Skip-Level

OpenAI Releases GPT 6.2 Astra and Sol with New Safety Thresholds

By ยท 18 Nov 2026

A top-down still life of a large marigold server module and a slim cream server module on a bright papaya cloth, with a closed brass padlock between them and a coiled sky blue cable, with no people and no text.
  • On Nov 17, OpenAI released GPT 6.2 Astra for the hardest professional work and GPT 6.2 Sol for everyday work, and said Astra scores 5.2 points higher than GPT 6 Astra on Agents' Last Exam at less than a third of the cost.
  • OpenAI prices GPT 6.2 Astra at $8 per million input tokens and $40 per million output tokens, and says Astra is its first externally released model to meet a High threshold for AI self improvement, while both models meet a Critical cybersecurity threshold.
  • The same day, Microsoft said MAI Code 2 is generally available in the GitHub Copilot model picker and in Copilot Fusion, and that, averaged over four coding benchmarks, Code 2 delivers 98% of GPT 6.1 Sol quality at 40% of the cost.
  • Black Duck's Nov 17 writeup of its 2027 AI Readiness Report, covering 65.5 million findings across about 13,100 applications, says organizations discovered 18 new findings for every 1 they closed in May 2026.

Three things engineering and IT managers should know this Wednesday, from a new pair of OpenAI models with tighter safety flags to a cheaper Microsoft coding model already sitting in Copilot, and a widening gap between finding vulnerabilities and closing them.

1. OpenAI Releases GPT 6.2 Astra and Sol with New Safety Thresholds

On Nov 17, OpenAI published "Introducing GPT 6.2 Astra and Sol." The post presents GPT 6.2 Astra as a frontier model for the hardest professional projects, including coding, documents, and multi step agent work, and GPT 6.2 Sol as a more efficient model for everyday work. OpenAI's GPT 6.2 system card, also dated Nov 17, says the company is releasing both models that day, and that they are the most capable models it has deployed externally to date.

On Agents' Last Exam, a test of multi step professional work, OpenAI says GPT 6.2 Astra scores 5.2 points higher than GPT 6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost. Those figures are OpenAI's, from the Nov 17 post, not an independent test. In the API, the post says GPT 6.2 Astra input and output prices are 20% lower than GPT 6 Astra, at $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. GPT 6.2 Sol, the post says, keeps GPT 6.1 Sol's API token prices.

The same post, and the system card, say both models meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework. The system card says GPT 6.2 Astra meets the High threshold for AI self improvement, while GPT 6.2 Sol falls below it, and that Astra is OpenAI's first externally released model to meet that High threshold. OpenAI describes a High designation as the equivalent of giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to a 2024 baseline. The system card says that, in response to stronger cybersecurity results, OpenAI tightened default restrictions on higher risk cybersecurity assistance and requires additional vetting for more permissive access. It also says OpenAI improved isolation in its research environments, expanded monitoring of frontier agents, and introduced Control Room, a system that can stop model workloads when safety or security concerns arise.

The Nov 17 post says GPT 6.2 Astra and Sol are available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, with Enterprise and Edu access depending on a workspace admin. Developers can use both models through the OpenAI API.

Why it matters: This is a cost and oversight change, not only a quality bump. Our read: treat OpenAI's benchmark and threshold claims as company assessments, then decide where Astra may run, who may turn on the higher access paths, and how you will log agent work before a cheaper, more capable model starts changing a process your policy has not approved. Sources: OpenAI and OpenAI system card

2. Microsoft Puts MAI Code 2 in GitHub Copilot at a Fraction of Frontier Cost

Also on Nov 17, at Ignite, Microsoft announced MAI Code 2 and MAI Code 2 Flash. In the official Microsoft blog post on the keynote, Chairman and CEO Satya Nadella says Code 2 is a flagship family of agentic coding models for everyday enterprise coding tasks, including taking code to production, deep debugging, multi file work, and repo wide reasoning, at a fraction of frontier cost. Code 2 Flash, he says, is for fast, high volume work such as explanations, targeted edits, routine fixes, and rapid iterations. The same post says both models are generally available that day on GitHub Copilot, in the model picker and as part of Copilot Fusion.

A Nov 17 Microsoft AI post gives the comparison figures. Averaged over four benchmarks that Microsoft says target everyday enterprise coding tasks seen in GitHub Copilot (DeepSWE 1.1, Terminal Bench 2.1, SWE Atlas, and FrontierCode v1.1), Microsoft says MAI Code 2 delivers 98% of the performance of GPT 6.1 Sol at 40% of the cost, and 96% of Claude Opus 5.5's performance at 5% of the cost. The post's notes say all models were run at max reasoning effort, and that the MAI Code 2 cost in that comparison is shown at a 50% promotional price. Microsoft also says these models are not yet tuned for harder problems, such as fundamental science research and very long horizon agentic work.

Why it matters: A cheaper coding model in the picker most developers already use will change volume, review load, and spend at the same time. Our read: treat the 98% and 40% figures as Microsoft's own benchmarks, including the promotional price note, then write down which Copilot model is the default, which changes still need a human review, and how you will watch cost if more of the work shifts into Code 2. Sources: Microsoft and Microsoft AI

3. Black Duck Says Teams Found 18 Security Findings for Every 1 They Closed

This item is a research report, not a product launch. On Nov 17, Black Duck published a blog post on its 2027 AI Readiness Report. The post says the report analyzes 65.5 million application security testing findings across roughly 13,100 applications at 533 organizations.

Black Duck says that in June 2025, organizations discovered 11 new findings for every finding they closed, and that by May 2026 the ratio had risen to 18 to 1, with a peak of 22 to 1 in December 2025. By the end of the study, it says, only about 1 in 10 findings had been closed. At the median, findings identified through static application security testing and software composition analysis took about 30 days to close. The post opens by saying that in 2018 the average time between a vulnerability's disclosure and its first exploit was 2.6 years, and that the window is now less than a day. Critical and high severity findings, it says, were 16.8% of new findings in May 2026, up from 13.1% in June 2025.

Why it matters: The report is about a widening backlog, not a claim that any one tool will close the gap. Our read: if discovery is outrunning fixes, pick a risk based order for what gets closed first, and decide where automated fix tools may draft a change and where a person still has to approve it. Source: Black Duck

Compiled daily with Grok from public sources and edited by the AIDevPulse team.

Get Skip-Level and the daily brief in your inbox: sign up for the AIDevPulse newsletter. For today's developer news, read today's main brief.

Sources:

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.

Prompt

I manage an engineering or IT team. Using only the three items in this readout, help me draft a one page plan for the next month that does three things: decides where GPT 6.2 Astra and Sol may run and what oversight applies when a model meets a High or Critical safety threshold, using only OpenAI's Nov 17 announcement and system card, sets a default and review rule for MAI Code 2 in GitHub Copilot, using Microsoft's Nov 17 posts and treating the 98% and 40% figures as Microsoft's own benchmarks with a promotional price note, and names a risk based order for closing a security backlog, using only Black Duck's Nov 17 report on the 18 to 1 finding ratio. Keep my team's context in mind: [describe your team's size, which models you already allow, and how you review agent work]. Do not add facts that are not in the readout, and flag anything I should check before sharing.