gpt-5.6-sol

OpenAI's official preview post, dated June 26, 2026, announces a limited preview of the GPT-5.6 series with GPT-5.6 Sol positioned as the flagship model, alongside Terra (a balanced everyday-work model) and Luna (a fast, affordable tier). The post explicitly states that Terra delivers competitive performance to GPT-5.5 The post confirms that, at the U.S. government's request, GPT-5.6 Sol, Terra, and Luna were initially previewed only to a small group of trusted partners whose participation was shared with the government, with broader availability planned in the coming weeks. OpenAI frames this as a short-term coordination step tied t

LayerLens's July 14, 2026 benchmark review runs GPT-5.6 Sol, Terra, and Luna through 12 benchmarks on its Stratix harness and reports Sol at 53.6% on Agents' Last Exam, 80 on the Artificial Analysis Coding Agent Index with max reasoning, 88.8% on Terminal-Bench 2.1, and 64.6% on SWE-Bench Pro. The review explicitly nam The piece highlights Sol's long-context performance — 91.5% on MRCR 8-needle recall at 256K–512K tokens within its 1.05M context window — and notes an "Ultra" multi-agent mode unique to Sol that pushes BrowseComp to 92.2% at roughly 4x token cost. Luna is reported at 84.7% on Terminal-Bench 2.1 at $1/$6 per million tok

BenchLM's profile for GPT-5.6 Sol, published July 9, 2026 and updated through September 14, 2026, lists the model at a capability score of 80.7/100 and rank 5 of 232 tracked models. It reports API pricing of $4 per million input tokens and $20 per million output tokens, with cached input at $0.40, and a context window Category percentile values place GPT-5.6 Sol in the 97th percentile for both Agentic and Coding tracks, 98th for Knowledge, 94th for Multimodal, and 42nd for Reasoning, with 48 source-displayable benchmark rows on file. The aggregator provides consolidated technical metadata useful for cross-checking other benchmark so

CodeRabbit's July 9, 2026 engineering review evaluates GPT-5.6 Sol and Terra for coding agents and code review, naming both variants explicitly. The post reports that Sol "follows through better" on messy repo tasks, takes long checklists seriously, and finds more review issues in CodeRabbit's harness, while Terra is p The review notes that Sol does not erase other frontier models — Fable 5 still feels stronger on architectural judgment and Sonnet 5 retains a cleaner comment-quality story in some review workflows — but Sol wins on practical follow-through. CodeRabbit's write-up is a practitioner-oriented assessment useful to engineer

The llm-stats.com model page for GPT-5.6 Sol catalogs aggregated benchmark performance, listing a composite LLM Stats Score of 55.0 at a blended price of $6.19 per million tokens, placing it between GPT-6 Astra (59.9 at $11.9) and Muse Spark 1.3 (54.4 at $0.10). It also reports per-turn LLM Stats Scores across conversa The page lists individual benchmark ranks with sources, including a 1.00/1 rank on OpenAI's internal Connectors production benchmark and a 0.97/1 rank on OpenAI's internal Capture-the-Flag Challenges at max reasoning effort, with HealthBench Consensus also surfaced among the tracked evaluations. The excerpt is truncate

OpenAI announced the public release of its GPT-5.6 series on Thursday, roughly two weeks after a government-coordinated limited rollout to trusted partners, according to CNBC reporting dated July 8, 2026. The release covers three variants — GPT-5.6 Sol, Terra, and Luna — and was framed by OpenAI CEO Sam Altman with a " Alongside the public release, OpenAI introduced a new generation of voice models called GPT-Live, with GPT-Live-1 and GPT-Live-1 mini rolling out globally to ChatGPT users on the same Wednesday. The CNBC piece notes these voice models can listen and speak simultaneously for more conversational interaction. While the vo

Digital Applied published a June 26, 2026 preview guide covering the GPT-5.6 family: Sol as the flagship for ambitious agentic work, Terra as a balanced everyday tier at roughly half of GPT-5.5's cost, and Luna as the fast, low-cost option for high-volume jobs. The guide frames the tiering as a permanent "good, better, The guide reports OpenAI-sourced preview figures including a Terminal-Bench 2.1 score of 91.9 for Sol in ultra mode, pricing held at $5 input / $30 output (the same as GPT-5.5), and red-teaming exceeding 700,000 A100-equivalent GPU hours. It also flags a new max reasoning effort, a multi-agent ultra mode, and more pred