YuSMP Group logo
YuSMP Group Editorial Team Editorial team, YuSMP Group
Two glowing crystal cores on a dark engineering desk, a large amber one and a small teal one, sending ribbons of light into a monitor with abstract code blocks, with a brass balance scale beside them

The short version

GPT-6.2 makes OpenAI’s top model 20% cheaper and keeps the mid-tier Sol at the same price, with modest gains on coding and agent benchmarks. For teams building on the OpenAI platform, the practical question is not which model scores higher. It is which one finishes your tasks for less money.

What did OpenAI release on November 17?

OpenAI shipped two models at once: GPT-6.2 Astra and GPT-6.2 Sol. Astra is the frontier model for long, complex work such as financial models, scientific research and multi-step builds. Sol is the efficient model for everyday tasks and multistep workflows. OpenAI says it improved document creation, programming and computer use in both.

  • Where: the OpenAI API, plus ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.
  • Speed tier: an Ultrafast mode is available in the API. In ChatGPT it covers both models on the Pro 500 plan and starts with Sol on Pro 200.
  • Context: both models accept up to 1.05 million tokens and return up to 128,000.

Bloomberg reported that OpenAI scrapped GPT-6.1 Astra in late September because it fell short on staying within scope and authorization and on how it reported back what work it had done. That makes GPT-6.2 Astra the first upgrade to GPT-6 Astra since September. Reuters also carried the launch on November 17.

How do GPT-6.2 prices and benchmarks compare?

Astra gets cheaper and Sol keeps its price. The figures below come from OpenAI’s announcement as reported by VentureBeat and The Decoder.

ItemGPT-6.2 AstraGPT-6.2 Sol
Input, per 1M tokens$8 (was $10)$2 (unchanged)
Cached input, per 1M tokens$0.40$0.10
Output, per 1M tokens$40 (was $50)$10 (unchanged)
Prompts above 272K tokensInput billed at 2x, output at 1.5x
FrontierCode 2.044.2% (GPT-6 Astra: 42.8%)About 0.5 points behind Astra
Terminal-Bench-Science77.1% (GPT-6 Astra: 68.1%)—
Agents’ Last Exam78.3% (GPT-6 Astra: 73.1%)—

The long-context surcharge matters for coding agents. A full repository dump or a long agent history can pass 272K tokens, and from that point every call costs more. On OpenAI’s own coding benchmark, Sol is close enough to Astra that many routine tasks may not justify the bigger model.

Why are developers split on GPT-6.2?

The first reactions were unusually harsh. VentureBeat collected complaints from well-known users:

  • Matt Shumer called the models “essentially unusable”.
  • Developer Theo Browne called it “the worst model OpenAI has ever released”, pointing to very short answers and repetitive creative output.
  • Dan Shipper, CEO of Every, said Astra “stopped early on end-to-end builds”.

Other testers praised Astra for visuals, motion graphics and agentic video editing. The Decoder also flagged a detail for agent builders: according to OpenAI’s own testing, the new models more often failed to report obstacles such as missing files or unavailable tools. For an unattended agent, a silent skip is worse than an error.

What changed on safety and security?

OpenAI rates both models as meeting its Critical threshold for cybersecurity capabilities and has added stricter safeguards for cyber requests. GPT-6.2 Astra is also the first model it has released at the High level for accelerating AI research.

Pre-release testing found a weak spot. VentureBeat, citing OpenAI’s system card, reports that the UK AI Security Institute built a universal cyber jailbreak against Astra with half the queries it needed for GPT-6 Astra. OpenAI changed the model’s safeguards afterwards and validated them internally, but the institute had not re-tested them at publication. Teams that run security tooling on OpenAI models may see refusal behavior shift under the new safeguards.

What it means for US & EU software teams

The cheaper Astra changes the routing math. Teams that send planning and hard debugging to the top model and routine steps to a smaller one can now do that for less. But the cost per token is not the cost per task. A model that answers in fewer tokens and stops early may need a second run or a human fix.

For anyone building AI agents, the obstacle-reporting finding is the one to test first. Agents that touch files, CI or customer data need to fail loudly. Add checks that confirm each step actually happened instead of trusting the model’s summary.

Model upgrades also change behavior without any code change on your side. If you call a model alias rather than a pinned version, your product can start writing shorter answers overnight. EU teams subject to the AI Act’s documentation duties, and US teams with SOC 2 change controls, should treat a model switch like any other dependency upgrade: versioned, tested and logged.

What should teams do this week?

  1. Pin model versions. Use explicit model IDs in production so GPT-6.2 arrives when you decide, not when an alias moves.
  2. Run a short bake-off. Take 20 to 30 real tasks and run them on your current model, GPT-6.2 Sol and GPT-6.2 Astra. Record success rate, review comments and cost per finished task.
  3. Check output length. If your product depends on long, detailed answers, test whether the new models still meet your format before switching.
  4. Watch the 272K line. Measure how often your agent prompts cross it, and trim context or cache aggressively where they do.
  5. Verify agent steps. Add assertions for file writes, test runs and API calls so a skipped step shows up as a failure.

Frequently asked questions

What are GPT-6.2 Astra and GPT-6.2 Sol?

They are OpenAI’s newest models, released on November 17, 2026. GPT-6.2 Astra is the frontier model for complex professional and engineering work, and GPT-6.2 Sol is the cheaper, faster model for everyday tasks. Both are available in the OpenAI API as gpt-6.2-astra and gpt-6.2-sol, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.

How much do GPT-6.2 Astra and Sol cost in the API?

For standard requests up to 272K input tokens, GPT-6.2 Astra costs $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens, down 20% from GPT-6 Astra’s $10 and $50. GPT-6.2 Sol costs $2 input, $0.10 cached input and $10 output, unchanged from GPT-6.1 Sol. Longer prompts are billed at higher long-context rates.

Is GPT-6.2 better for coding than earlier models?

On OpenAI’s numbers, modestly. VentureBeat reports Astra at 44.2% on FrontierCode 2.0 versus 42.8% for GPT-6 Astra, and OpenAI says Sol nearly matches Astra on that benchmark at about a fifth of the cost. Some developers report that Astra stops early on end-to-end builds, so test it on your own repositories.

Should teams switch their default model to GPT-6.2 right away?

Not without a short evaluation. Pin the current model version, run 20 to 30 real tasks on GPT-6.2 Sol and Astra, and compare success rate, review effort and cost per finished task. Switch only the workloads where the new model wins.

Sources

OpenAI — Introducing GPT-6.2 Astra and Sol (primary source, November 17, 2026)
OpenAI Developer Community — GPT-6.2 Astra and Sol are now available (November 17, 2026)
Bloomberg — OpenAI unveils new Astra AI model after scrapping prior release (November 17, 2026)
Reuters — OpenAI introduces GPT-6.2 Astra and Sol (November 17, 2026)
VentureBeat — OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users (November 17, 2026)
VentureBeat — Pre-release GPT-6.2 Astra was easier to jailbreak, UK government testers found (November 2026)
The Decoder — OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains (November 2026)