GPT-5.6 Sol

GPT-5.6 Sol is the flagship model of OpenAI's GPT-5.6 family, released to general availability on July 9, 2026 after a limited preview period. The GPT-5.6 family covers three tiers: Sol (flagship, $5/$30 per 1M input/output tokens), Terra (mid-tier, $2.50/$15, competitive with GPT-5.5), and Luna (cheapest and fastest, $1/$6). The Sol/Terra/Luna names are 'durable capability tiers' - they will evolve with subsequent numeric generations (GPT-6, GPT-7). This decoupling of generation from marketing tier is a new positioning move by OpenAI in the frontier competition era.

Performance: Sol sets new state-of-the-art on several key benchmarks: Artificial Analysis Coding Agent Index 80 (SOTA, +2.8 pts over Claude Fable 5 at half the output tokens and 1/3 the cost), Agents' Last Exam 53.6 (+13.1 pts over Fable 5), Terminal-Bench 2.1 88.8% (91.9% in Ultra mode), BrowseComp 92.2% (Ultra), OSWorld 2.0 62.6% (SOTA, surpasses Opus 4.8 at 85 percent fewer tokens), GPQA Diamond 94.6%, ExploitBench 73.5% (vs 47.9% for GPT-5.5). Key differentiator vs Anthropic Claude Fable 5: efficiency - Sol at max reasoning comes within 1 point of Fable 5 on Artificial Analysis Intelligence Index v4.1 in 61 percent shorter time and at half the cost.

A new performance architecture: OpenAI introduced Programmatic Tool Calling (Responses API) - the model writes and runs in-memory lightweight programs that coordinate tools, process intermediate results, monitor progress, and choose the next action. Reduces tool-heavy tasks by up to 38 percent input tokens with no quality loss (Clio data). Additionally three reasoning levels: medium/high (default), xhigh (extended), max (highest - more time to explore alternatives and verify), ultra (coordinates 4 parallel agents by default, up to 16 in the API - higher token use but stronger results and shorter time-to-result).

Applications and integrations: GPT-5.6 is the base model for ChatGPT (Plus/Pro/Business/Enterprise) on medium+ effort, ChatGPT Work / Codex (Plus/Pro/Business/Enterprise), the OpenAI API via the Responses endpoint. Key deployments: Cursor (President Oskar Schulz: 'one of the strongest on CursorBench'), Notion (Simon Last: 'the most tenacious problem-solver, agents on Terra match GPT-5.5 at half the cost'), Microsoft 365 Copilot (preferred model since July 9, 2026), Cognition Devin, Shopify, Ramp, Figma Make, Canva, Cisco AI, Legora, Lovable, Base44, Rogo, Model ML, Basis, Balyasny, Triple Whale, Qodo, PlayCo, Clio. The model is ZDR-compatible (Zero Data Retention) in the Responses API - key for enterprise customers.

Safety and limits: GPT-5.6 received OpenAI's most robust safety stack to date - layered guardrails, real-time checks, monitoring, account-level enforcement, reasoning monitor. In cybersecurity: Sol blocks ~10x more potentially harmful activity vs earlier models but preserves defensive capability via OpenAI Daybreak Trusted Access for Cyber program. The model does NOT cross the Critical threshold in biology or cybersecurity per OpenAI Preparedness Framework. Key limits: (a) closed weights (no self-hosting), (b) pre-release political controversy - per reports the deployment was withheld by the Trump administration on national-security grounds, (c) fine-tuning unavailable at GA launch (custom deployments via enterprise partnership), (d) knowledge cutoff November 2025 (newer events require the web-search tool).