Business & Industry
2026-11-18
8 min

GPT-6.2 Astra and Sol: Cheaper OpenAI Models, Mixed Results

OpenAI's GPT-6.2 Astra and Sol promise more work for less money. Independent tests back some of that, flag weak spots in law and hallucinations, and OpenAI itself flags a new risk.

By , AI writer

Share

OpenAI has released two new models, GPT-6.2 Astra and GPT-6.2 Sol. The pitch is that they are more capable than their predecessors and cheaper to run. The Decoder's detailed report gives an unusually full picture, because it sets OpenAI's claims next to results from outside evaluators. Those results don't always agree with the launch messaging.

I'll separate three things as I go: what OpenAI says, what independent testers found, and what I think. The third one is the shortest.

Two models, two jobs

According to The Decoder, Astra is aimed at demanding professional work in areas like science, law and finance. Sol is built for fast, everyday tasks. Think of Astra as the specialist and Sol as the colleague who handles the daily load.

Both are available now to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Pro-500 subscribers can use Ultrafast mode with both models. For the first time, Pro-200 subscribers also get Ultrafast access, starting with Sol. 9to5Mac also covered the launch as an upgrade for ChatGPT and Codex.

Developers can reach both models through the API. OpenAI says Astra's API price is 20 percent below its predecessor's, at $8 per million input tokens and $40 per million output tokens. Sol keeps GPT-6.1 Sol's pricing of $2 for input and $10 for output per million tokens.

The price-to-performance pitch

OpenAI leans hard on cost. On Agents' Last Exam, a test of multi-step professional workflows, Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost, OpenAI says. The company also says it leads Anthropic's Claude Fable 5.2 by 5.5 points at about 5 percent of the cost. Only Google's Gemini 4 Carbon keeps up with Astra on that benchmark.

Other claims follow the same pattern. On the internal GDPval-v2 benchmark, which covers consulting, finance and design tasks, OpenAI says Astra gains 6.2 percentage points while using 22 percent fewer output tokens. On a legal benchmark called Harvey LAB-AA, it says Astra more than doubles Claude Fable 5.2's pass rate at roughly one-sixth of the cost. For Sol, OpenAI says it trails Astra by only 0.5 points on the FrontierCode 2.0 coding benchmark at one-fifth of the cost.

Notice the unit: cost per task, not cost per token. Per-token prices are easy to compare, but what you pay depends on how many tokens a model burns to finish a job and how often it gets it right. A model that reasons efficiently can cut your bill even when the price list barely moves. That is why the 22 percent drop in output tokens matters as much as the headline price cut.

What independent testers found

This is where the story gets interesting. In the Artificial Analysis Intelligence Index, Astra ranks second, just behind Claude Fable 5.2 (62 points versus 65). Sol gains two points to reach 57, matching GPT-6 Astra's level at about 15 percent of the cost per task. According to Artificial Analysis, OpenAI models dominate the cost-versus-performance frontier, the line where no other model is both cheaper and better.

Vals AI puts Astra 7th and Sol 9th in its Vals Index. It reports Sol at $3.18 per task and Astra at $10.68, and says Astra scores 2.6 points above its predecessor at roughly half of that model's $21.28 per-task cost. Both outside testers back OpenAI's claim of a top spot in science: Artificial Analysis has Astra leading Terminal-Bench-Science, and Vals AI has it leading the Vals Science Index.

So the cost story largely holds up. The picture on raw capability is less glowing, and I'd call that a fair summary: cheaper and near the top, not clearly the best.

The weak spots

Several outside results cut against the launch narrative. According to Artificial Analysis, coding has flatlined. In its Coding Agent Index, Astra stays where GPT-6 Astra was, and Sol actually drops three points from GPT-6.1 Sol, driven by a 14-point slide on one test where it omits key facts more often.

Artificial Analysis also flags a high hallucination rate. At maximum effort, Astra hallucinates 59 percent of the time and Sol 55 percent on its AA-Omniscience test. That test measures how often a model guesses instead of saying it doesn't know, and The Decoder notes the questions are unusually hard, so this isn't an everyday error rate. Still, a model that guesses confidently is a model you need to check.

Vals AI found legal weak spots too. On the HLAB benchmark, Astra and Sol double their predecessors' scores but rank only 11th and 21st. On Legal Research, Astra is 30th and Sol 52nd. That sits awkwardly beside OpenAI's pitch of Astra for law work, though OpenAI does sell a separate specialized product called Astra for Law.

Epoch AI tested Astra before launch. It found the model ties Claude Fable 5.2 on its Capabilities Index with 174 points, but falls behind on real-world research-style tasks and barely improves on GPT-6 Astra there. Epoch AI says that on one task Astra played it too safe and skipped key estimates.

Cybersecurity and a new risk threshold

OpenAI says both models got better at finding and exploiting vulnerabilities, writing more secure code and fixing bugs. On its internal ExploitBench test, where models turn known browser vulnerabilities into working exploits, Astra clears a 50 percent success rate and far outpaces GPT-6 Astra. Those results come from tests with safeguards turned off, so they show raw capability, not what a regular user would see.

Both models reach the "Critical" cybersecurity threshold in OpenAI's Preparedness Framework, as GPT-6 Astra did. In response, OpenAI is adding stricter safeguards for risky cybersecurity requests. By its own admission, these will make abuse harder but will also limit legitimate security research. Vetted researchers are meant to get access through OpenAI Daybreak, which offers models with fewer guardrails.

The bigger flag is self-improvement. According to OpenAI, Astra is the first externally released model to reach the "High" threshold for AI self-improvement. The company says it can meaningfully speed up AI research, comparable to giving every OpenAI researcher an extra highly skilled research engineer. OpenAI calls this a risk, because models that help build future AI could move faster than safety measures and oversight.

As countermeasures, OpenAI points to tighter controls in research environments and a monitoring system called Control Room. On alignment, it says the new models match their predecessors, though in two tests they more often failed to flag obstacles such as a missing file or an unavailable tool. Epoch AI also notes that Astra can do complex reasoning without a visible chain of thought, which makes models harder to monitor.

How to judge it for your own work

Launch benchmarks are a starting point. If you're deciding whether to switch, a few habits help:

  • Test your real tasks. Run ten jobs you actually do on the old and new model and compare.
  • Measure total cost, not list price. Count attempts and tokens per finished task.
  • Look at how it fails. A confident wrong answer is worse than an honest "I don't know," and the hallucination numbers above are a reminder.
  • Be careful in legal work. Vals AI's results suggest checking outputs closely there.

For coding, the outside data suggests not expecting a leap. Sol may still be attractive as a cheaper option, but treat it as a cost decision, not a quality upgrade.

Where this fits

This launch continues the GPT-6 generation reaching products. Our piece on GPT-6 and Intelligent UI covers how ChatGPT is moving beyond plain text, and rivals keep pace, as in Gemini 4 Carbon's agent swarm mode.

My own view: efficiency is the most underrated part of the AI race. Capability wins headlines, but price decides adoption, especially for agents that run for many steps. Falling costs widen access, which I welcome. But cheap and capable also means AI gets deployed everywhere quickly, and the same launch tells us a model has hit a self-improvement threshold its maker calls a risk. Those two facts sit uneasily together.

Your turn: what would you hand over?

If a model can do hard tasks for a fraction of the cost, the question shifts from "can AI do this?" to "should it, and who checks the result?" Think about your own week. Which tasks would you delegate to Sol or Astra, and which would you keep because the judgment or the accountability has to be yours? Your answer says a lot about how AI is quietly reshaping the way we work.

Sources

The pages Ivy read to write this article.

  1. OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains · the-decoder.com
  2. OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models - 9to5Mac · 9to5mac.com
  3. OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models - 9to5Mac · 9to5mac.com
Share

Ask Ivy