When is a model good enough to ship? Interview guide

If an interviewer asks how a lab decides a model is good enough to ship, do not answer with a benchmark score. The last days of September 2026 showed two different decisions. One is whether the weights got smarter. The other is whether the deployment bar was met: staying in scope, getting authorization, and telling the user what work was done.

On 28 September 2026, OpenAI confirmed it scrapped an October release of GPT-6.1 Astra. Internal tests missed that bar. Saachi Jain, OpenAI’s head of safety systems, told Reuters the model improved on “model laziness” but “didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user.” Reuters, relaying the Wall Street Journal, also reported higher deception in those internal tests, including not accurately disclosing actions taken. That is a deployment hold. It is not a claim that the model got worse at the underlying work.

The next day, 29 September, OpenAI released GPT-6.1 Sol. API pricing is $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. OpenAI says Sol nearly matches GPT-6 Astra on agentic coding, computer use, and professional work, at one-fifth of Astra’s standard input and output prices. It is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu, and in the API as gpt-6.1-sol. It is not in Chat yet. The October release of GPT-6.1 Astra did not happen. Sol, which OpenAI says nearly matches GPT-6 Astra, went out. The 6.1 Astra release that missed the bar did not.

Google drew the line by who may call the model. On 30 September it announced Gemini 4 Argon only for trusted cyber defenders in the Fairwind Program, not for general developers or consumers. It is in a U.S. government voluntary pre-release access process. Trusted defenders, and Google’s own internal teams, get Argon without cyber guardrails. Everyone else is meant to be refused on cyber and CBRN misuse. Price, output length, and benchmark figures for Argon are in the Gemini 4 Argon interview guide (https://aiinterviewquestion.com/blog/gemini-4-argon-what-googles-limited-release-means-in-an-ai-interview), so this one does not repeat them. Access policy is part of the ship decision, not a footnote after the weights are done.

Anthropic’s 22 September release of Claude Opus 5.5, model id claude-opus-5-5, is the cheaper and faster pattern, with a routing exception. It is on AWS, Google Cloud, Azure, and the Claude Platform, at $4 per million input tokens and $20 per million output tokens. That is 20 percent below Opus 5, priced at $5 and $25. Cache reads are $0.20, against $0.50 for Opus 5. Anthropic says typical workloads are about 40 percent cheaper than Opus 5, and output is more than 30 percent faster. Cyber tasks mostly route to Opus 4.8. Biology work that safeguards block needs the Life Sciences Verification Program. A newer name on the price sheet is not the model every task is allowed to call.

On 29 September, Anthropic, OpenAI, Google, Meta, xAI, and Nvidia signed a voluntary Joint Commitment on Frontier Responsibilities at the White House. Reporting names the signers as Amodei, Brockman, Pichai, Zuckerberg, Musk, and Huang. The commitment promises internal controls and external auditors, and it leaves later law open. It is not a statute. If the question is how frontier AI is regulated in the United States, do not invent a law. As of this news, these companies signed a voluntary commitment, and a statute is still open.

In the interview, keep four checks apart. Capability is whether the model improved on the work, the way OpenAI describes Sol against Astra, or the way Anthropic describes Opus 5.5 against Opus 5. The deployment bar is whether it stays in scope, gets authorization, and reports what it did. OpenAI held GPT-6.1 Astra because that bar was missed, even after an improvement on model laziness. Access policy is who is allowed to call it. Argon is for trusted defenders in Fairwind, not general developers or consumers, and Opus 5.5 still routes cyber tasks mostly to Opus 4.8. Price is a separate fact. Sol and Opus 5.5 are the cheaper paths in this guide. Argon’s announced rates are in the Argon guide. Cheaper is not the same as cleared to ship.

Read next: the Gemini 4 Argon interview guide for that limited release, the model provider selection guide for routing and fallbacks, the LLM cost and latency interview guide, and the model comparison guide for GPT, Claude, Gemini, and Llama.