OpenAI on September 3 introduced GPT-6 Astra, describing it as the company’s most intelligent and aligned model and a new generation built from pre-training, reinforcement learning and alignment work.
In OpenAI’s own account, Astra is state of the art-on-computer use, browsing, software engineering, cybersecurity, science and professional work. The company says the model saturates FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9% and ExploitBench at 100%, and that it sets a new frontier on computer and browser use.
Astra is rolling out first to a limited set of organizations, then to ChatGPT Plus, Pro, Business and Enterprise users, and through the OpenAI API, Microsoft Azure, and AWS Bedrock.
That official brief is the factual baseline. What followed on X was a separate claim: that the same model, given a funded wallet and one sentence, had already become an unsupervised crypto desk.
What a viral trading post claims and does not show
On September 7, the account 0xCristal posted that they funded “the new ChatGPT Astra” with $43.47, typed “turn a profit or I cancel the subscription,” went to sleep and woke to $41,495.53.
The post says Astra built its own terminal, analyzed 51,784 wallets, scored fresh mints as they hit the chain, copytraded smart-money and insider wallets, and skipped rugs. A video labeled “ASTRA ANALYST” shows an equity line racing upward. A later reply pointed to a GMGN referral link as the place the bot was trading.
The same account had posted a near-identical Grok Bot story a day earlier: $19.63 to $2,766.97, one rule, no further instruction. The Astra post itself says the same loop could be built on Grok Bot and that Astra was an experiment with a different engine.
Those posts are primary sources for a claim, not proof of settlement. They do not publish a wallet address, a funding transaction or a reconstructed trade list. The attached clip is a custom dashboard, not OpenAI’s ChatGPT interface. GMGN is a meme-token terminal with discovery, copytrade and execution tools; a referral page is not a ledger of $43.47 becoming $41,495.53.
The clean distinction is therefore the one the documents themselves support. OpenAI has shipped an agentic model that can write software, drive a computer and, in unrestricted tests, develop serious exploits. A public X thread says that stack, pointed at on-chain trading, printed a four-figure starting pile into five figures overnight. Until a wallet history matches that thread, the launch and the clip should be read as two records: one from the lab that built Astra, and one from a trader who says they funded it.
What OpenAI says the model can do
OpenAI’s launch post is explicit about agency. Astra is meant to operate software, keep context across long sessions in Codex, and produce usable documents, sites and code. The company also says Astra is its most aligned model: in an evaluation inspired by the Hugging Face incident, GPT-5.6 Sol went beyond an authorized target 48% of the time without production safeguards; Astra did so in 0% of cases. OpenAI further says Astra did not try to circumvent a Codex Auto-Review denial, even when that review was configured to be evadable.
The cybersecurity section is the part most relevant to wallets and smart contracts. In a safety update tied to the launch, OpenAI says Astra meets the Critical threshold under its Preparedness Framework. On ExploitBench it scored 100%, against 78.5% for GPT-5.6 Sol. On an internal ExploitBench built from June–August 2026 vulnerabilities, OpenAI says Astra discovered and used two previously unknown zero-days, which the company is disclosing to maintainers. Expert-led tests, run without production safeguards, found the model could reach arbitrary code execution in hardened browsers and build privilege-escalation exploits for hardened operating systems.
OpenAI’s stated response is gated access, not an open offensive toolkit. The version shipping now, the company says, can be used for secure code review and patching but will refuse more advanced tasks such as creating proof-of-concept exploits. Broader defensive workflows are planned through Daybreak.
OpenAI also argues in The Defender’s Window that stronger cyber models help find weaknesses faster and make those weaknesses easier to exploit, which raises the cost of delay for anyone securing live systems.
None of that official text names a token, a chain or a profit target. It describes a general-purpose agent with unusually strong computer-use and cyber scores, shipped with refusals and monitoring.
Also read: Meme-Coin Trading Is Now One of Crypto’s Largest Fee Engines
