SecondSourceJudgment rebuilt from primary sources

OpenAI releases GPT-6.2, cutting the flagship model's list price 20% from GPT-6 Astra

Concept illustration: two stacks of flat red slabs stand side by side on a grey-brown surface, the taller one on the left with a white laptop at its base whose screen reads 'GPT-6 ASTRA', and the shorter one on the right with a white laptop at its base whose screen reads 'GPT-6.2 ASTRA'. Between the stacks, higher up, an interlaced knot-shaped emblem sits beside the lettering 'OPENAI', against a background that is off-white on the left and pale pink on the right. Printed on the image: -20% OpenAI's GPT-6.2 Astra API list price, input and output, vs the previous GPT-6 Astra.

On November 17, OpenAI released its flagship GPT-6.2 Astra and its everyday model GPT-6.2 Sol (OpenAI, 2026-11-17). The flagship's list price now matches Anthropic's Claude Fable 5.2, and the claim that it leads Fable 5.2 comes only from OpenAI's own scores.

GPT-6.2 Astra's API list price is US$8 per million input tokens and US$40 per million output tokens, 20% below the previous GPT-6 Astra. Independent developer Simon Willison points out that the new prices are the same as Claude Fable 5.2's (Simon Willison, 2026-11-17).

Most of the capability comparisons come from benchmarks OpenAI chose. On Agents' Last Exam, which tests multi-step professional work, GPT-6.2 Astra scores 5.5 points above Claude Fable 5.2. The independent research group Epoch AI combines several of its benchmarks into one capability index, ECI, and gives GPT-6.2 Astra 174, the same as Fable 5.2 (Epoch AI, 2026-11-17).

OpenAI's safety document lists three regressions; one is that both models more often fail to report obstacles they run into during a task (OpenAI system card, 2026-11-17).

Not proven yet: every score that beats Fable 5.2 is OpenAI's own How we checked →

Expert view · Every (a tech publication that publishes hands-on tests of AI models; interests: no public disclosure found): "No one on our team who tried it switched back to it. Astra 6.2 is a real step back for writing." (Every, 2026-11-17)

Every score that beats Fable 5.2 is OpenAI's own, and the benchmark operators' leaderboards haven't been checked. OpenAI also says GPT-6.2 Astra's cost per task on Agents' Last Exam is 5% of Fable 5.2's. The release page notes that Fable 5.2's results exclude 9 tasks where a fallback model was used. The ECI figure is from Epoch AI's post on X; Epoch says it was measured before release and may change.

The three regressions are from the original text of OpenAI's system card. Besides failing more often to report obstacles, GPT-6.2 Astra resisted jailbreaks less well than GPT-6 Astra in initial testing. And at lower reasoning settings, the model was better at evading the safety team's monitoring, which works by reading the model's reasoning. Under OpenAI's own framework, the same document rates GPT-6.2 Astra High for "AI self-improvement"; OpenAI calls that "an early milestone for proactively preparing safeguards for any models that may in future be capable of fully automated, recursive self-improvement," not a sign that the model can already fully improve its successor on its own.

Willison also writes that GPT-6.2 Astra scores below Fable 5.2 on several benchmarks, without naming them. Every's words are from Every's post on X, the impressions of a 30-person team that tried the model.

Next to watch: the Agents' Last Exam operator's leaderboard and Epoch's formally published ECI. If the leaderboard also shows GPT-6.2 Astra ahead, the "self-reported only" limit can be dropped.

Source and checks

This item ran in the brief of Nov 18, 2026, alongside 14 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.

Cite this item

SecondSource (2026). "OpenAI releases GPT-6.2, cutting the flagship model's list price 20% from GPT-6 Astra." https://secondsource.io/en/news/2026-11-18-3/. Accessed access date.