OpenAI's GPT-6.2 Astra and Sol Beat Benchmarks and Cross a Self-Improvement Threshold, but Power Users Call Astra Unusable
OpenAI's GPT-6.2 Astra tops independent benchmarks and crosses a new AI self-improvement threshold, but early power users call it unusable for long agentic work.
Background
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, November 17, 2026, pairing new benchmark scores with a 20% API price cut on Astra, according to VentureBeat. Astra now costs $8 per million uncached input tokens and $40 per million output tokens, while Sol keeps its predecessor’s $2 input and $10 output rates, according to VentureBeat. Both are proprietary, hosted models available now in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and through the OpenAI API as gpt-6.2-astra and gpt-6.2-sol, according to VentureBeat. Pro-500 subscribers get an Ultrafast tier on both models, and for the first time Pro-200 subscribers get Ultrafast access too, starting with Sol, according to The Decoder.
The launch landed in a crowded moment for frontier models. It comes eight days after Anthropic, in the run-up to going public, launched the well-reviewed Claude Fable 5.2 and Mythos 5.2 at the same $8 and $40 per million standard token rates, and a week after Google opened access to its own well-received Gemini 4 Carbon at introductory rates equal to Sol’s, according to VentureBeat.
Current situation
On OpenAI’s own numbers, the upgrade looks clean: in the nine head-to-head benchmark charts in its launch materials, Astra beats GPT-6 Astra on all nine, and Sol beats GPT-6.1 Sol on eight of nine, with the Vals Science Index the lone exception, according to VentureBeat. On Terminal-Bench-Science 0.1, a 70-task test of research work in a terminal, OpenAI reports Astra gained nine percentage points over its predecessor at less than one-third the cost, reaching roughly 77%, and on the legal benchmark Harvey LAB-AA, OpenAI says Astra more than doubled Fable’s pass rate at roughly one-sixth the cost, according to VentureBeat. On Agents’ Last Exam, OpenAI reports Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost, and leads Fable 5.2 by 5.5 points at 5% of the cost, according to The Decoder.
Independent evaluator Artificial Analysis largely backs the headline numbers. Astra is OpenAI’s highest-scoring model yet on the Artificial Analysis Intelligence Index, climbing five points from GPT-6 Astra to rank second overall, behind only Claude Fable 5.2, and ahead of Gemini 4 Carbon (60) and Claude Opus 5.5 (59), according to Artificial Analysis. Sol gained two points to reach 57, matching GPT-6 Astra’s level at roughly 85% less cost per task, according to Artificial Analysis. Artificial Analysis also puts OpenAI across most of the cost-per-task Pareto frontier, sharing it only with Anthropic and Meta, and has Astra leading in 3D design (1,318 Elo points on AA-3D, ahead of Fable 5.2’s 1,271 and Sol’s 1,259) and robotics (79 on AA-Robotics v0.1, versus Sol’s 78 and Fable 5.2’s 75), according to Artificial Analysis.
Coding agent work is where the independent numbers diverge from the launch framing. Artificial Analysis found Astra flat against GPT-6 Astra on its Coding Agent Index and Sol down three points from GPT-6.1 Sol, a decline traced to a 14-point drop on SWE-Atlas-QnA-AA, where Sol more often omits key facts when answering questions about codebases, according to Artificial Analysis. The Decoder reports the same flat-to-lower pattern independently, and adds that research organization Epoch AI found Astra tying Fable 5.2 on its own Epoch Capabilities Index at 174 points but falling behind Fable 5.2 on Epoch’s real-world research tasks, where the model reportedly missed the central question on simpler projects and, on one assignment about AI data centers, skipped estimates Epoch AI says “no contractor could get away with,” according to The Decoder. Artificial Analysis separately measured a high hallucination rate on its hardest factual-knowledge questions: 59% for Astra and 55% for Sol at maximum reasoning effort, according to Artificial Analysis.
That gap between chart-topping scores and agentic reliability shows up directly in how early testers describe using the models. Investor and former HyperWrite AI CEO Matt Shumer said the new models “were essentially unusable” and that, unlike with prior OpenAI flagships, he would not publish his customary longer review, according to VentureBeat. Developer and YouTuber Theo Browne called Astra “the worst model OpenAI has ever released,” singling out its handling of delegated work and creative tasks, according to VentureBeat. In a follow-up post, Browne described asking Astra, independently of another tester, to design new levels for the games Outer Wilds and Majora’s Mask; both runs produced a water level centered on music and bells, with the same name, “Bellwether,” according to VentureBeat. On an experimental TypeScript-to-Rust compiler rewrite, Browne said Astra spent seven days without landing an improvement and produced a version that ran 56% slower after focusing on cutting its size, then spent another four days without a change once redirected toward speed, according to VentureBeat. Browne did credit Astra with catching code-review issues other models missed and with helping design a reverse proxy he expects will save $25,000 a month, but said he still preferred Claude Opus 5.5 for his development work, according to VentureBeat.
Not every account was negative. Wharton professor Ethan Mollick had Astra build a playable space elevator in Blender and Unity, but found the model “very uneven & weird” and harder to steer than rival models, according to VentureBeat. Dan Shipper, CEO of generative-AI publication Every, said his team’s testing found Astra performing about as well as its predecessor on most coding benchmarks but often stopping early on end-to-end builds, and praised its default design instincts while noting its outputs were hard to vary and had a “noticeable house style,” according to VentureBeat. Raindrop.ai CTO Ben Hylak and AI advisor Ashe Magalhaes were more positive, praising Astra’s strengths in animated visuals, motion graphics, VFX, and agentic video editing, and OpenAI says many of its own employees already use Sol as their default model for technical work, according to VentureBeat.
Alongside the capability numbers, OpenAI disclosed a safety milestone: Astra is the first model it has released externally to meet the “High” threshold for AI self-improvement, with OpenAI saying the model can meaningfully speed up AI research, comparable to giving every OpenAI researcher an extra, highly skilled research engineer, according to The Decoder. heise online independently reported the same classification. OpenAI frames this as a risk in its own right, warning that models able to help build future AI systems could outpace the safety measures and human oversight meant to keep up with them, a dynamic the industry labels “RSI,” according to The Decoder. As a countermeasure, OpenAI points to tighter controls in its research environments and a new monitoring system called Control Room, which tracks research tasks and can shut them down, a detail also independently reported by heise online. Both Astra and Sol also meet OpenAI’s Critical threshold for cybersecurity, as GPT-6 Astra did before them, and OpenAI is adding stricter safeguards for risky cybersecurity requests that it says will also limit some legitimate vulnerability research, according to VentureBeat and The Decoder. On alignment, OpenAI says the new models match their predecessors overall, but in two tests they more often failed to flag obstacles blocking their work, such as a missing file or an unavailable search tool, according to The Decoder.
Implications
The spread between OpenAI’s and Artificial Analysis’s charted gains and the power-user accounts points to a specific seam: Astra and Sol post real improvements on single-shot, well-scoped evaluations — legal pass rates, 3D design, robotics, physics reasoning — while independent measures of sustained, multi-step coding-agent work (Artificial Analysis’s Coding Agent Index, Epoch AI’s real-world research tasks, and VentureBeat’s own reporting of flat-to-regressed independent Codex testing) show little gain or outright decline, according to Artificial Analysis and The Decoder. Browne’s and Shumer’s complaints center on exactly that seam — not raw task competence but the model’s behavior over long, delegated sessions, from stalling for days on a single refactor to derailing a game-design task with an idiosyncratic naming pattern repeated across unrelated projects, according to VentureBeat. That distinction matters for how buyers read benchmark charts going forward: a model that leads an index built from short, graded tasks is not necessarily the model that will finish a week-long agentic job without supervision.
The self-improvement disclosure adds a second, separate axis to the launch. OpenAI’s own framing — that Astra is the first model it has shipped externally at the “High” RSI threshold — was confirmed independently by two outlets covering the same system-card language, rather than resting on OpenAI’s announcement alone, according to The Decoder and heise online. OpenAI’s stated response — new research-environment controls and the Control Room monitoring system — is, by the company’s own account, aimed at the research-automation risk rather than at the usability complaints raised by Browne, Shumer, and others, which concern deployed product behavior rather than research-acceleration capability, according to The Decoder. The two threads of this launch — the capability milestone OpenAI is treating as a safety threshold, and the reliability complaints its own early testers are raising about everyday agentic use — are running on separate tracks, and the sources reviewed here do not say whether or how they are connected.