OpenAI launches GPT-6.2 Astra and Sol with 20 percent lower API prices
OpenAI's GPT-6.2 Astra costs 20 percent less than its predecessor, ties Fable 5.2 on Epoch AI's index, and is the first released model to hit OpenAI's High AI self-improvement threshold.

Why it matters
- OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol; Astra's API pricing is 20 percent lower than its predecessor at $8 per million input tokens.
- GPT-6.2 Astra ranks second in the Artificial Analysis Intelligence Index, five points above GPT-6 Astra, just behind Anthropic's Claude Fable 5.2.
- GPT-6.2 Astra is the first externally released model to reach OpenAI's High threshold for AI self-improvement.
- GPT-6.2 Astra solved 7 of 68 FrontierMath Erdős problems, up from the previous record of 3, each verified in Lean.
- Vals AI prices Sol at $3.18 per task against Astra's $10.68, making Sol the cheapest model in the top 10.
OpenAI released two new language models, GPT-6.2 Astra and GPT-6.2 Sol, and cut the API price of the flagship model by 20 percent: Astra now costs $8 per million input tokens, $0.40 for cached inputs, and $40 per million output tokens. Both models are available now to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and through the OpenAI API. Sol keeps GPT-6.1 Sol's pricing at $2 for input, $0.10 for cached inputs, and $10 per million output tokens.
The release matters for two reasons. First, it puts price pressure directly on Anthropic, whose Claude Fable 5.2 still ranks first in the Artificial Analysis Intelligence Index but at a far higher cost per task. Second, GPT-6.2 Astra is the first externally released model to hit OpenAI's "High" threshold for AI self-improvement, a milestone the company itself flags as a risk.
What do the new models cost and where are they available?
OpenAI positions GPT-6.2 Astra for demanding professional work in science, law, finance, and accounting, and GPT-6.2 Sol for fast, everyday tasks. Pro-500 subscribers can use Ultrafast mode with both models. For the first time, Pro-200 subscribers also get Ultrafast access, starting with GPT-6.2 Sol.
According to Artificial Analysis, an independent evaluation service, OpenAI owns the Pareto frontier on cost versus performance — the line where no other model is both cheaper and better. Only Anthropic and Meta models also land on that frontier. OpenAI holds most of the top spots on efficiency, measured by output tokens per task; Claude Fable 5.2 is the only outside model that appears there.
In the Artificial Analysis Intelligence Index, which rolls up nine evaluations across knowledge work, coding, science, and general knowledge, GPT-6.2 Astra finishes second, five points above GPT-6 Astra and just behind Claude Fable 5.2. It passes Google's Gemini 4 Carbon, at 60 points, and Anthropic's Claude Opus 5.5, at 59. GPT-6.2 Sol picks up two points to reach 57, matching GPT-6 Astra's level at roughly 15 percent of the cost per task.
How strong are Astra's benchmark claims?
OpenAI calls Astra "the world's best model for scientists." On Agents' Last Exam, a benchmark for multi-step professional workflows, the company says GPT-6.2 Astra scores 5.2 points above its predecessor GPT-6 Astra at less than a third of the cost, and leads Claude Fable 5.2 by 5.5 points at 5 percent of the cost. Only Google's Gemini 4 Carbon keeps pace on that benchmark.
On the legal benchmark Harvey LAB-AA, covering tasks such as contract clause analysis, Astra more than doubles Claude Fable 5.2's pass rate at about one-sixth of the cost, according to OpenAI. On the internal GDPval-v2 benchmark, which includes consulting, finance, and design tasks, Astra gains 6.2 percentage points over its predecessor while using 22 percent fewer output tokens.
The model also sets records in scientific work:
- Terminal-Bench-Science, testing scientific workflows in a terminal: 9 percentage points above GPT-6 Astra, again at less than a third of the cost.
- Vals Science Index, measuring interpretation of biological data and experiment execution: a new high, according to OpenAI.
- FrontierMath Erdős, a benchmark of unsolved math problems: 7 of 68 tasks cracked, each formally verified in the Lean proof assistant, up from the previous record of 3 of 68 set by GPT-6.1 Sol. One of the seven solutions had no known answer before.
OpenAI reports gains in computer use as well. In internal testing, the model built a 3D car model, wrote a physics solver in Python, animated the results in Blender, and in another demo modeled a remote-controlled car part by part and produced an animated assembly guide. On the 3D design benchmark AA-3D, where human judges compare blinded interactive experiences, Astra takes first with 1,318 Elo points, ahead of Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259.
Artificial Analysis backs the science claims. Astra leads Terminal-Bench-Science 0.1, places second behind Claude Fable 5.2 on the physics test CritPt v2, the factual benchmark AA-Omniscience, and HLE-Diamond, and leads AA-3D and AA-Robotics v0.1, scoring 79 on robotics against Sol's 78 and Fable 5.2's 75.
Where do the models fall short?
Coding is the weak spot. According to Artificial Analysis, both models have flatlined. In the Coding Agent Index, measured with the Codex harness at max compute, GPT-6.2 Astra stays where GPT-6 Astra was. GPT-6.2 Sol drops three points from GPT-6.1 Sol, driven by a 14-point slide on SWE-Atlas-QnA-AA, where the model omits key facts more often than its predecessor.
Artificial Analysis also flags a high hallucination rate on the AA-Omniscience Hallucination metric, which measures how often a model guesses on factual questions instead of saying it does not know. At maximum effort, GPT-6.2 Astra hallucinates 59 percent of the time, Sol 55 percent. The questions are unusually hard, so the figures do not reflect everyday rates.
Epoch AI, which tested Astra before launch, gives a mixed verdict. The model scores 174 on the Epoch Capabilities Index, tying Fable 5.2; both beat the pre-Astra trend line for reasoning models. But on real-world research tasks, Astra fell behind Fable 5.2 and barely improved over GPT-6 Astra. Epoch AI says the model missed the central question on simpler research projects and "spun its wheels" fixing things. On a task about researching AI data centers, Astra played it too safe and skipped key estimates — something Epoch AI says no contractor could get away with. Epoch AI rated Astra's writing as relatively strong among tested models, but still below its own bar.
Vals AI places Astra 7th and Sol 9th in its index, separated by 1.6 points. Sol runs $3.18 per task, less than a third of Astra's $10.68, making it the cheapest model in the top 10. Astra scores 2.6 points above GPT-6 Astra at roughly half its predecessor's $21.28 per-task cost and uses the fewest output tokens in the top 10, averaging 91,000 per task. Legal work lags: on HLAB the models rank 11th and 21st despite doubling their predecessors' scores, and on Legal Research Astra places 30th and Sol 52nd. The specialized "Astra for Law" product announced in September reaches 6th on the Legal Research Bench.
What about safety and self-improvement?
Both models hit the "Critical" threshold for cybersecurity in OpenAI's Preparedness Framework, matching GPT-6 Astra. OpenAI says they find and exploit vulnerabilities more effectively, write more secure code, and deliver better bug fixes, and is adding stricter safeguards for risky cybersecurity requests in response — safeguards the company says will also limit some legitimate security research. The published cybersecurity numbers come from internal tests with safeguards and classifiers turned off. OpenAI plans to offer the models through OpenAI Daybreak, which gives vetted security researchers access to models with fewer guardrails.
GPT-6.2 Astra is the first externally released model to reach the "High" threshold for AI self-improvement. OpenAI says the model can meaningfully speed up AI research, on par with giving every OpenAI researcher an extra, highly skilled research engineer, and explicitly names this as a risk: models that accelerate AI development could outpace safety measures and human oversight, the concern the industry labels "RSI." Countermeasures include tighter research-environment controls and the Control Room monitoring system, which can track research tasks and shut them down. OpenAI has also expanded monitoring of how models reason and act.
Epoch AI highlights one related capability: Astra runs complex reasoning within a single output token, with no visible chain of thought, at 12.3 serial steps — up from 10.5 for GPT-6 Astra and 4.2 for Claude Fable 5.1. Models that reason without showing their work are harder to monitor. On alignment, 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 search tool.
With Astra matching Fable 5.2 on capability at a fraction of the price, and Sol matching last generation's flagship at 15 percent of the cost per task, the release shifts the frontier-model market toward a price war — while leaving OpenAI to prove that RSI-rated models can stay under meaningful oversight.
Original: deploymentsafety.openai.com
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
728 articles
Related articles
- OpenAI launches GPT-6.2 Astra and Sol, cuts API prices 20%
- OpenAI ships GPT-6.2 Astra and Sol with 20% cheaper API pricing
- OpenAI Launches GPT-6.1 Sol at One-Fifth of Astra's Price
- OpenAI’s GPT-6.2 Astra earns a High self-improvement rating
- GPT-6.1 Terra halves API prices and closes the gap on OpenAI's biggest models