OpenAI ships GPT-6.2 Astra and Sol with 20% cheaper API pricing
OpenAI released GPT-6.2 Astra and Sol, cutting Astra API prices 20 percent. Astra ranks second on the Artificial Analysis index and tops cost-performance charts, but testers flag coding flatlines and legal weaknesses.

Why it matters
- GPT-6.2 Astra API costs $8 per million input tokens and $40 per million output tokens, 20 percent below its predecessor.
- Astra ranks second in the Artificial Analysis Intelligence Index, behind Anthropic's Claude Fable 5.2.
- GPT-6.2 Astra is the first released model to reach OpenAI's 'High' threshold for AI self-improvement.
- At maximum effort, Astra hallucinates 59 percent of the time on AA-Omniscience; Sol 55 percent.
- On FrontierMath Erdős, Astra solved 7 of 68 unsolved problems, up from the previous record of 3.
OpenAI cut API pricing for GPT-6.2 Astra by 20 percent and released the model alongside GPT-6.2 Sol, aiming both at Anthropic's flagship Claude Fable 5.2 on performance while undercutting it sharply on cost. The two models are available now to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, with API access for developers.
Astra targets demanding professional work in science, law, and finance. Sol serves fast, everyday tasks. OpenAI says both beat their predecessors on benchmarks and cost less to run. Astra's API pricing is $8 per million input tokens, $0.40 for cached inputs, and $40 per million output tokens. Sol keeps GPT-6.1 Sol's pricing: $2 input, $0.10 cached, $10 output per million tokens. Pro-500 subscribers can use both models in Ultrafast mode; for the first time, Pro-200 subscribers also get Ultrafast access, starting with Sol.
How do the new models perform on benchmarks?
OpenAI's announcement leans on price-to-performance. On Agents' Last Exam, a benchmark for multi-step professional workflows, GPT-6.2 Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost, according to OpenAI. It leads Anthropic's Claude Fable 5.2 by 5.5 points at just 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, OpenAI says. On the internal GDPval-v2 benchmark, spanning consulting, finance, and design tasks, Astra gains 6.2 percentage points over its predecessor while using 22 percent fewer output tokens.
OpenAI calls Astra "the world's best model for scientists." On Terminal-Bench-Science, which tests scientific workflows in a terminal, the model scores 9 percentage points above GPT-6 Astra, again at less than a third of the cost. Astra also sets a new high on the Vals Science Index, which measures how well models interpret biological data and run experiments.
The model also improved at computer use, already a strength of its predecessor. In OpenAI's testing, it built a 3D car model, wrote a physics solver in Python, and animated the results in Blender. In another demo, it modeled a remote-controlled car part by part and produced an animated assembly guide.
On the 3D benchmark AA-3D, where human judges compare two blinded interactive experiences, GPT-6.2 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.
Where does Astra rank overall?
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, just behind Claude Fable 5.2. Astra jumps five points over GPT-6 Astra, passing Google's Gemini 4 Carbon (60 points) and Anthropic's Claude Opus 5.5 (59 points). Sol gains two points to reach 57, matching GPT-6 Astra's level at roughly 15 percent of the cost per task.
Artificial Analysis finds that 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 efficiency spots, measured by output tokens per task; Claude Fable 5.2 is the only outside model appearing there.
Artificial Analysis also confirms gains in science and factual knowledge. Astra leads Terminal-Bench-Science 0.1 and places second behind Claude Fable 5.2 on the physics reasoning test CritPt v2, the factual benchmark AA-Omniscience, and HLE-Diamond. Astra also leads in 3D design (AA-3D) and robotics (AA-Robotics v0.1), scoring 79 on robotics against Sol's 78 and Claude Fable 5.2's 75.
What did independent testers find?
Coding is the weak spot. According to Artificial Analysis, both models have flatlined there. 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 from answers more often than its predecessor did.
Artificial Analysis also flags high hallucination rates. At maximum effort, GPT-6.2 Astra hallucinates 59 percent of the time on AA-Omniscience Hallucination Rate, Sol 55 percent. The metric measures how often a model guesses on factual questions instead of saying it does not know. Artificial Analysis notes the questions are unusually hard, so the figures do not reflect everyday rates.
Epoch AI tested Astra before launch. The model scores 174 on the Epoch Capabilities Index, tying Claude Fable 5.2; both beat the previous trend line for reasoning models. 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 on a task about researching AI data centers, it 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 below its own bar.
One finding stands out: Astra runs complex reasoning within a single output token, with no visible chain of thought. Epoch AI clocks 12.3 serial steps for Astra, up from 10.5 for GPT-6 Astra and 4.2 for Claude Fable 5.1, the strongest non-OpenAI model at the time. Epoch AI notes that models reasoning invisibly are harder to monitor, a concern raised with the first Astra.
On FrontierMath Erdős, a benchmark of unsolved math problems, Astra solved 7 of 68 tasks, each formally verified in the proof assistant Lean. The previous record was 3 of 68, set by GPT-6.1 Sol. One of the seven had no known solution before.
Vals AI ranks Astra 7th and Sol 9th in the Vals Index, separated by 1.6 points — but the price gap is wide. 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 improved to 2.6 points above GPT-6 Astra at roughly half of that model's $21.28 per-task cost, and uses the fewest output tokens in the top 10, averaging 91,000 per task. Vals AI backs Astra's top spot on the Vals Science Index and first place on Terminal-Bench Science and VCB 1–100, at about half the cost and runtime of GPT-6 Astra. On Terminal-Bench 4 for agentic coding, Astra places 5th and Sol 7th.
Legal tests expose gaps. On HLAB, Astra and Sol double their predecessors' scores but rank only 11th and 21st. On Legal Research, Astra sits 30th and Sol 52nd. Vals AI says the models make nearly twice as many tool calls as their predecessors while writing shorter answers. In September, OpenAI announced "Astra for Law," a specialized product that reaches 6th place on the Legal Research Bench. Vals AI tested with a one-million-token context window and 128,000 max output tokens at full reasoning effort.
What are the safety implications?
OpenAI says GPT-6.2 Astra is the first externally released model to reach the "High" threshold for AI self-improvement in its Preparedness Framework. The company says the model can meaningfully speed up AI research, on par with giving every OpenAI researcher an extra, highly skilled research engineer.
OpenAI flags this as a risk: models that help build future AI systems could outpace safety measures and human oversight, an issue the industry debates under the label "RSI." Countermeasures include tighter controls in research environments and the Control Room monitoring system, which can track research tasks and shut them down. OpenAI says it expanded monitoring that examines how models reason and act and automatically flags unauthorized behavior.
Both models also improved at cybersecurity, finding and exploiting vulnerabilities more effectively, writing more secure code, and delivering better bug fixes with higher token efficiency. They hit the "Critical" threshold for cybersecurity in OpenAI's Preparedness Framework, as GPT-6 Astra did. OpenAI is adding stricter safeguards for risky cybersecurity requests — safeguards it says will also limit legitimate vulnerability 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.
On alignment, the new models match their predecessors, OpenAI says, though in two tests they more often failed to flag obstacles blocking their work, such as a missing file or an unavailable search tool. With Astra now reasoning invisibly at 12.3 serial steps and reaching the "High" self-improvement threshold, OpenAI's monitoring and Daybreak access policies will carry as much weight as the benchmark scores.
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 Launches GPT-6.1 Sol at One-Fifth of Astra's Price
- OpenAI launches GPT-6.2 Astra and Sol with 20 percent lower API prices
- OpenAI Ships GPT-6.1 Sol at a Fifth of Astra's Price
- OpenAI ships GPT-5 to developers at $1.25 per million input tokens