bundles five tracks, using Canvas and WebGL for visuals and Web Audio for sound. The post points to the shared HTML in the replies.

ShadeLurk ↗
ShadeLurk
CASE 14Evaluation
3D & animation

Turning map data into a walking Three.js mech

Simon Willison gave Astra and Sol the same Three.js brief using OpenStreetMap data. Yonkers was to rise as a mech and stomp through several New York City boroughs. The original brief gives viewers a basis for comparing how the outputs handle map data, movement and scenery.

Turning map data into a walking Three.js mech

Simon Willison gave Astra and Sol the same Three.js brief using OpenStreetMap data. Yonkers was to rise as a mech and stomp through several New York City boroughs. The original brief gives viewers a basis for comparing how the outputs handle map data, movement and scenery.

simonw ↗
simonw
CASE 15Evaluation
3D & animation

Blender toy-cowboy short film comparison

ChrisGPT gave Astra and Fable 5.2 the same brief for a 10-second toy-cowboy film in Blender. The author reports about two hours and 60,000 generated tokens for Astra, and about 20 hours and 150,000 to 300,000 generated tokens for Fable. The post gives these run-level figures without an independent quality score.

Blender toy-cowboy short film comparison

ChrisGPT gave Astra and Fable 5.2 the same brief for a 10-second toy-cowboy film in Blender. The author reports about two hours and 60,000 generated tokens for Astra, and about 20 hours and 150,000 to 300,000 generated tokens for Fable. The post gives these run-level figures without an independent quality score.

ChrisGPT ↗
ChrisGPT
CASE 16Evaluation
Games & scenes

Pixel flight game from the same prompt

Densitylucky asked Astra and Fable 5.2 to build separate pixel-art flight games from the same prompt. The creator reports costs of about $10.03 for Astra and $22.78 for Fable, describes a 143 KB game code file and points to the games and prompt in the original thread. The figures apply to this comparison.

Pixel flight game from the same prompt

Densitylucky asked Astra and Fable 5.2 to build separate pixel-art flight games from the same prompt. The creator reports costs of about $10.03 for Astra and $22.78 for Fable, describes a 143 KB game code file and points to the games and prompt in the original thread. The figures apply to this comparison.

Densitylucky ↗
Densitylucky
CASE 17Evaluation
3D & animation

Checking asset provenance in a T-Rex modeling task

luckeyfaraday asked Astra to build a 3D T-Rex. The author reports that the run downloaded an existing asset and shares a screenshot. This failure report suggests an acceptance check: trace the model file and verify whether reuse was allowed for the task.

Checking asset provenance in a T-Rex modeling task

luckeyfaraday asked Astra to build a 3D T-Rex. The author reports that the run downloaded an existing asset and shares a screenshot. This failure report suggests an acceptance check: trace the model file and verify whether reuse was allowed for the task.

luckeyfaraday ↗
luckeyfaraday
CASE 18Evaluation
3D & animation

3D disassembly of two interlocked parts

PhysAIAnalysis tested models on separating two interlocked claw-shaped pieces. The evaluator reports that Astra took its interlock measure from 71 to zero in 24 minutes at $10.24, and that 12 of the 16 models failed. The measures and costs come from the evaluator.

3D disassembly of two interlocked parts

PhysAIAnalysis tested models on separating two interlocked claw-shaped pieces. The evaluator reports that Astra took its interlock measure from 71 to zero in 24 minutes at $10.24, and that 12 of the 16 models failed. The measures and costs come from the evaluator.

PhysAIAnalysis ↗
PhysAIAnalysis
CASE 19Demo
Front-end & motion

Adapting one creator clip into three ads

Magnific’s own product demo shows a workflow for adapting one creator clip into three ads. The process uses GPT-6.2 Astra, Magnific MCP and Magnific Illusion. Magnific says it required no reshoot and shares the process and prompts in its thread.

Adapting one creator clip into three ads

Magnific’s own product demo shows a workflow for adapting one creator clip into three ads. The process uses GPT-6.2 Astra, Magnific MCP and Magnific Illusion. Magnific says it required no reshoot and shares the process and prompts in its thread.

magnific ↗
magnific
CASE 20Evaluation
Coding & tool use

Tool-call errors and completion time in ts-bench

gosrum evaluated Astra and Sol with ts-bench and compared them with Claude and Gemini 4 Hydrogen. The author reports tool-call errors in the GPT-6.2 runs and none in the compared Claude and Hydrogen runs. GPT-6.2 reportedly completed the task faster than Hydrogen. These observations come from the author’s test.

Tool-call errors and completion time in ts-bench

gosrum evaluated Astra and Sol with ts-bench and compared them with Claude and Gemini 4 Hydrogen. The author reports tool-call errors in the GPT-6.2 runs and none in the compared Claude and Hydrogen runs. GPT-6.2 reportedly completed the task faster than Hydrogen. These observations come from the author’s test.

gosrum ↗
gosrum
CASE 21Evaluation
Creative writing

Short-story evaluation with ten required elements

Lech Mazur’s benchmark asks for 600–800-word stories shaped by ten required elements and compares pairs using judges from three model families. The publisher reports a 4.30 score for Astra, with more restrained endings and recurring themes of family and craft. The observations apply to this short-fiction setup.

Short-story evaluation with ten required elements

Lech Mazur’s benchmark asks for 600–800-word stories shaped by ten required elements and compares pairs using judges from three model families. The publisher reports a 4.30 score for Astra, with more restrained endings and recurring themes of family and craft. The observations apply to this short-fiction setup.

LechMazur ↗
LechMazur
CASE 22Demo
Games & scenes

Browser-based brick-breaker prototype

goyozi used a brick-breaker game as a personal comparison and posted links to Astra, Sol, Fable and Opus versions on Hacker News. The cover shows AX Ball, the Astra version linked by the author. The original comment leads to the demos so viewers can compare controls and feedback.

Browser-based brick-breaker prototype

goyozi used a brick-breaker game as a personal comparison and posted links to Astra, Sol, Fable and Opus versions on Hacker News. The cover shows AX Ball, the Astra version linked by the author. The original comment leads to the demos so viewers can compare controls and feedback.

goyozi ↗
goyozi
CASE 23Evaluation
Coding & tool use

Robot manipulation code on LIBERO

Playground compared programs written by different models for robot-arm tasks and showed simulated runs. It reports that Astra’s Python program passed 96.8% and that a separate neural-network approach passed 79% of 500 LIBERO trials. The figures cover different implementations in the evaluator’s simulation.

Robot manipulation code on LIBERO

Playground compared programs written by different models for robot-arm tasks and showed simulated runs. It reports that Astra’s Python program passed 96.8% and that a separate neural-network approach passed 79% of 500 LIBERO trials. The figures cover different implementations in the evaluator’s simulation.

withplayground ↗
withplayground
CASE 24Evaluation
3D & animation

Detail and build time in 3D race-car models

ZryMiller shows the race-car models from Fable 5.2 and Astra. The creator reports about three hours for Astra and two days of work on the Fable version, and describes the Astra car as less detailed. The previews let viewers compare the wheels, livery and bodywork across two differently paced builds.

Detail and build time in 3D race-car models

ZryMiller shows the race-car models from Fable 5.2 and Astra. The creator reports about three hours for Astra and two days of work on the Fable version, and describes the Astra car as less detailed. The previews let viewers compare the wheels, livery and bodywork across two differently paced builds.

ZryMiller ↗
ZryMiller
CASE 25Evaluation
Games & scenes

Exploring Atlantis in a submarine

Higgsfield’s own product demo shows Fable 5.2 and Astra versions of an underwater exploration game. The task is to pilot a submarine through ruins. The clip lets viewers compare the viewpoint, lighting and scene layout. Higgsfield credits its own platform in the post.

Exploring Atlantis in a submarine

Higgsfield’s own product demo shows Fable 5.2 and Astra versions of an underwater exploration game. The task is to pilot a submarine through ruins. The clip lets viewers compare the viewpoint, lighting and scene layout. Higgsfield credits its own platform in the post.

higgsfield_ai ↗
higgsfield_ai
CASE 26Evaluation
3D & animation

A tennis court in the VoxelBench scene test

VoxelBench shows completed voxel scenes. This card uses the tennis court labeled GPT-6.2 Astra in the original image, showing court markings, seating and surroundings. The same post also contains Sol outputs, and the evaluator reports an elevated error rate for Astra.

A tennis court in the VoxelBench scene test

VoxelBench shows completed voxel scenes. This card uses the tennis court labeled GPT-6.2 Astra in the original image, showing court markings, seating and surroundings. The same post also contains Sol outputs, and the evaluator reports an elevated error rate for Astra.

voxelbench ↗
voxelbench
CASE 27Evaluation
3D & animation

From a meme drawing to a printable 3D figure

Lukas Petersson shares a Wojak-Bench 3D leaderboard and a bicycle example. The task is to convert a flat drawing into a printable figure. Astra led the posted leaderboard, while the author called out the bicycle prompt as a difficult case for Astra.

From a meme drawing to a printable 3D figure

Lukas Petersson shares a Wojak-Bench 3D leaderboard and a bicycle example. The task is to convert a flat drawing into a printable figure. Astra led the posted leaderboard, while the author called out the bicycle prompt as a difficult case for Astra.

lukaspet ↗
lukaspet
CASE 28Evaluation
Reasoning & analysis

Visual questions on EyeBench

The EyeBench leaderboard posted by Adonis Singh shows GPT-6.2 Astra first and Sol second. The publisher reports that both finished within two questions of a perfect result. The screenshot records that benchmark version and setup and includes timing and cost columns.

Visual questions on EyeBench

The EyeBench leaderboard posted by Adonis Singh shows GPT-6.2 Astra first and Sol second. The publisher reports that both finished within two questions of a perfect result. The screenshot records that benchmark version and setup and includes timing and cost columns.

adonis_singh ↗
adonis_singh
CASE 29Evaluation
Reasoning & analysis

Long-horizon financial analysis on DeepFinance

DeepFinance evaluates models on end-to-end private-equity and investment-banking tasks. The evaluator’s posted snapshot puts Astra second at 61.4%, 3.8 percentage points ahead of the preceding Astra. The figures reflect the benchmark’s task set and scoring rubric.

Long-horizon financial analysis on DeepFinance

DeepFinance evaluates models on end-to-end private-equity and investment-banking tasks. The evaluator’s posted snapshot puts Astra second at 61.4%, 3.8 percentage points ahead of the preceding Astra. The figures reflect the benchmark’s task set and scoring rubric.

0interestrates ↗
0interestrates
CASE 30Demo
Creative writing

Using manga to think about people and AI agents

Suzacque shows Astra elaborating on an analogy between Stand users in JoJo and people who direct AI agents. The reply explores the division of work between the user and the Stand. The cover is the creator’s screenshot of the Japanese reply.

Using manga to think about people and AI agents

Suzacque shows Astra elaborating on an analogy between Stand users in JoJo and people who direct AI agents. The reply explores the division of work between the user and the Stand. The cover is the creator’s screenshot of the Japanese reply.

Suzacque ↗
Suzacque
CASE 31Demo
3D & animation

Sailing ship at sunset

Da7_Tech labels a demonstration of a sailing ship on the ocean as GPT-6.2 Astra. The clip shows sunset lighting, waves and moving sails. The post does not give the prompt or production tools.

Sailing ship at sunset

Da7_Tech labels a demonstration of a sailing ship on the ocean as GPT-6.2 Astra. The clip shows sunset lighting, waves and moving sails. The post does not give the prompt or production tools.

Da7_Tech ↗
Da7_Tech
CASE 32Evaluation
Reasoning & analysis

Two harnesses on ARC-AGI-3

A Hacker News thread discussed the Astra results published by ARC Prize. ARC Prize reports 100% at xhigh under the Provider Adapter harness, which retains reasoning state, and a best Standard-harness score of 70.3% at max. The scorecard comes from ARC Prize and shows both harnesses by effort level.

Two harnesses on ARC-AGI-3

A Hacker News thread discussed the Astra results published by ARC Prize. ARC Prize reports 100% at xhigh under the Provider Adapter harness, which retains reasoning state, and a best Standard-harness score of 70.3% at max. The scorecard comes from ARC Prize and shows both harnesses by effort level.

rzk ↗
rzk
CASE 33Evaluation
3D & animation

A split-screen pelican ride by the sea

Octoberry posted a separate seaside cycling comparison. The upper panel carries a Fable 5.5 / Opus 5.5 routing label and the lower panel is labeled GPT-6.2 Astra. The clip lets viewers compare lighting and motion under the creator’s prerelease labels.

A split-screen pelican ride by the sea

Octoberry posted a separate seaside cycling comparison. The upper panel carries a Fable 5.5 / Opus 5.5 routing label and the lower panel is labeled GPT-6.2 Astra. The clip lets viewers compare lighting and motion under the creator’s prerelease labels.

Octoberry ↗
Octoberry
CASE 34Demo
Illustration & SVG

A fashion poster built around model settings

Masimo_Blue shares a fictional clothing poster built around Astra’s reasoning and speed settings. The blue shorts feature a garment tag that repurposes those settings as care instructions. The source includes the chat and the poster.

A fashion poster built around model settings

Masimo_Blue shares a fictional clothing poster built around Astra’s reasoning and speed settings. The blue shorts feature a garment tag that repurposes those settings as care instructions. The source includes the chat and the poster.

Masimo_Blue ↗
Masimo_Blue
CASE 35Evaluation
Illustration & SVG

Drawing a unicorn in TikZ

Angaisb_ shares a comparison of GPT-4 and GPT-6.2 Astra drawing the same subject with LaTeX and TikZ. The Astra panel shows a running unicorn against a starry sky. The model labels come from the published comparison.

Drawing a unicorn in TikZ

Angaisb_ shares a comparison of GPT-4 and GPT-6.2 Astra drawing the same subject with LaTeX and TikZ. The Astra panel shows a running unicorn against a starry sky. The model labels come from the published comparison.

Angaisb_ ↗
Angaisb_
CASE 36Demo
Front-end & motion

Layer-based video editing in Higgsfield Miracle

Higgsfield demonstrates a workflow that separates a shot into editable elements. The company says GPT-6.2 Astra powers pixel-level changes to parts of a shot. The preview documents Higgsfield’s own private-beta demonstration.

Layer-based video editing in Higgsfield Miracle

Higgsfield demonstrates a workflow that separates a shot into editable elements. The company says GPT-6.2 Astra powers pixel-level changes to parts of a shot. The preview documents Higgsfield’s own private-beta demonstration.

higgsfield ↗
higgsfield
CASE 37Evaluation
Reasoning & analysis

Pass rates and output tokens on Harvey legal tasks

OpenAI shares a chart comparing results on Harvey LAB-AA v1.1, which includes tasks such as reviewing contract terms against source materials. The chart lists GPT-6.2 Astra and its separately configured Astra for Law with their pass rates and mean output tokens.

Pass rates and output tokens on Harvey legal tasks

OpenAI shares a chart comparing results on Harvey LAB-AA v1.1, which includes tasks such as reviewing contract terms against source materials. The chart lists GPT-6.2 Astra and its separately configured Astra for Law with their pass rates and mean output tokens.

sherwinwu ↗
sherwinwu
CASE 38Evaluation
Coding & tool use

Permission dialogs after a long computer-use run

Theo posted a desktop screenshot after leaving computer use running, with multiple system permission prompts visible. His follow-up attributes the run to GPT-6.2 Astra. The image records the visible end state; the task and duration are unspecified.

Permission dialogs after a long computer-use run

Theo posted a desktop screenshot after leaving computer use running, with multiple system permission prompts visible. His follow-up attributes the run to GPT-6.2 Astra. The image records the visible end state; the task and duration are unspecified.

theo ↗
theo
CASE 39Demo
Front-end & motion

A stylized character in a vertical short

henriqueim0veis shares a vertical short featuring a stylized character in sunglasses indoors. The creator attributes the clip to GPT-6.2 Astra. The post gives no production steps or companion tools.

A stylized character in a vertical short

henriqueim0veis shares a vertical short featuring a stylized character in sunglasses indoors. The creator attributes the clip to GPT-6.2 Astra. The post gives no production steps or companion tools.

henriqueim0veis ↗
henriqueim0veis
CASE 40Demo
Coding & tool use

Selecting Astra and reasoning effort in Orca

Orca’s team shows GPT-6.2 Astra in its model selector, alongside reasoning effort and speed controls. The publisher reports support in both its terminal and chat interfaces. The screenshot illustrates model selection.

Selecting Astra and reasoning effort in Orca

Orca’s team shows GPT-6.2 Astra in its model selector, alongside reasoning effort and speed controls. The publisher reports support in both its terminal and chat interfaces. The screenshot illustrates model selection.

orca_build ↗
orca_build
CASE 41Evaluation
Reasoning & analysis

Trade-offs in parallel agent research

gnukeith shows an orchestrated research run with multiple agents. The author reports that Astra and Sol cut steps the author considered important. The screenshot shows the task tree, and the quality assessment comes from this run.

Trade-offs in parallel agent research

gnukeith shows an orchestrated research run with multiple agents. The author reports that Astra and Sol cut steps the author considered important. The screenshot shows the task tree, and the quality assessment comes from this run.

gnukeith ↗
gnukeith
CASE 42Evaluation
3D & animation

Portal and machinery

VoxelBench labels this result GPT-6.2 Astra. The scene shows a ring portal, mechanical arms and a tiled floor, giving viewers a way to inspect composition and details for this voxel task.

Portal and machinery

VoxelBench labels this result GPT-6.2 Astra. The scene shows a ring portal, mechanical arms and a tiled floor, giving viewers a way to inspect composition and details for this voxel task.

voxelbench ↗
voxelbench
CASE 43Evaluation
3D & animation

A pink-haired figure in a round frame

VoxelBench labels this result GPT-6.2 Astra. The scene shows a pink-haired figure, a gold frame and a stage base, giving viewers a way to inspect composition and details for this voxel task.

A pink-haired figure in a round frame

VoxelBench labels this result GPT-6.2 Astra. The scene shows a pink-haired figure, a gold frame and a stage base, giving viewers a way to inspect composition and details for this voxel task.

voxelbench ↗
voxelbench
CASE 44Evaluation
3D & animation

A voxel figure with twin tails

VoxelBench labels this result GPT-6.2 Astra. The scene shows green twin tails, a raised leg and a circular stand, giving viewers a way to inspect composition and details for this voxel task.

A voxel figure with twin tails

VoxelBench labels this result GPT-6.2 Astra. The scene shows green twin tails, a raised leg and a circular stand, giving viewers a way to inspect composition and details for this voxel task.

voxelbench ↗
voxelbench
CASE 45Evaluation
3D & animation

A voxel portrait in a green dress

VoxelBench labels this result GPT-6.2 Astra. The scene shows a slit dress, light hair and folds in the fabric, giving viewers a way to inspect composition and details for this voxel task.

A voxel portrait in a green dress

VoxelBench labels this result GPT-6.2 Astra. The scene shows a slit dress, light hair and folds in the fabric, giving viewers a way to inspect composition and details for this voxel task.

voxelbench ↗
voxelbench
CASE 46Evaluation
3D & animation

A voxel living-room scene

VoxelBench labels this result GPT-6.2 Astra. The scene shows a figure, a couch, an arched window and furniture, giving viewers a way to inspect composition and details for this voxel task.

A voxel living-room scene

VoxelBench labels this result GPT-6.2 Astra. The scene shows a figure, a couch, an arched window and furniture, giving viewers a way to inspect composition and details for this voxel task.

voxelbench ↗
voxelbench
CASE 47Evaluation
3D & animation

A voxel robot and a hanging Earth

VoxelBench labels this result GPT-6.2 Astra. The scene shows a boxy robot, a mechanical arm and a suspended globe, giving viewers a way to inspect composition and details for this voxel task.

A voxel robot and a hanging Earth

VoxelBench labels this result GPT-6.2 Astra. The scene shows a boxy robot, a mechanical arm and a suspended globe, giving viewers a way to inspect composition and details for this voxel task.

voxelbench ↗
voxelbench
CASE 48Evaluation
3D & animation

A sea creature and a sailing ship

VoxelBench labels this result GPT-6.2 Astra. The scene shows a winged sea creature, a sailing ship and waves, giving viewers a way to inspect composition and details for this voxel task.

A sea creature and a sailing ship

VoxelBench labels this result GPT-6.2 Astra. The scene shows a winged sea creature, a sailing ship and waves, giving viewers a way to inspect composition and details for this voxel task.

voxelbench ↗
voxelbench

Showing 48 / 84

BUILD WITH EVOLINK / 04

Start with a concrete task

Explore the GPT-6.2 Astra API on EvoLink. Use the cases to choose inputs, tools and acceptance criteria, then test the outcome on your own task.