Coderefercoderefer
CoursesWebinarsBlogAbout
Coderefercoderefer

Discover. Learn. Automate. Grow.

Learn

  • Courses
  • Webinars
  • Blog
  • Search

Company

  • About
  • Contact
  • FAQ

Legal

  • Privacy Policy
  • Terms of Service
  • Refund Policy
© 2026 Coderefer. All rights reserved.
HomeBlogGPT-6.2 Astra and Sol: $8/$40 Pricing, Codex Access, and Opus 5.5 Comparison
ai-toolsNovember 18, 20266 min read

GPT-6.2 Astra and Sol: $8/$40 Pricing, Codex Access, and Opus 5.5 Comparison

GPT-6.2 Astra ($8/$40 per 1M tokens) and Sol ($2/$10) are live in ChatGPT, Codex, and GitHub Copilot. Here's the pricing, the backlash, and whether to switch from Opus 5.5.

V

Vamsi Tallapudi

Manager, Architect Technology at Cognizant

GPT-6.2 Astra and Sol: $8/$40 Pricing, Codex Access, and Opus 5.5 Comparison

OpenAI shipped two new models on November 17, 2026: GPT-6.2 Astra and GPT-6.2 Sol. Astra is the flagship, priced at $8 per million input tokens and $40 per million output tokens through the API, a 20% cut from GPT-6 Astra. Sol is the cheap option at $2/$10 per million tokens, unchanged from GPT-6.1 Sol. Both are live now in ChatGPT, Codex, GitHub Copilot, and the API.

That's the official pitch. What early testers are reporting is messier: cheaper tokens, genuinely strong demo tasks, and also a pile of complaints about models that quit early or drift off-task on anything that takes more than an afternoon.

What are GPT-6.2 Astra and Sol?

Astra and Sol replace GPT-6 Astra and GPT-6.1 Sol as OpenAI's two-tier lineup inside ChatGPT Work and Codex. Astra is pitched as "a frontier model built for your hardest projects" — professional documents, spreadsheets, slide decks, research work, and agentic coding. Sol is the everyday workhorse: multistep tasks, writing, and routine coding at a fraction of the cost.

Both rolled out the same day to Plus, Pro, Enterprise, and Edu users, plus developers through the API. GitHub confirmed the same release date for Copilot access, so this wasn't a staggered rollout — it hit every surface at once.

How much do GPT-6.2 Astra and Sol cost?

Here's the pricing, lined up against the model it replaced and against Claude Opus 5.5 for scale:

ModelInput (per 1M)Output (per 1M)Cached input
GPT-6.2 Astra$8.00$40.00$0.40
GPT-6 Astra (old)$10.00$50.00$1.00
GPT-6.2 Sol$2.00$10.00$0.10
Claude Opus 5.5 (standard)$4.00$20.00$0.20
Claude Opus 5.5 (fast mode)$8.00$40.00—

That last row is the interesting one. Astra's standard rate is exactly what Opus 5.5 charges for its fast mode. At Opus 5.5's normal speed, Anthropic is still half the price of Astra. OpenAI's 20% discount is real, but it's a discount off a model that was already more expensive than its closest Claude competitor.

Where can you actually use them?

Both models are available in ChatGPT Work and Codex right now, and through the OpenAI API as gpt-6.2-astra and gpt-6.2-sol. GitHub Copilot picked them up the same day for Pro+, Business, and Enterprise plans, across VS Code, Visual Studio, the Copilot CLI, GitHub.com, JetBrains IDEs, Xcode, Eclipse, and mobile. Admins can disable either model from the Copilot settings panel, but both ship enabled by default.

There's also a new wrinkle in ChatGPT's speed tiers. Ultrafast mode, which generates tokens at up to 4.5x the standard rate, is now available to Pro 200 subscribers for Sol only, and to Pro 500 subscribers for both models. It burns through your usage allowance faster in exchange for the speed, so it's a trade, not a free upgrade.

What do OpenAI's own benchmarks actually show?

This is where I'd slow down before getting excited. OpenAI's headline numbers compare Astra to GPT-6 Astra (its own predecessor) and to Claude Fable 5.2 — not Claude Opus 5.5, the model most people actually run coding agents on.

  • Agents' Last Exam: Astra beats GPT-6 Astra by 5.2 points at under a third of the cost, and beats Claude Fable 5.2 by 5.5 points at roughly 5% of the per-task cost.
  • Terminal-Bench-Science 0.1: 9 points higher than GPT-6 Astra.
  • ExploitBench (cybersecurity): up 22 points.
  • Harvey LAB-AA (legal): roughly double the pass rate of Claude Fable 5.2 at about one-sixth the cost.
  • Sol lands within half a point of Astra on Cognition's FrontierCode 2.0 coding benchmark, at about a fifth of Astra's cost.

Fable 5.2 is Anthropic's lighter, cheaper Claude tier, not its flagship. Comparing a frontier-priced model against someone else's budget tier and calling it a win is a pretty old trick in this industry, and OpenAI isn't the first to pull it. If Astra actually beat Opus 5.5 head to head, that's the chart I'd expect to see front and center. It isn't there.

Why are power users unhappy?

The HN thread on this launch (and the wider reaction across X) is not kind. A few specifics worth knowing before you flip your default model:

  • Matt Shumer said he wouldn't publish his usual review because the models were "essentially unusable" in his testing.
  • Theo Browne called it the worst model OpenAI has released, pointing to a TypeScript-to-Rust rewrite where Astra spent seven days without improving the code, then another four days after being told to prioritize speed, and still shipped code 56% slower than the original.
  • Ethan Mollick, who had pre-release access, described Astra as "very uneven & weird" and harder to steer than other current models.
  • Multiple HN commenters reported the models stopping early mid-task, refusing to keep going when asked, and occasionally making changes nobody authorized.
  • On a specific benchmark cited in the thread, Sol's pass rate reportedly dropped from 100% (on 6.1) to 93% on 6.2 — a small-looking percentage gap that's actually a sevenfold jump in error rate.

Not every reaction was negative. Raindrop.ai's CTO and a few others praised Astra's work on motion graphics, VFX, and agentic video editing, where it does seem to have made a real jump. So this isn't a universal flop, it's lopsided: good at the flashy creative demos, shaky on the long, grinding agentic coding work that's supposed to be its selling point.

Should you switch from Claude Opus 5.5?

Here's my honest take: not for coding, not yet. The price story sounds good until you check the actual numbers, Astra's standard rate matches Opus 5.5's fast rate, and Opus 5.5 at normal speed is still cheaper per task. More importantly, the pattern in the complaints (stopping early, drifting off instructions, burning days without progress) is exactly what kills a model for agentic work, regardless of what it scores on a one-shot eval. A benchmark win against Fable 5.2 doesn't tell you anything about how Astra behaves on day four of a real refactor.

Sol is a different story. At $2/$10 per million tokens, it's cheap enough to try for routine tickets, docs, and lightweight scripting, and OpenAI's within-half-a-point claim against Astra on coding benchmarks is at least testable without much risk. Just don't expect it to replace Opus 5.5 on anything that needs sustained judgment over a long session.

If you're already deep in the Claude ecosystem and curious how the broader lineup stacks up, our breakdown of Opus 5 pricing and the Opus 5 vs Fable 5 comparison are worth reading before you decide where GPT-6.2 fits. And if you want the history on how OpenAI got here, our GPT-6.1 Terra writeup covers the model this one replaced.

Bottom line

GPT-6.2 Astra and Sol are cheaper than their predecessors, and Sol in particular looks like a solid budget pick for everyday coding and writing. Astra, though, is landing with the same complaint every OpenAI "frontier" release has gotten lately: strong on curated benchmarks, inconsistent on real multi-day work. If you're on Opus 5.5 for serious agentic coding, I wouldn't move. If you're paying full price for GPT-6 Astra today, at least the switch to 6.2 is a straight 20% discount for roughly the same (or better, in spots) model.

Watch

GPT-6.2 Astra and GPT-6.2 Sol — Models built for the real world
ai-tools openai chatgpt 2026

Comments

    Get AI tricks that save you hours every week

    New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.

    Read Next

    MAI Code 2 vs Claude Opus 5.5 in GitHub Copilot: Worth Switching?
    ai-tools

    MAI Code 2 vs Claude Opus 5.5 in GitHub Copilot: Worth Switching?

    MAI Code 2 claims 96% of Claude Opus 5.5's coding performance at 5% of the price in GitHub Copilot. We checked the benchmarks, the promo pricing, and real-world reports.

    November 18, 20267 min read
    Java4Shell (CVE-2026-114938): How to Check Exposure and Patch Java 17-27
    ai-tools

    Java4Shell (CVE-2026-114938): How to Check Exposure and Patch Java 17-27

    Java4Shell (CVE-2026-114938) is a critical OpenJDK Base64 bug hitting Java 17, 21, 25 and 27 on AVX-512 CPUs. How to check exposure and patch it today.

    November 18, 20267 min read
    Windows 11 Tracks Whether You Use Claude Code, Codex, or Cursor: How to Check and Turn It Off
    ai-tools

    Windows 11 Tracks Whether You Use Claude Code, Codex, or Cursor: How to Check and Turn It Off

    A researcher found Windows 11 checks for claude.exe, codex.exe, cursor.exe and copilot.exe and reports which ones you run. Here's how to check and disable it.

    November 17, 20267 min read
    View all posts →