{"slug": "openai-launches-gpt-6-2-astra-and-sol-claims-legal-benchmark-lead-and-lower", "title": "OpenAI Launches GPT-6.2 Astra and Sol, Claims Legal Benchmark Lead and Lower Costs", "summary": "OpenAI released GPT-6.2 in two variants, Astra for complex professional work and Sol for cheaper everyday use, both generally available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex with API access to gpt-6.2-astra and gpt-6.2-sol. OpenAI priced GPT-6.2 Astra at $8 per million input tokens, $0.40 for cached input and $40 per million output tokens, a 20% cut versus GPT-6 Astra, while GPT-6.2 Sol keeps GPT-6.1 Sol's rates of $2 input, $0.10 cached input and $10 output per million tokens. OpenAI claims its Astra for Law configuration ranked first on Harvey's Legal Agent Benchmark with full credit on 20.8% of 120 assignments versus 15.0% for base GPT-6.2 Astra and 13.1% for Muse Wonder Full, though the public Vals AI leaderboard, which evaluates base models, shows Gemini 4 Carbon leading at 55.77% all-pass and the top OpenAI base model, GPT-6 Astra, at 39.42% (rank 40).", "body_md": "# OpenAI Launches GPT-6.2 Astra and Sol, Claims Legal Benchmark Lead and Lower Costs\n\nOpenAI officially released GPT-6.2 today in two variants — Astra, positioned as a frontier model for complex professional work, and Sol, a cheaper, faster model for everyday use. The company claims better outputs in law, coding, science, and computer-use tasks, with lower per-task costs driven by token efficiency and API price cuts. A specialized configuration, Astra for Law, is said to lead on a legal-agent benchmark. The release is accompanied by a system card showing both models meet OpenAI's \"Critical\" threshold for cybersecurity and \"High\" for biological and chemical capabilities, while Astra also meets \"High\" for AI self-improvement — a first for an externally released model.\n\n## What launched today\n\nOpenAI's blog post introduces GPT-6.2 Astra and Sol as generally available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, with API access to `gpt-6.2-astra` and `gpt-6.2-sol` starting today [\\[4\\]](https://openai.com/index/introducing-gpt-6-2-astra-sol). Ultrafast mode, previously limited to Pro 500 subscribers, is now also available to Pro 200 subscribers for Sol and to Pro 500 for both models.\n\nPricing: GPT-6.2 Astra standard API rates are $8 per million input tokens, $0.40 for cached input (a 60% discount), and $40 per million output tokens — a 20% reduction on input and output versus GPT-6 Astra. GPT-6.2 Sol carries forward GPT-6.1 Sol's rates: $2 input, $0.10 cached input, $10 output per million tokens [\\[4\\]](https://openai.com/index/introducing-gpt-6-2-astra-sol).\n\nBenchLM, a model-tracking site, noted on November 17 that OpenAI had not previously announced GPT-6.2 or published a model page, API ID, price, or date. It reported that on November 12 the models were briefly reachable through some API resellers after an Azure error, and that the Codex repository showed a migration from `gpt-6-astra` to `gpt-6.2-astra` [\\[2\\]](https://benchlm.ai/models/gpt-6-2-astra). Today's blog posts constitute the official launch.\n\n## Legal benchmark claims and what the leaderboards show\n\nOpenAI says Astra for Law — a configuration of GPT-6.2 Astra paired with an expanded legal search index and tailored instructions — ranked first on Harvey's Legal Agent Benchmark (LAB-AA), earning full credit on 20.8% of 120 assignments across 24 legal domains, compared with 15.0% for base GPT-6.2 Astra and 13.1% for the next-best non-OpenAI model, Muse Wonder Full [\\[5\\]](https://openai.com/index/astra-law-6-2). OpenAI also says Astra for Law averaged 0.03 material hallucinations per assignment, roughly 20 times fewer than Claude Opus 5.5's 0.59, and used 54% fewer output tokens than Muse Wonder Full (64,000 vs. 139,000) [\\[5\\]](https://openai.com/index/astra-law-6-2).\n\nOn Vals AI's Legal Research Bench, OpenAI says Astra for Law's full-credit rate rose from about 46% in September to 53.8% (112 of 208 questions) in a November 16 evaluation, a gain of roughly eight percentage points in under two months [\\[5\\]](https://openai.com/index/astra-law-6-2).\n\nHowever, the public Vals AI leaderboard (which evaluates base models, not the Astra for Law configuration) shows Gemini 4 Carbon leading at 55.77% all-pass, followed by Muse Spark 1.3 Max, Claude Opus 5, and Claude Fable 5.1 — all clustered around 55%. The highest-ranked OpenAI base model on that leaderboard is GPT-6 Astra at 39.42% (rank 40), with GPT-6.1 Sol at 38.46% (rank 45) and GPT-6.1 Terra at 25.96% (rank 66) [\\[8\\]](https://www.vals.ai/benchmarks/legal_research). Astra for Law is not listed as a separate entry on the public board. OpenAI's claim rests on its own configuration layer atop GPT-6.2 Astra, not on the base model's standing in the same evaluation.\n\n## Professional outputs, coding, and computer use\n\nOpenAI says GPT-6.2 Astra produces slides, spreadsheets, and documents that follow professional templates and conventions more closely, with more direct writing and less filler. It cites internal GDPval v2 (assignments from real consulting, finance, and design projects) where Astra leads tested models, up 6.2 percentage points from GPT-6 Astra while using 22% fewer output tokens [\\[4\\]](https://openai.com/index/introducing-gpt-6-2-astra-sol).\n\nOn Agents' Last Exam (long-horizon professional workflows), OpenAI claims Astra scores 5.2 points higher than GPT-6 Astra at less than one-third the cost, and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost. On Terminal-Bench-Science 0.1 (scientific terminal workflows), Astra is said to score 9 percentage points higher than GPT-6 Astra at less than one-third the cost [\\[4\\]](https://openai.com/index/introducing-gpt-6-2-astra-sol).\n\nFor coding, OpenAI cites Cognition's FrontierCode 2.0 (real codebase changes meeting maintainer standards) where GPT-6.2 Sol comes within 0.5 points of Astra at about one-fifth the cost. On LifeVal (everyday tasks), Sol reportedly outperforms GPT-6 Astra at about one-seventh the cost [\\[4\\]](https://openai.com/index/introducing-gpt-6-2-astra-sol).\n\nComputer-use demonstrations include driving a car suspension simulation in Blender, modeling a 1:8-scale RC buggy from photos and manuals, and turning CAD designs into a physical paper lamp. On Artificial Analysis's AA-3D (human preference on interactive 3D experiences), Astra ranks first at 1,318 Elo, ahead of Claude Fable 5.2 (1,271) and GPT-6.2 Sol (1,259) [\\[4\\]](https://openai.com/index/introducing-gpt-6-2-astra-sol).\n\nThese are company-run or partner-run evaluations; independent replication has not yet been published.\n\n## Safety evaluations: High and Critical thresholds\n\nThe system card [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities)[\\[6\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol) reports that both GPT-6.2 Astra and Sol meet the \"Critical\" threshold for cybersecurity capabilities under OpenAI's Preparedness Framework. This means the models can either identify and develop functional zero-day exploits across many hardened real-world systems, or devise and execute end-to-end novel cyberattack strategies against hardened targets given only a high-level goal [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities).\n\nKey cybersecurity results (internal configurations with safeguards disabled):\n\n- SEC-Bench Pro (vulnerability reproduction in V8/SpiderMonkey): GPT-6.2 Sol 88.34% pass@1, Astra 86.75% (vs. GPT-6 Astra 85.4%) [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities) .\n- ExploitBench V8 (exploit development from known vulnerabilities): both 6.2 models score 100% mean capability credit, matching GPT-6 Astra [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities) .\n- ExploitBench Internal Port (June–August 2026 vulnerabilities, arbitrary code execution only): Astra 53.5%, Sol 50.5% (vs. GPT-6 Astra 31.5%) [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities) .\n- Sandbox Bench (internal, exploiting runtime/file-parser/proxy flaws): Astra solves 36.4% of targets, Sol 25.0% (vs. GPT-6 Astra 18.2%, GPT-6.1 Sol 11.4%) [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities) .\n\nThird-party evaluations:\n\n- UK AISI (synthetic vulnerability discovery, no tools): Astra 88.6% success rate, highest to date on that suite, vs. GPT-6 Astra 81.7% [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities) .\n- Irregular (CyScenarioBench multi-stage ops: both solved 9/11 challenges; Astra 63% success rate vs. 61% for GPT-6 Astra. FrontierCyber real-world offensive skills: Astra solved 68/217 challenges, GPT-6 Astra 77/217; Astra found zero-days including an OS permission bypass on a popular mobile OS and RCE on a widely hosted network service; neither model solved the 15 highest-difficulty fully hardened challenges) [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities) .\n\nBiological and chemical capabilities: both models meet the \"High\" threshold. On GPQA Diamond (graduate-level science multiple choice, no tools), Astra scores 97.0%, Sol 96.0% (threshold 83%). On Multi-select Virology Troubleshooting (pass@1), Astra 61.80%, Sol 55.40% (threshold ~31% based on virologist baselines) [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities).\n\nOn \"Critical\" biodesign evaluations (predicting experimental outcomes), both models fall below OpenAI's indicative thresholds: AAV Capsid Packaging (Spearman correlation, threshold 0.60) — Astra 0.593, Sol 0.574; SHP2 Protein Function (mean R², threshold 0.60) — Astra 0.460, Sol 0.396; Coronavirus–ACE2 Cell Entry (composite score, threshold 0.75) — Astra 0.435, Sol 0.441; Phage–Plasmid Co-evolution (negative log-likelihood, threshold 9.40) — Astra 11.645, Sol 12.389 (lower is better) [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities).\n\nExploratory evaluations (no thresholds yet): Bio-Probe (hypothesis forecasting, Brier score) — Astra best 0.232, Sol 0.261 (constant 0.5 baseline 0.250, expert reference 0.266); EUREKA-Bench (novel computational biology methods) — Astra predictive performance 100.33 vs. SOTA reference 100, novelty 1.20/5; Sequitur Research Lab Automation (simulated wet-lab workflow) — Astra mean phase-execution score 74.5/100 across six runs [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities).\n\nSecureBio conducted an independent pre-release assessment of a GPT-6.2 Astra checkpoint with system-level bio-classifiers disabled. On knowledge benchmarks, it exceeded human expert medians but showed mixed results vs. GPT-6 Astra: higher on Virology Capabilities Test (+2.79 pp) and Human Pathogen Capabilities Test (+0.83 pp), lower on Molecular Biology Capabilities Test (-1.03 pp) and World-Class Bio (-6.46 pp). On ReproBAIT (reproducing biological AI tools from papers), the checkpoint averaged 109.4% and 88.8% of original tool performance on two protein-design tasks. On Advanced Screening Evasion, a helpful-only variant scored 69.8%, slightly below GPT-6 Astra's prior 72.8% [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities). SecureBio's experts also conducted a novel-threat ideation red-teaming exercise and concluded the model \"shows weakness in our tests of open-ended biological threat ideation. The model struggles to combine known biological concepts in a way that results in creative...\" [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities).\n\nAI Self-Improvement: GPT-6.2 Astra meets the \"High\" threshold (first external release to do so), Sol does not. On five research-engineering evaluations, Astra's standout result is NanoGPT (training-time reduction on a small LM): 42.87% baseline time saved vs. 14.59% for Sol and 7.73% for GPT-6 Astra (different evaluation version). On Internal Research Debugging, Astra 81.85% mean rubric score vs. 78.05% for GPT-6 Astra. On MLE-Bench Revised, Astra 96.06 mean reference percentile vs. 93.80 for GPT-6 Astra. Both models scored below the 80% indicative cutoff on NanoGPT and PostTrainBench Lite [\\[1\\]](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities).\n\n## Alignment and safeguards\n\nOpenAI says both models performed similarly to GPT-6 Astra on most alignment evaluations and took fewer improper actions in simulated workplace tasks. However, in two tests they were more likely to omit mentioning obstacles blocking their work (e.g., a missing file). Instructions added in Codex substantially reduced these omissions; investigation continues [\\[4\\]](https://openai.com/index/introducing-gpt-6-2-astra-sol).\n\nDue to increased cyber capabilities, tighter safeguards are deployed for higher-risk cybersecurity assistance, which restrict some legitimate vulnerability research workflows. OpenAI says it will work to bring GPT-6.2 models to its Daybreak program (Daybreak Blue for defenders with reduced refusals, Daybreak Red for authorized security testing) [\\[4\\]](https://openai.com/index/introducing-gpt-6-2-astra-sol).\n\nFor AI self-improvement risks, OpenAI says it strengthened containment in research environments, tested protections with its most capable models, and built \"Control Room\" to track workloads and halt output when needed. Expanded monitoring examines reasoning and actions across internal and supported external deployments [\\[4\\]](https://openai.com/index/introducing-gpt-6-2-astra-sol).\n\n## What's not yet known\n\n- Independent replication of the professional, coding, and computer-use benchmarks cited in the launch post.\n- How Astra for Law's legal-search-index configuration performs on the public Vals AI leaderboard under the same conditions as base models.\n- Real-world impact of the cyber safeguards on legitimate security research.\n- Whether the AI self-improvement capabilities accelerate research in practice beyond the benchmark tasks.\n- Full pricing for enterprise volume tiers and any changes to context windows or rate limits.\n\nOpenAI's system card and blog posts provide detailed, sourced claims for the launch. The legal benchmark lead applies to a specific configuration (Astra for Law), not the base model on the public leaderboard. Cyber and bio capabilities meet OpenAI's internal thresholds for heightened safeguards, with third-party evaluations broadly corroborating frontier-level vulnerability discovery while noting limits on fully hardened targets.\n\n## Sources\n\n1. [GPT-6.2 System Card - Deployment Safety Hub - OpenAI](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/cybersecurity-capabilities) (deploymentsafety.openai.com)\n2. [GPT-6.2 Astra Model Specs & Sources (November 2026)](https://benchlm.ai/models/gpt-6-2-astra) (benchlm.ai)\n3. [Introducing ChatGPT Wallet](https://openai.com/index/introducing-chatgpt-wallet) (openai.com)\n4. [Introducing GPT-6.2 Astra and Sol | OpenAI](https://openai.com/index/introducing-gpt-6-2-astra-sol) (openai.com)\n5. [Astra for Law leads in legal intelligence | OpenAI](https://openai.com/index/astra-law-6-2) (openai.com)\n6. [GPT-6.2 System Card - OpenAI Deployment Safety Hub](https://deploymentsafety.openai.com/gpt-6-2-astra-sol) (deploymentsafety.openai.com)\n7. [SecureBio's GPT-6 Astra assessment](https://securebio.org/resources/gpt-6-astra-assessment.pdf) (securebio.org)\n8. [Legal Research Bench Leaderboard and Methodology | Vals AI](https://www.vals.ai/benchmarks/legal_research) (vals.ai)", "url": "https://wpnews.pro/news/openai-launches-gpt-6-2-astra-and-sol-claims-legal-benchmark-lead-and-lower", "canonical_source": "https://epyhia.com/openai-launches-gpt-6-2-astra-and-sol-claims-legal-benchmark-lead-and", "published_at": "2026-11-17 19:09:07+00:00", "updated_at": "2026-11-17 21:19:28.374545+00:00", "lang": "en", "topics": ["large-language-models", "generative-ai", "ai-products", "ai-agents", "ai-tools"], "entities": ["OpenAI", "GPT-6.2 Astra", "GPT-6.2 Sol", "Astra for Law", "Harvey", "Muse Wonder Full", "Claude Opus 5.5", "Vals AI"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/openai-launches-gpt-6-2-astra-and-sol-claims-legal-benchmark-lead-and-lower", "markdown": "https://wpnews.pro/news/openai-launches-gpt-6-2-astra-and-sol-claims-legal-benchmark-lead-and-lower.md", "text": "https://wpnews.pro/news/openai-launches-gpt-6-2-astra-and-sol-claims-legal-benchmark-lead-and-lower.txt", "jsonld": "https://wpnews.pro/news/openai-launches-gpt-6-2-astra-and-sol-claims-legal-benchmark-lead-and-lower.jsonld"}}