Pull to refresh
Logo
OpenAI releases GPT-6.2 Astra and Sol with cheaper frontier pricing

OpenAI releases GPT-6.2 Astra and Sol with cheaper frontier pricing

New Capabilities

Astra's token prices fall 20% while Sol keeps GPT-6.1 Sol rates, and both ship with tighter cybersecurity safeguards

Overview

Updated 1 hour ago

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. Astra is built for long, multistep professional work; Sol is the cheaper model for everyday tasks.

Astra's input and output token prices are 20% lower than GPT-6 Astra's, and its cached input costs 60% less. OpenAI says the model beats its predecessor by 5.2 points on Agents' Last Exam, a test of long professional workflows, at less than a third of the cost.

Why it matters

Cheaper frontier tokens lower the cost of automating long professional tasks, and the new cyber safeguards decide which security work the models can do.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

$8
Astra input price per million tokens
Down 20% from GPT-6 Astra's input rate.
$40
Astra output price per million tokens
Down 20% from GPT-6 Astra's output rate.
$2
Sol input price per million tokens
Unchanged from GPT-6.1 Sol's input rate.
20%
Cut to Astra's token prices
Applies to both input and output tokens versus GPT-6 Astra.
5.2
Astra's point lead over GPT-6 Astra on Agents' Last Exam
OpenAI says the gain comes at less than a third of the cost.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

People Involved

Organizations Involved

Timeline

2026 November 2026

4 events Latest: Today
Tap a bar to jump to that date
  1. OpenAI launches GPT-6.2 Astra and Sol

    Today Product launch

    Astra targets complex professional work; Sol targets everyday tasks at a lower price.

  2. Microsoft Foundry lists both models as generally available

    Today Distribution

    Azure customers can access Astra and Sol through Microsoft's AI platform.

  3. GitHub adds both models to coding and code review tools

    Today Distribution

    GitHub says early testing caught more real code issues than earlier GPT-6 models.

  4. OpenAI scraps GPT-6.1 Astra over safety concerns

    Safety

    Internal researchers raised concerns, and OpenAI dropped the model less than two months before the 6.2 launch.

Scenarios

1

OpenAI brings GPT-6.2 to Daybreak cyber defenders

Likely Resolves by Q1 2027

Discussed by: OpenAI's launch post, which says the company will work to bring the models to Daybreak.

OpenAI currently gives defenders reduced cyber refusals through Daybreak Blue and advanced testing through Daybreak Red, using GPT-6 Sol and GPT-6 Cyber. The launch post says GPT-6.2 access is planned but gives no date. If it ships, security teams get the new models without the standard refusals that the launch imposed.

2

Independent leaderboard confirms Astra's first-place finish

Possible Resolves by Feb 28, 2027

Discussed by: OpenAI's benchmark table; independent leaderboard results will show whether the claimed lead holds.

OpenAI says Astra beats GPT-6 Astra by 5.2 points on Agents' Last Exam and leads Claude Fable 5.2 by 5.5 points at 5% of the cost. Those figures come from OpenAI's own testing configuration. A public leaderboard placement would confirm the ranking under outside rules.

3

A rival model retakes the benchmark lead

Possible Resolves by Q2 2027

Discussed by: OpenAI's own comparison to Claude Fable 5.2; no rival response has been announced.

OpenAI measured Astra against Claude Fable 5.2 and reported a 5.5-point lead at 5% of the cost. Anthropic, Google, or another lab could answer with a model that closes the gap. A rival launch would reset both the capability comparison and the price-per-task math that OpenAI is selling.

4

Security researchers document blocked vulnerability work

Possible Resolves by Jan 31, 2027

Discussed by: OpenAI's safety section and trade coverage note that the safeguards will restrict some legitimate research.

OpenAI tightened safeguards for higher-risk cybersecurity requests. It says the protections reduce misuse but also limit legitimate vulnerability research and security testing. If researchers show the classifiers blocking ordinary work, OpenAI faces pressure to narrow the refusals.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

February-November 2019

GPT-2 staged release (2019)

OpenAI announced GPT-2 in February 2019 and initially withheld the full model, citing concern that it could generate misleading text at scale. The company released smaller versions first and the full model in November 2019.

Then

Researchers criticized the decision as overstated and hard to enforce, since the code and smaller models were public.

Now

Staged release became a common pattern for labs introducing systems with potential misuse risks.

Why this matters now

GPT-6.2's cybersecurity safeguards follow the same logic: ship the model, restrict the highest-risk uses, and adjust the limits later.

March 2024

Claude 3 model tiers (2024)

Anthropic released Claude 3 in three sizes, Opus, Sonnet, and Haiku, priced and positioned for different workloads. The lineup let customers trade cost against capability within one model generation.

Then

Customers could pick a cheaper tier without switching model families.

Now

Tiered lineups became standard across AI labs, including OpenAI's own model families.

Why this matters now

GPT-6.2 Astra and Sol follow the same split: a frontier model and a cheaper sibling sharing one generation and one context limit.

January 2025

DeepSeek-R1 release (2025)

DeepSeek released R1, a reasoning model with open weights and API prices well below those of US labs. The release briefly wiped hundreds of billions of dollars from AI hardware stocks.

Then

US labs faced new questions about pricing and model efficiency.

Now

Cheaper reasoning models pushed price cuts across the industry and made cost per task a standard comparison.

Why this matters now

GPT-6.2's lower token prices and OpenAI's cost-per-task claims belong to the same competition over how cheaply a model can finish a task.

Sources

(10)