Evidence note: this article separates three kinds of evidence and labels each one where it appears. Vendor and customer announcements are reported as announcements, and none of them were independently cross-checked for this article. External evaluations published by benchmark organizations are named as such, with their publishers. Figures drawn from PlayAgit’s and Vision’s own records carry the date range and the sample size inside the sentence, and those are internal measurements, not independent verification of anyone’s product claims.
"Welcome to the AGI era": the launch and the safety documentation
OpenAI’s GPT-6 Astra was announced on 2026-09-03, with same-day coverage from CNBC and Axios and a model page on OpenAI’s own site. A Wikipedia entry has since been created for the model.
Two things widely repeated about that launch remain open here. The first is the framing: the phrase "Welcome to the AGI era" has circulated as OpenAI’s own characterization of the debut, and that attribution was not cross-checked for this article. The second is the safety documentation: a system card for the model is said to have been published on OpenAI’s deployment-safety site, and that too was not verified. Evidence level for both: uncorroborated launch-week attribution, open. Neither should be treated as established, and neither is used as a premise anywhere below.
What OpenAI itself says the model does
Alongside the launch coverage, OpenAI published a rollout description positioning the model for work across code, browsers, and professional applications. Two capability claims sit in it. The first is computer use: the model is described as working through applications directly, without an API. The second is advanced reasoning across complex professional tasks.
Evidence level: provider-published product description, not independently confirmed; open. Both are worth stating plainly because they are the claims everything downstream is measured against, and neither is a benchmark result. The question an implementer actually faces is whether either description survives contact with one workflow’s constraints, and a product page is not evidence either way.
The external scoreboards: a coding agent index and an ARC Prize evaluation
Two evaluations from outside OpenAI appeared in the same week. Artificial Analysis published a benchmarking write-up of GPT-6 Astra, which includes its Coding Agent Index. The ARC Prize organization published its own evaluation of the model against ARC-AGI-3.
No specific score from either is restated here. PlayAgit’s records contain no cross-check of those numbers, and a benchmark figure repeated without a second source is just the first source again. Evidence level: published by the evaluating organizations themselves; the individual figures were not independently confirmed for this article. What the two publications do establish is that independent evaluation happened quickly, which is more than most model launches get in week one.
What each evaluator says about the scope of its own finding is a different matter, and it can be reported without restating a number. Artificial Analysis reports gains on its Coding Agent Index over GPT-5.6 Sol, while the same breakdown reports weaker performance on DeepSWE. ARC Prize reports results under more than one execution harness and distinguishes between them, which makes the harness part of the finding rather than a detail of how it was obtained. Evidence level: each evaluator’s own characterization of its own published work; the underlying figures remain uncorroborated here. Both reservations point the same way: an aggregate improvement is not a uniform improvement, and a reader who takes one headline number as the model’s coding ability is reading past what the publishers themselves wrote.
What Perplexity is said to have stopped doing
An OpenAI customer page describes Perplexity’s use of the model. The account circulating from it is specific and unusually broad: that Perplexity has the model writing communications, changing software, and monitoring production systems, and that the company checks in on it far less often than it did with earlier models.
Evidence level: vendor-published customer account, not independently corroborated; open. The supervision claim is the load-bearing one and deserves its own reservation. "Checks in less often" is a statement about a company’s internal review habits, which no external party can observe and no benchmark can measure. It is the kind of claim that is repeated far more confidently than it is evidenced, and nothing reviewed here confirms it.
Cognition’s claim is about proof, not output
A second customer page covers Cognition’s use of the model in Devin. The reported claim is worth reading carefully, because it is not the usual one. It is not that the agent writes more code. It is that the agent is better at testing its own work and demonstrating that the work functions, with the stated goal that engineers review less code and ship more.
Evidence level: vendor-published customer account, not independently corroborated; open. If it held up, it would matter more than a throughput number. The distance between an agent finishing and an agent proving it finished correctly is the gap that makes review expensive, a point examined in PlayAgit’s earlier study of 137 runs where changing a game’s source was not the same as delivering it. But the claim arrives without published test data, and it stays open.
Where the model is reachable: OpenRouter, financial services, and code review
Three availability claims circulated in the first days. The model is listed on the OpenRouter platform. OpenAI is reported to have introduced ChatGPT for Financial Services, pairing built-in financial data with the model for research, modeling, and client-ready materials. CodeRabbit published an evaluation of the model in code review covering gains, privacy, and cost.
Evidence level for all three: announcement and vendor-blog claims, not cross-checked; open. Availability claims are usually the easiest to confirm and the least interesting to get wrong, but they were not confirmed here, so they are listed as reported rather than as fact. CodeRabbit’s own framing is worth carrying across with the link: it calls its review results early and directional, and reports a larger advantage on difficult cross-file reviews than across its overall review sample. That is a task-specific claim, and it should keep that scope rather than be read as a general code-review gain.
What builders published in the first days
A cluster of public repositories appeared quickly, and their existence is easier to observe than their contents are to verify. They include work applying a computer-use capability to Minecraft gameplay, single-file HTML games generated in one shot, broad project collections spanning web apps, games, creative tools, and developer utilities, a website-generation collection, an advisor project, a skill package, and curated lists of use cases, prompts, integrations, evaluations, and safety notes (one, another). Reports also describe the model driving robot arms.
Evidence level: repository and demonstration claims, none reproduced or measured for this article; all open. A repository that says a game was generated in one shot is not a record that it was. Whether these artifacts hold up under a second attempt is exactly what nobody published, and readers interested in how that distinction plays out in practice may find PlayAgit’s developer-facing measurements of the model more useful than the demo lists.
Twelve runs, 2026-09-09 to 2026-09-10
Against all of that, here is the far smaller thing that was actually measured. In Vision’s own operator runs touching GPT-6 or Astra across 2026-09-09 to 2026-09-10, twelve runs were recorded, all twelve on the same command-line tool, the Codex CLI.
| Measure | Value |
|---|---|
| Runs | 12 |
| Dates | 2026-09-09 to 2026-09-10 |
| Wall-clock range | 0.4 to 20.7 minutes |
| Median wall-clock | 4.0 minutes |
| Timed runs behind the median | 12 |
Source: Vision’s own operator run records for those two days, n=12. This is an internal measurement of one deployment, not independent verification of any vendor claim. The spread is the finding, not the median. A fifty-fold range between the shortest and longest run means a single average would describe almost none of them.
Three errors, nine finished runs, 43 changed paths
Of the twelve runs, three ended in an error. The recorded error values were item.completed once, ResourceBudgetExceeded once, and once a message about Code Mode being unavailable after a failure to spawn the code-mode process, which the record stores truncated. The nine runs that finished left changes behind: 43 changed paths in total, and 9 commits.
Source: Vision’s own run records for 2026-09-09 to 2026-09-10, n=12, of which 9 completed. A 25 percent error rate over twelve runs is a weak estimate, and the three failures are not one failure repeated. One is a resource ceiling, one is a tool-spawn failure, and one is a bare event name where an error message should be. That last one is a record-keeping defect as much as a run outcome.
What the run records could not say about themselves
The honest limit on everything above is attribution. Vision’s own gap ledger holds a run of measured-and-fixed entries dated 2026-09-09 that name GPT-6 or Astra; four of them bear directly on how the runs above should be read, and they are described one at a time below rather than counted, because the count depends entirely on whether an entry has to name the model in its title or merely in its body. The first records that Vision neither chose nor wrote down which model each of its two command-line providers was actually running. The fix added an explicit per-provider model setting and made each run record carry the model it used.
That fix landed on the same day the twelve runs began. The consequence is unavoidable: these twelve runs were selected because they touch GPT-6 or Astra, not because each record proves which model produced it. Treating them as clean per-model measurements would overstate them. The same shape of problem appears in PlayAgit’s study of 303 agent runs whose records could not say whether anyone played the games: the measurement is only as good as what the log was built to capture.
It is worth naming what these aggregates therefore cannot be asked, because the questions are the obvious ones. They cannot yield an Astra-only success rate, a speed comparison against another model, or a cost per correct outcome. Every one of those needs confirmed model execution, comparable tasks, costs, and a validated result joined together at the level of the individual run, and these records join none of them. The totals describe activity and artifacts in one operating environment. Read as anything more, they would be answering a question nobody measured.
The second correction of the same week: a scoreboard that counted held work as failure
The second ledger entry from 2026-09-09 is about scoring rather than attribution. Vision’s provider scoreboard had been counting held work with the same weight as failure. Work that was completed but deliberately not landed was being folded into the denominator as though it had broken. The fix removed held work from the application-rate denominator and narrowed the gate to work that actually failed.
Source: Vision’s own gap ledger, 2 measured-and-fixed entries naming GPT-6 or Astra, both 2026-09-09. The reason this belongs in an adoption report is that it was a scoreboard producing a real number that meant something other than what it appeared to mean. Any comparison published from it before that date understated the providers it scored. PlayAgit’s review of efficiency claims against Vision’s agent logs ran into the same class of problem from the other direction.
Two more corrections from the same day: a comparison and a matcher
The same 2026-09-09 run of ledger entries holds two more that bear on the figures above. One records the first side-by-side provider comparison Vision ran: the same set of tasks given to Claude Fable 5.1 and to GPT-6 Astra, with the outputs kept. The other records a defect in the matcher that decides which of Vision’s own runs count as material for an article — it read a number carrying a unit as though it were a product name, and attached an unrelated run to an article as supporting evidence.
Source: Vision’s own gap ledger, entries dated 2026-09-09. The second one is the more uncomfortable of the pair, because it is a fault in the evidence-gathering step rather than in the work being measured: for as long as it stood, an article could carry a run that had nothing to do with its subject. The first matters for a different reason. It is the only record here in which the model was given a task next to another model on the same terms, which is the one shape of measurement that none of the run aggregates above can be turned into.
The claims that carry no measurement
Three further claims circulated in the first week with no measurement attached to any of them: that the model is being used in multi-agent orchestration setups with dynamic subagents; that extracted system prompts for it and for Codex are published in a public leaks repository alongside those of Claude Fable 5.1, Opus 5, and Gemini models; and that technical commentary connects the model to looped transformers and hidden reasoning.
Evidence level: all three uncorroborated and open. The architecture commentary deserves the sharpest reservation, because it is an inference about a system whose internals the vendor has not published. Interesting as analysis, it is not evidence about how the model is built.
What this asks of every adoption report
PlayAgit’s own pipeline supplies the last figure, and it is the least flattering one. Across 10 pipeline runs from 2026-09-02 to 2026-09-11, 14 distinct subjects naming GPT-6 or Astra were researched into 31 dossiers holding 245 claims. Of those 245 claims, 21 were cross-checked. The quality gate reviewed 14 drafts and published 13, sent 1 back to revise, and rejected none.
Source: PlayAgit’s own pipeline records for 2026-09-02 to 2026-09-11, n=10 runs. Under nine percent of gathered claims survived a cross-check, which is why almost every vendor and community claim in this article is marked open rather than reported as fact.
Implementation Note
Everything below is from PlayAgit’s and Vision’s own records. None of it verifies a vendor claim, and each figure carries what produced it.
- Timing. Vision’s own operator runs touching GPT-6 or Astra across 2026-09-09 to 2026-09-10 numbered 12, all on the Codex command-line tool, with wall-clock times from 0.4 to 20.7 minutes and a median of 4.0 minutes across all 12 timed runs. Planning around the median would have been wrong for the tail.
- Failures. In those same 12 runs over those two days, 3 ended in an error: one resource-budget ceiling, one failure to spawn the code-mode process, and one that recorded only an event name where a message belonged.
- Output. Of the same 12 runs, 9 left changes behind, totalling 43 changed paths and 9 commits.
- Corrections. Vision’s own gap ledger, 2026-09-09: four measured-and-fixed entries naming GPT-6 or Astra bear on the figures above — provider models that were never chosen or recorded, a scoreboard counting held work as failure, the first same-task comparison between Claude Fable 5.1 and GPT-6 Astra, and a material matcher that read a number with a unit as a product name.
- Sourcing discipline. PlayAgit’s own pipeline records for 2026-09-02 to 2026-09-11, across 10 runs, show 14 subjects researched into 31 dossiers holding 245 claims, 21 of them cross-checked, and 14 drafts reviewed by the quality gate: 13 published, 1 returned for revision, 0 rejected.
The practical rule these figures support is narrow. Record which model ran before counting anything about it, because a fix dated the same day the runs began means the runs cannot fully speak for the model. And check what a scoreboard’s denominator contains before quoting its output, because a metric can be arithmetically correct and still describe the wrong thing.