Skip to content
← All news

GPT-6.2 Astra and Sol: coding results by effort

46/50 native-setting artifacts and 10/10 requested-ultra diagnostics and 9/10 eligible delegation-enabled artifacts are published. Latest native-setting phase: low.


The low phase has 10 published artifacts from 10 requested cells. 0 cells await publication or reconciliation; 0 attempt records have no published artifact. Across publication phases, 46/50 native-setting artifacts and 10/10 requested-ultra diagnostic artifacts and 9/10 eligible delegation-enabled artifacts are available (65 total). Published requested settings: max, xhigh, high, medium, low, requested ultra (diagnostic).

The five tasks were selected by historical decisive voting activity recorded on November 17, 2026: landing page (1,192 votes), dashboard (251), palette tool (218), arcade game (202), and news front page (152). These counts describe accumulated task voting activity.

9/10 separate delegation-enabled Ultra artifacts are published. These use a single parent request with bounded native collaboration, persistent history, and the same requested model and Ultra setting for child agents. Parent and child settings, completed turns, and observed tool types are checked separately. Local AGENTS.md instructions were loaded and privately archived as additional context beyond the displayed runner brief and wrapper. The actual wire tool schema and provider-resolved effort were not captured. The per-thread usage table uses anonymous Parent/Child ordinals. An aggregate is shown only when the latest cumulative usage of every distinct thread can be counted once without inherited usage. This protocol differs from the single-turn diagnostics and native-setting comparisons.

Delegation-enabled taskGPT-6.2 AstraGPT-6.2 Sol
Landing pagePublished; 2 completed child agentsProtocol-limited response; excluded
DashboardPublished; no child agents observedPublished; 2 completed child agents
Palette toolPublished; no child agents observedPublished; no child agents observed
Arcade gamePublished; no child agents observedPublished; no child agents observed
News front pagePublished; no child agents observedPublished; no child agents observed
TaskGPT-6.2 Astra: lowGPT-6.2 Sol: low
Landing pagePublishedPublished
DashboardPublishedPublished
Palette toolPublishedPublished
Arcade gamePublishedPublished
News front pagePublishedPublished
EffortAstra publishedSol published
max4/55/5
xhigh3/55/5
high5/54/5
medium5/55/5
low5/55/5
Requested ultra: delegation disabled, diagnostic5/55/5

The native-setting artifacts and requested-ultra diagnostics use explicit model and effort requests through the Codex CLI subscription route under a single-turn protocol. Selected tool capabilities were disabled, and no tool items were observed in the CLI event stream. The actual wire tool schema was not captured. Each published artifact in these two groups comes from one completed turn and one final assistant response. The first response is preserved. Exact prompts and output hashes are recorded in the arena metadata. The event stream does not expose a provider-resolved model identity. Native and delegation-enabled runners omit a maximum output-token request, leaving the provider default in effect; the effective provider ceiling is unknown. These attempts do not use a verified matched output-token budget. The configured client timeout is 90 minutes, measured on the runner monotonic clock; it is separate from the unknown provider output-token ceiling.

Desktop and mobile browser checks cover rendering, runtime exceptions, layout overflow, and limited control interactions. Each artifact uses a fresh browser profile; its desktop and mobile visits share storage within that profile. Mobile checks can therefore include state retained from the desktop visit to that artifact. These checks do not establish that every requested feature is correct. 1 published artifact currently has recorded browser issues or layout observations in their metadata.

Ultra, max, xhigh, high, medium, and low are requested settings within each model. The local model catalog describes max as maximum reasoning depth and ultra as maximum reasoning with automatic task delegation. For the ten diagnostics, ultra is the requested label under a fixed single-turn protocol with client delegation disabled. CLI 0.167.0 source and the current model catalog imply that this request maps to xhigh inference effort. This is a source-derived inference, not captured per-call wire evidence. Max remains the highest one-turn reasoning setting in that source/catalog. No client delegation was observed in the ten ultra streams. Provider-resolved effort and provider-internal behavior are not exposed. Effort labels are not a verified ordering of token budgets or elapsed time, and labels across model families do not imply equal compute. Historical comparisons keep their original settings and are labelled as earlier runs.

Recorded incremental cash cost for the new subscription runs is $0. This is distinct from token-based API pricing and does not make the subscription free. No latency or cost ranking is inferred from historical artifacts.

27 published native-setting artifacts use fresh persistent local session history to retain evidence across archive-connection failures. Memory use, memory generation and dedicated memory tools were disabled in CLI configuration; no earlier conversation was resumed. The exact task brief, runner wrapper and requested model/effort settings were retained. This does not establish identical internal context to the earlier ephemeral runs. The two high-effort landing pilots keep their original ephemeral protocol and completion dates.

20 published medium or low artifacts explicitly pin unchanged entries from the official CLI 0.167.0 model catalog snapshot fetched on November 18 at 00:43:58 UTC. Each saved session was checked against the selected model instruction template, requested model and effort, exact prompt, persistent history, read-only settings and disabled memory features. This avoids the earlier dynamic-metadata fallback. The model-specific template differs from the generic high-run template; full internal-context equality across phases and provider-resolved settings remain unverified.

7 published high-effort artifacts used generic client metadata after the CLI could not find model-specific metadata at startup. Saved session records preserve the requested model and high setting; pinned client source passes that explicit effort through. The fallback changes the bundled client instruction template and other model metadata. These are intact first responses with a separate context qualification, not evidence that effort alone explains differences from earlier runs. The exact warning and evidence hashes are shown with each affected artifact.

11 published artifacts overlapped controller sleep. Public generation-time values for those attempts are withheld. Original observed durations remain in the run evidence, but they do not establish uninterrupted model latency. First outputs were preserved; no benchmark attempt was rerun.

A separate delegation-enabled Sol landing pilot returned a complete raw HTML response, but both child-agent spawn requests were blocked by missing session history and no children were created. Its response remains archived off-arena and is excluded from verified delegation-enabled result counts and comparisons. The public attempt record below separates output existence from protocol eligibility.

Missing published artifacts have separate request, protocol or evidence-capture records. Generated-app quality was not assessed for those cells. Available original request and raw-stream evidence is preserved; the benchmark attempt was not rerun.

  • Unpublished attempt record: 01-landing-gpt62-astra-max: Provider capacity and connection errors were followed by the configured runner timeout. No final assistant response was recorded.
  • Unpublished attempt record: 01-landing-gpt62-astra-xhigh: The final response and usage could not be recovered after the controller sleep and archive-connection interruption. The model outcome and whether an output existed remain unknown.
  • Unpublished attempt record: 02-game-gpt62-astra-xhigh: The final response and usage could not be recovered after the controller sleep and archive-connection interruption. The model outcome and whether an output existed remain unknown.
  • Unpublished attempt record: 09-news-gpt62-sol-high: The controller lost its parent process during the daemon restart, terminating the CLI before a final assistant response was recorded. Persistent history contains no final response or terminal usage. The provider-side outcome remains unknown; this is not an app-quality assessment.
GPT-6.2 Sol · max
Fable 5.2 · max
GPT-6.2 Sol at max beside an earlier Fable 5.2 max result for the same published Driftwood brief. Historical request settings remain separate. Open “Landing” in the arena to run both full-screen, go blind, and vote.
More AI news in Polish at nowosci.ai