AIニュース：OpenAI・5・6・ローンチ

概要

The biggest AI model release entering July 10, 2026 is OpenAI's general launch of GPT-5.6. The family is available across ChatGPT, Codex, and the OpenAI API after a limited preview, giving users three capability and price tiers instead of a single flagship choice.

The important change is not only a higher benchmark score. OpenAI is positioning GPT-5.6 around end-to-end work: coding, design, research, computer use, tool coordination, and professional documents. For users and developers, the practical question is which tier completes a workflow with the best mix of quality, speed, and total cost.

OpenAI

GPT-5.6 has three models. GPT-5.6 Sol is the flagship for difficult reasoning, coding, cybersecurity, science, and professional work. GPT-5.6 Terra is the balanced option for everyday tasks, while GPT-5.6 Luna is the fastest and most affordable tier.

OpenAI says the family gets more useful work from each token than GPT-5.5. Sol adds stronger computer use and design judgment, while Terra and Luna aim to bring agentic performance to workloads that do not need the most expensive model.

The release also introduces higher-effort modes for complex tasks. The max setting gives a model more time to explore and verify an answer. The ultra setting coordinates multiple agents across parallel workstreams, trading higher token use for stronger results and a faster time to completion on demanding jobs.

開発者

Programmatic Tool Calling lets GPT-5.6 write and run lightweight programs that coordinate tools, filter intermediate results, and decide what information should return to the model. This can reduce repeated model round trips when an agent needs to inspect many records, tool outputs, or files.

The Responses API also adds a multi-agent beta. Developers can run concurrent subagents and combine their results in one request. That makes the release relevant to coding agents, research systems, financial analysis, document workflows, and other jobs that can be split into independent workstreams.

5・6

At launch, GPT-5.6 Sol costs $5 input / $30 output per 1 million tokens. Terra costs $2.50 input and $15 output, while Luna costs $1 input and $6 output per 1 million tokens.

GPT-5.6 is rolling out through ChatGPT, Codex, and the OpenAI API. Access varies by plan: paid ChatGPT users receive Sol access, while ChatGPT Work and Codex expose different combinations of Sol, Terra, Luna, max, and ultra depending on the subscription tier.

For developers, the headline token price is only the starting point. A model that uses fewer tool calls, produces fewer correction loops, or finishes a task with fewer output tokens can be cheaper even when its listed rate is higher.

実用AIワークフロー

• Compare Sol, Terra, and Luna on the same real task instead of relying only on model benchmarks.
• Measure completion quality, latency, output tokens, tool calls, and review time together.
• Use Sol for difficult coding, research, design, or high-stakes professional drafts.
• Route routine summaries, extraction, classification, and simple edits to Terra or Luna when quality remains acceptable.
• Treat ultra as a targeted mode for work that benefits from parallel investigation, not as a default for every prompt.

開発者が注目すべき点

Builders should watch how Programmatic Tool Calling changes agent architecture. If a model can process tool output inside a lightweight program, applications may need fewer orchestration loops and less context duplication. Teams should still keep evaluations, cost limits, permission boundaries, and human review around consequential actions.

Prompt caching also matters. Cache reads retain a large discount, while cache writes for GPT-5.6 and later models carry a premium. Products with long system prompts or repeated context should test cache behavior before estimating production costs.

検索意図の分解

People searching for GPT-5.6 release news are usually asking when the model launched, how Sol, Terra and Luna differ, whether GPT-5.6 is available in ChatGPT or Codex, and what the API costs.

People comparing GPT-5.6 with Claude or Gemini are asking a workflow question: which model completes coding, research, design, and agent tasks with the least correction and the best performance per dollar? The right answer depends on representative evaluations, not one benchmark.

Goodiebase の視点

GPT-5.6 makes model selection more explicit. Sol, Terra and Luna are not simply good, better, and best; they are routing choices for different budgets and workloads.

For Goodiebase users, the practical takeaway is to upgrade by task. Test the strongest tier on work where quality and autonomy save meaningful review time, then use lower-cost models for repeatable jobs that do not benefit from maximum reasoning.