OpenAI launches GPT-5.6 with Sol, Terra and Luna tiers

Why it matters

GPT-5.6 is OpenAI's clearest commercial push toward agentic work: higher capability on demand, cheaper tiers for scale, and pricing mechanics that force developers to manage tokens like infrastructure costs.

Sam Altman (@sama) used a post on X on July 9th to point users to OpenAI's GPT-5.6 launch post, calling it "the best model we have ever produced" and pairing the claim with one of OpenAI's most detailed recent attempts to explain how it wants frontier AI to be bought, measured and deployed.

The launch brings GPT-5.6 into general availability after a limited preview, with three named model tiers: Sol, the flagship model; Terra, a balanced model for everyday work; and Luna, the lower-cost option. OpenAI says GPT-5.6 is available starting July 9th across ChatGPT, Codex and the OpenAI API, with a global rollout planned over 24 hours.

Altman's post matters because OpenAI is framing GPT-5.6 less as a chatbot upgrade than as an operating layer for long-running work. The company says Sol improves coding, professional knowledge work, cybersecurity and science, while using fewer tokens and reducing estimated cost against some rival and prior models. Those are OpenAI's own claims, delivered through OpenAI's own benchmark presentation, and the post leans heavily on a metric that has become central to the AI buyer's spreadsheet: performance per dollar.

The pricing makes that strategy explicit. OpenAI says GPT-5.6 Sol costs $5 per 1M input tokens and $30 per 1M output tokens. Terra is priced at $2.50 input and $15 output. Luna is priced at $1 input and $6 output. Cache writes are billed at 1.25 times the uncached input rate for GPT-5.6 and later models, while cache reads keep the 90% cached-input discount. OpenAI is also adding explicit cache breakpoints and a 30-minute minimum cache life, a small infrastructure detail that will matter to developers trying to keep agent workflows from becoming unpredictable line items.

The product shift is the new compute ladder. OpenAI says GPT-5.6 supports higher reasoning settings, including max, and adds ultra, a mode that coordinates four agents in parallel by default. In the API, OpenAI says developers can build similar patterns through a multi-agent beta in the Responses API. That puts the company squarely into the market for AI systems that can run through research, coding, document production and other multi-step jobs with less human orchestration.

OpenAI is also pushing Programmatic Tool Calling in the Responses API. The feature lets GPT-5.6 write and run lightweight in-memory programs that coordinate tools, process intermediate data and decide what to do next. The company's pitch is cost control: fewer model round trips, fewer tokens passed back into the model, and less manual scripting by developers. In practice, that is OpenAI trying to make agents cheaper to operate before customers conclude that long-running autonomy is too expensive for production use.

The benchmark table is broad and aggressive. OpenAI says GPT-5.6 Sol reaches 80 on the Artificial Analysis Coding Agent Index, scores 88.8% on Terminal-Bench 2.1 and 72.7% on DeepSWE v1.1. It says Sol reaches 90.4% on BrowseComp, 62.6% on OSWorld 2.0 and 73.5% on ExploitBench. The company also says GPT-5.6 Sol Ultra reaches 92.2% on BrowseComp and 91.9% on Terminal-Bench 2.1. Those figures should be read as OpenAI's launch framing unless and until buyers test the models on their own workloads.

OpenAI included customer comments from Qodo, Rogo, Clio, Notion, Ramp, Shopify, Cisco, Lovable, ModelML, Triple Whale, PlayCo, Canva, Microsoft, Base44 and Legora. The named users are useful because they show where OpenAI wants GPT-5.6 judged: code review, financial research, legal workflows, app generation, presentations, frontend development and office productivity. They also leave the harder commercial question intact. OpenAI gave detailed token pricing and benchmark comparisons, but enterprise adoption will depend on whether those gains survive contact with real permissions, messy files, compliance checks and procurement budgets.

Safety is a large part of the launch post. OpenAI says GPT-5.6 is more capable than earlier models in biology and cybersecurity, while remaining below its Critical threshold in both areas. The company says it ran extensive red teaming before general availability, including about 700,000 A100e GPU hours of black-box automated testing. It also says Sol's cyber safeguards block roughly 10 times more potentially harmful activity than previous models, with more sensitive cyber capabilities reserved for verified users through OpenAI Daybreak's Trusted Access program.

That safety framing gives away the tension inside the release. OpenAI wants GPT-5.6 to be strong enough for defensive security work, code automation and internal AI research, while keeping the most dangerous use cases behind identity checks, monitoring and tiered access. The launch post says GPT-5.6 has already changed how OpenAI itself works: over the past six months, the company says the share of research compute devoted to internal coding inference grew 100-fold, while internal agentic token usage rose about 22-fold.

For founders building on top of OpenAI, the release adds another decision layer. GPT-5.6 is no longer a single model choice. It is a menu of capability tiers, reasoning settings, cache economics and agent patterns. The teams that win cost-sensitive AI workflows will be the ones that know when Sol is worth paying for, when Terra is enough, when Luna can carry the cheap path, and when ultra is a margin problem disguised as a product feature.