📅 2026-07-09 · 12 min read

GPT-5.6 Release Guide: Sol vs Terra vs Luna, Pricing, Benchmarks, and Availability

GPT-5.6 is a three-tier model family spanning flagship reasoning, balanced everyday work, and lower-cost high-throughput use. This guide turns OpenAI's launch data into practical model-selection decisions.

What launched

OpenAI released GPT-5.6 as a three-tier model family rather than one universal model. Sol is the flagship tier for the hardest reasoning and agentic work, Terra is positioned as a balanced model for everyday work at lower cost, and Luna is the fastest and most affordable tier for high-throughput use.

The launch matters because the family spans consumer, coding, and developer workflows at the same time. OpenAI says GPT-5.6 is available across ChatGPT, ChatGPT Work, Codex, and the OpenAI API, with access depending on plan, product surface, and rollout timing.

Availability

OpenAI says the GPT-5.6 rollout starts globally across ChatGPT, Codex, and the API, continuing gradually toward full availability. Access differs by product surface.

• Chat: Plus, Pro, Business, and Enterprise users can access GPT-5.6 Sol through medium and higher effort settings; Pro and Enterprise users can also select Sol Pro for the highest-quality results on complex tasks.
• ChatGPT Work and Codex: Free and Go users can access Terra, while Plus and higher plans can choose among Sol, Terra, and Luna and set effort levels. OpenAI says max is available to users with GPT-5.6 access in these surfaces, while ultra has narrower plan availability.
• API: developers can access Sol, Terra, and Luna. OpenAI also introduced Programmatic Tool Calling and a multi-agent beta for concurrent subagents and synthesis.

Rollout and plan entitlements can change. Check the official product and pricing pages before committing a production workflow to a specific access assumption.

GPT-5.6 pricing

OpenAI lists the following API rates per one million tokens.

GPT-5.6 also changes prompt-caching economics. OpenAI says cache writes for GPT-5.6 and later models are billed at 1.25× the uncached input rate, while cache reads continue to receive a 90% cached-input discount. This makes cache design part of cost modeling for repeated long-context workflows.

Sol vs Terra vs Luna: which model should you test first?

Benchmark context

OpenAI published a broad benchmark table covering professional work, coding, science, computer use, cybersecurity, self-improvement, multimodal tasks, academics, tool use, long context, and abstract reasoning. The key lesson is not that one model wins every row. GPT-5.6 is strongest where agentic work, computer use, coding, and efficiency matter, while competitor rankings vary by benchmark.

The official table also shows important non-wins. For example, GPT-5.6 Sol's SWE-Bench Pro result is 64.6%, below several Claude results in the same table. Benchmark decisions should therefore stay task-specific instead of collapsing into one universal ranking.

Coding and agent workflows

GPT-5.6's biggest practical story is the combination of stronger agentic capability with better efficiency. OpenAI highlights Terminal-Bench 2.1 and multi-agent settings, while early partners report fewer tokens, lower latency, fewer steps, and fewer tool calls on production workflows.

For coding teams, that means model evaluation should measure more than benchmark score. Run one repo task and track accepted changes, test pass rate, retries, tool calls, total tokens, and review burden. Sol may win on the hardest task, but Terra can be the better default if it reaches acceptable quality at substantially lower cost.

Knowledge work, design, and computer use

OpenAI positions GPT-5.6 as a stronger end-to-end operator for documents, presentations, spreadsheets, browsing, and computer-use workflows. The launch page reports Sol at 90.4% on BrowseComp and 62.6% on OSWorld 2.0, with an Ultra result of 92.2% on BrowseComp.

The practical implication is that a model-selection test should include the final artifact, not only the intermediate text. For presentations, spreadsheets, dashboards, and frontend work, inspect layout quality, completeness, adherence to reference files, and the number of correction cycles required before the output is usable.

Safety and cybersecurity

OpenAI says GPT-5.6 launches with its strongest safeguards to date and that the models remain below the company's Critical capability thresholds in biology and cybersecurity. The launch page emphasizes defensive uses such as secure code review, patching, vulnerability triage, threat modeling, malware analysis, and detection engineering.

The cyber benchmark gains are substantial: OpenAI reports Sol at 73.5% on ExploitBench versus 47.9% for GPT-5.5, and 71.2% on SEC-Bench Pro versus 45.8% for GPT-5.5. These results should be read together with access controls, safeguards, and the intended defensive workflow.

Who should use which GPT-5.6 model?

• Choose Sol first for hard coding agents, long-running research, complex computer use, demanding professional analysis, and tasks where failure is expensive.
• Choose Terra first for broad everyday work, scaled agent fleets, knowledge work, and teams that want a better price-performance balance before paying for maximum capability.
• Choose Luna first for high-volume extraction, classification, routing, quick assistance, and latency-sensitive workflows where the lowest cost matters.
• Test max or ultra selectively when a task is genuinely hard enough to justify more reasoning or parallel agents. Do not make the most expensive setting the default without measuring incremental value.

For most teams, the best rollout is a tiered routing strategy: Luna for simple work, Terra as the default, and Sol for escalation. That recommendation is an inference from the official capability and pricing tiers, not a claim that one routing policy fits every workload.

Primary source: OpenAI's official GPT-5.6 launch page. Benchmark results and pricing above are reproduced from the official launch data; benchmark methods differ and should be interpreted within each evaluation context.

FAQ

Is GPT-5.6 one model?

No. OpenAI presents GPT-5.6 as a family with Sol, Terra, and Luna capability tiers, each with different price and performance positioning.

Is GPT-5.6 available in ChatGPT?

Yes. OpenAI says the family is rolling out across ChatGPT, ChatGPT Work, Codex, and the API, with model access depending on plan and product surface.

Which GPT-5.6 model is cheapest?

Luna has the lowest listed API price at $1 per million input tokens and $6 per million output tokens.

Does GPT-5.6 win every coding benchmark?

No. The official launch table shows very strong Terminal-Bench and DeepSWE results, but model ordering varies across benchmarks such as SWE-Bench Pro.