OpenAI’s GPT-6.2 Undercuts Claude at 5% of the Cost

Human Verified 0

Share to: , and

A developer's desk at dusk with two glowing orbs above a laptop showing abstract colorful patterns, a notebook and a steaming mug of coffee

OpenAI launched GPT-6.2 Astra and Sol on Tuesday, pitching frontier work at a fraction of the cost. Astra is the flagship for hard projects, while Sol is the everyday model. Both are live now in ChatGPT Work, Codex and the API.

Two Models, One Pitch

Astra is built for the hardest jobs: slides, spreadsheets and documents that follow a template down to the font size, cleaner code and frontends for websites, gaming and 3D design. OpenAI says it sets a new high on science tests too, including terminal research.

Sol is the efficient sibling, aimed at everyday work and agents. OpenAI says it nearly matches Astra on Cognition’s FrontierCode test at about a fifth of the cost, and beats GPT-6 Astra on everyday tasks at a seventh of the price.

The Numbers OpenAI Is Quoting

On Agents’ Last Exam, OpenAI says Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost. It also claims 5.5 points over Claude Fable 5.2 at 5% of the cost — though Fable’s score excludes nine tasks that used a fallback model.

Astra’s API prices fall 20% from GPT-6 Astra, to $8 per million input tokens and $40 for output, with cached input 60% cheaper at $0.40. Sol keeps GPT-6.1 Sol’s $2 and $10 rates, and both finish tasks with fewer tokens, so users get more done before hitting limits.

Built for Professional Work

Astra is trained to match conventions in law, finance and accounting, with direct writing, less filler. On Harvey’s legal test it more than doubles Fable 5.2’s pass rate at a sixth of the cost, and built a valuation workbook answering 62.5% of questions correctly on all five tries.

Astra’s sharper computer use reaches into CAD and simulation tools, letting engineers check fits and test revisions before building. In 3D design it ranks first on Artificial Analysis’s AA-3D test, ahead of Claude Fable 5.2 and its own sibling Sol.

Stronger Cyber, Tighter Safeguards

Both models mark another jump in cyber capability, finding and exploiting flaws more effectively and writing more secure code. OpenAI calls Astra its most capable model ever for cyber, jumping 22 points over GPT-6 Astra on its internal ExploitBench test.

The stronger cyber skills come with tighter safeguards for higher-risk security help, which OpenAI says can also restrict some legitimate research. Approved defenders get access through OpenAI’s Daybreak programme, and OpenAI says it will bring both 6.2 models to Daybreak as well.

Both models also meet the High threshold for biological and chemical capabilities, and Astra is the first OpenAI model released externally at the High bar for AI self-improvement. In two tests, the models were more likely to skip mentioning obstacles that blocked their work.

Our Read

This is OpenAI answering on price, not just capability. Astra beats Fable 5.2 on OpenAI’s own tests at 5% of the cost, and undercuts GPT-6 Astra by a fifth — days after Microsoft’s Code 2 did the same in coding.

The pattern is now unmistakable: every frontier lab is selling the same quality for less, and the buyer wins. The catch: all these numbers come from the vendors themselves, so independent tests will decide whether the price war is real.

Why It Matters

For paying users, the launch means more work per subscription: Astra for the hardest projects, Sol for everyday tasks, both using fewer tokens. For rivals, the bar rises — Anthropic’s Fable 5.2 now faces a cheaper challenger in its IPO week.

Background

GPT-6.2 arrives two months after GPT-6 Astra’s September debut, days after Microsoft launched its own coding models. OpenAI’s cyber push builds on GPT-6 Cyber, while the models are not in regular ChatGPT — only in Work, Codex and the API.

Source: OpenAI announcement.

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.