New: GPT-6.2 benchmark report is here.Read the report ↗

GPT-6.2 Astra and Sol: everything you need to know

GPT-6.2 is a strange release. On paper, Astra is OpenAI’s strongest model for complex professional and scientific work. In the first hours after launch, several experienced testers said almost the opposite about using it for everyday work.

My read is that those signals are not mutually exclusive. Astra looks less like a universally better collaborator and more like a model optimized around research workflows, scientific problem-solving, and improving AI systems. That is my inference from the evidence, not a design goal OpenAI has stated.

What OpenAI released

OpenAI launched two related models. GPT-6.2 Astra is the higher-capability model for difficult professional and scientific work. GPT-6.2 Sol is the lower-cost model for everyday agentic workflows. Both expose a 1.05-million-token context window and a 128,000-token maximum output in their current API listings.

Standard API specifications at launch
SpecificationGPT-6.2 AstraGPT-6.2 Sol
Input price / 1M tokens$8$2
Cached input / 1M tokens$0.40$0.10
Output price / 1M tokens$40$10
Context window1,050,0001,050,000
Maximum output128,000128,000
Knowledge cutoffMay 18, 2026June 1, 2026

That makes Sol one quarter of Astra’s standard input and output price. A workload using 100,000 uncached input tokens and 20,000 output tokens would cost about $1.60 on Astra and $0.40 on Sol, before tools, caching, retries, and any premium serving mode.

Why the early reviews are so negative

The launch benchmarks are strong. The first public reviews are not. These are early reports from individual testers, not controlled evaluations, but they matter because they describe the friction benchmarks often miss: whether the model communicates clearly, stays with a task, and feels useful across a full work session.

Matt Shumer

Shumer said he would not publish a full review because his testing experience was so poor.

Dan Shipper

Shipper reported weaker writing, premature stopping on end-to-end coding tasks, and a recognizable visual house style.

Matthew Berman

Berman found the models good, but not competitive with the strongest Claude releases in his testing.

Dan Shipper’s detail is especially useful. His complaint was not that Astra cannot code. It was that it stops early in real projects and often needs another instruction to finish. That difference between isolated capability and sustained work is exactly what a benchmark score can hide.

What the benchmarks say

OpenAI’s launch data tells a much more favorable story. At its strongest reported settings, Astra reached 78.3% on Agents’ Last Exam v1.1, 44.2% on FrontierCode 2.0, and 62.1 on the Artificial Analysis Intelligence Index v5.0.1. Higher reasoning effort generally improved the score, while also increasing cost.

Professional work78.3%

Agents’ Last Exam v1.1

Astra at xhigh effort. OpenAI says this is 5.5 points above Fable 5.2, with a caveat about nine Fable tasks that used a fallback model.

Repository work44.2%

FrontierCode 2.0

Astra at max effort, with a reported average cost of $5.97 per attempted task.

Composite capability62.1

AA Intelligence Index

Astra at max effort across the independent index’s knowledge, coding, science, and general tasks.

Those numbers are not a rebuttal to the negative reviews. They answer a different question. The benchmarks ask whether the model can solve defined evaluations under a particular harness. The reviewers are asking whether they would choose it for actual writing, coding, design, and project work.

Why Astra feels optimized for research

This is the most interesting part of the release. Astra’s strongest and most distinctive evidence clusters around scientific workflows and AI research rather than general collaboration.

  • Scientific work: OpenAI reports new highs on Terminal-Bench-Science 0.1 and the Vals Science Index.
  • Research debugging: Astra scored 81.85% in OpenAI’s system-card evaluation, compared with 74.66% for Sol.
  • AI systems: Astra reached 76.92% on KernelGen and saved 42.87% of the reference time in NanoGPT experiments.
  • Model training: Astra scored 74.79% on PostTrainBench Lite and reached the 96.06th percentile on the revised MLE-Bench evaluation.
  • AI self-improvement: OpenAI classifies Astra as High capability in this category and compares its impact with adding a research engineer.

That pattern makes me think Astra is being optimized for doing research work: formulating and debugging experiments, operating in technical environments, and improving model-training systems. Again, this is an inference. OpenAI describes Astra more broadly as a model for complex professional and scientific work.

If that interpretation is right, the early reaction makes more sense. A model can be unusually strong at scientific and AI research while feeling terse, rigid, or frustrating as a writing partner and product-building collaborator. “Better” has become too broad a word for models this specialized.

Which model should you use?

My default recommendation is Fable 5.2 for most writing, coding, and collaborative work. Choose Astra when research is the job, and Sol when cost is the constraint.

Research pick · Astra

Research, science, and hard technical experiments

Use Astra when the work resembles the evaluations where its evidence is strongest: research debugging, scientific terminal workflows, model training, difficult codebase changes, and long technical investigations.

Budget pick · Sol

Lower-cost agentic work

Use Sol when you need the same large context and a much lower token price. Test it before moving a high-volume workflow, especially if the job depends on sustained follow-through.

Default pick · Fable 5.2

Writing and end-to-end collaboration

Based on the early reports, Fable is the safer first trial for writing, polished communication, and coding tasks where finishing the complete job matters. See the full Astra vs Fable 5.2 report.

The right evaluation is a small set of your own complete tasks. Keep the prompt, source material, tool access, and acceptance criteria fixed. Measure whether the result is usable, how often the model stops early, how much editing it needs, and the total cost after retries.

Safety and limitations

OpenAI classifies both models at Critical capability for cybersecurity and High capability for biological and chemical domains. Astra also crosses the High threshold for AI self-improvement. These classifications describe capability and deployment safeguards, not a claim that ordinary use is unsafe.

The system card also records two practical limitations. Safeguards may restrict legitimate vulnerability research and testing, and in two evaluations the models were more likely to omit obstacles that could block the work. That second behavior is worth watching in research and professional workflows, where an unreported failure condition can invalidate the result.

Sources and methodology

Specifications, launch benchmarks, and positioning come from OpenAI’s launch article and deployment safety report. The public reactions are individual first-day reports, quoted briefly and linked to their original posts. They are included as qualitative evidence, not averaged into a model score.