BuilderBuilder
GPT-6 Astra · Blossom arena
The episode says the Claude runs were stopped around $300 API-equivalent before the scenes were finished, and that the GPT models finished, with Astra at 6 hours 28 minutes. Opus attempt 1 is PARTIAL at $433.73 after a subscription-limit stop. Sonnet is partial_user_stopped at $312.27 over about three hours. GPT-6.1 Sol is partial_user_stopped with verdict STOPPED_PARTIAL. Those statuses stay. Astra's clean wall clock is unavailable; a supplemental span of 6h 28m 15.8s is recorded with an explicit scope warning and is not promoted to the measured clock. Supplied source capture Unaided model Open this round
Measured harness ledgerPublic result
GPT-6 Astra

Courtyard Blossom Stage 2 — GPT-6 Astra Max

Create a second cherry-blossom temple courtyard arena for the fixed fighting game under the pinned reference-guided Stage 2 contract.

Max reasoningHeadline result
Workflow cost
$74.50
Wall-clock
Not recorded
Processed tokens
53.63M processed
Record state
partial_root_snapshot_metrics_with_independent_in_app_browser_replay_and_unverified_model_geometry_evidence
Public summary

GPT-6 Astra Max partial_root_snapshot_metrics_with_independent_in_app_browser_replay_and_unverified_model_geometry_evidence ledger: Not recorded wall-clock, 53.63M processed, and $74.50 OpenAI Standard API-list-price equivalent, not an itemized Codex subscription charge.

Run identity and stack
  • Result ID: courtyard-arena-blossom-stage-2-reference-exposed-gpt-6-astra-max
  • Technical model: GPT-6 Astra
  • Provider: OpenAI Codex
  • Stack: OpenAI Codex
  • Stack: Technical model/configuration: GPT-6 Astra
  • Stack: Blender / browser arena workflow
  • Stack: Reference-guided Stage 2 fixture
Cost basis
  • Separately priced tools and non-token services are excluded.
Recorded caveats
  • The raw parser has no timing result because its user-message-schema expectation did not match this transcript. The displayed 6h 28m 15.8s figure is a qualified supplemental root-session span, not a clean generation-only duration.
  • The cumulative root-task snapshot includes tool time and idle gaps, and does not establish a clean public one-user-turn boundary.
  • Output tokens include reasoning, visible prose and code, and tool-call JSON.
  • Cache reads are discounted, so processed-token volume overstates effective cost.
  • This is an API-equivalent estimate; separately priced tools and non-token services are excluded.
  • Archive-operator replay establishes functional routes and input response only. It does not establish standardized FPS, transition hitch, cross-hardware performance, blind preference, or independent textures-off geometry quality.
  • No blind evaluation is recorded.
Public result only

This result keeps its public summary and evidence, but it does not currently have a matching Builder test with prompts, projects, Harness workflows or skills.

Browse BuilderReturn to GPT-6 Astra
RemakeBenchResearch console