Claude Opus 5.5, GPT-6 Sol and Luna are here: the new metric is cost per task

Anthropic and OpenAI announced new models within hours of each other on September 22, 2026. Anthropic released Claude Opus 5.5; OpenAI released GPT-6 Sol and GPT-6 Luna. Their shared message matters more than another benchmark record: bringing near-frontier work to a lower cost per completed task.

Claude’s X announcement frames Opus 5.5 as Anthropic’s first model since calling for pacing the frontier and emphasizes pre-release testing by external evaluators. ChatGPT’s announcement says Sol and Luna are rolling out in ChatGPT Work and Codex for paid plans. This article is a first reading of official announcements and technical documentation; I have not run the models on the same independent task suite.

Three models, three jobs

Claude Opus 5.5: aims to deliver high quality in long coding, enterprise knowledge work and hours-long agent sessions with fewer tokens and lower total cost.

GPT-6 Sol: the balanced GPT-6 option for complex coding and agentic work without Astra pricing.

GPT-6 Luna: targets very low unit cost for high-volume work such as classification, data processing and focused tool tasks.

This split makes a routed model architecture more sensible than assigning the strongest model to everything.

Claude Opus 5.5: a claim of Fable-level work at lower cost

Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40% less than Opus 5 on typical workloads. Standard API pricing is $4 per million input tokens and $20 per million output tokens. Cache reads cost $0.20 and a five-minute cache write costs $5. Fast mode promises up to 2.5 times the speed at $8 input and $40 output.

The company also reports output more than 30% faster than Opus 5. These figures come from Anthropic’s tests and typical-workload assumptions. The useful metric for a team is not token price but total cost per completed task: steps, retries, human intervention and whether the acceptance tests pass.

Price is only part of the release. Anthropic says Opus 5.5 improves on long, sprawling migrations, audits, computer use and knowledge work while communicating more clearly and concisely. Its benchmark table leads on several coding and professional-work tests, but Anthropic also says small benchmark margins can overstate the real-world gap.

Why safety sits at the center of the launch

Claude’s X post highlights alignment rather than raw performance. Anthropic says Opus 5.5 earns its strongest automated behavioral-audit result so far, is less likely to take hard-to-reverse actions or exceed assigned boundaries, and is more resistant than Opus 5 to prompt injection.

Those are vendor evaluations, not permission for unlimited autonomy. The model carries safeguards similar to Fable 5.1 for biology and cybersecurity, with verification programs for some sensitive work. Human approval and an audit trail remain necessary for money, deletion, production changes, personal data and broad system access. The Claude Fable 5.1 review covers the previous frontier tier and its data boundaries.

GPT-6 Sol: a new balanced-default candidate

OpenAI positions GPT-6 Sol for complex coding and agentic workflows. API pricing is $2 per million input tokens and $10 per million output tokens; cache reads cost $0.20 and cache writes $2.50. Sol provides a 1.05-million-token context window, up to 128,000 output tokens and a broad tool set including web and file search, code execution, hosted shell, computer use, MCP and patch application.

Sol’s practical advantage is bringing a similar tool surface below GPT-6 Astra’s $10/$50 pricing. Astra remains the top tier for the hardest science, cybersecurity and end-to-end professional work, but Sol may become the more practical default for daily coding agents and complex workflows. The GPT-6 Astra review covers that model’s technical and safety profile.

GPT-6 Luna lowers the floor for high-volume work

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. Cache reads cost $0.01 and cache writes $0.125. It still supports the same 1.05-million-token context window, 128,000-token output limit and core tool surface as Sol.

That makes Luna interesting for support classification, document tagging, structured extraction, batch QA and clearly specified subtasks. A cheap context window does not make it sensible to send a million tokens with every request. Above 272,000 input tokens, OpenAI doubles input and cache rates and applies a 1.5-times output rate to the whole request. Retrieval and selection still matter.

Where to try each model

• Opus 5.5: large code migrations, long research, audits and professional deliverables where accuracy matters.
• GPT-6 Sol: daily coding agents, multi-tool workflows and complex tasks that need a strong quality-cost balance.
• GPT-6 Luna: high-volume, well-specified, easy-to-verify tasks with limited failure impact.
• Astra or Fable 5.1: only when the task demonstrably exceeds the verified limits of cheaper models.

A pilot should measure first-pass success, elapsed time, API cost, human intervention, rework and wrong tool calls together. A cheap token is not cheap if the task must be repeated three times; a frontier model is not a good default if it overthinks simple work.

Note

The common theme of the September 22 launches is not simply higher peak intelligence but better price-performance. The decision should come from completion rate, verification burden and total cost on the same real task—not a vendor leaderboard.