This week (22–23 Sept 2026), the AI race isn't just about which model is strongest, but about cost and the ability to work as an 'agent'. Anthropic released Claude Opus 5.5, OpenAI moved GPT-6 Sol and Luna to cheaper tiers, and MCP got its first official certification.
Claude Opus 5.5: stronger, but aimed at cheaper agents
Anthropic announced Claude Opus 5.5 on Sept 22, positioning it as the strongest Opus model for coding, AI agents and knowledge work. The highlight isn't only benchmarks: the company says typical workloads cost about 40% less than Opus 5 (it finishes the same work with fewer tokens) and generates output more than 30% faster.
| Claude Opus 5.5 | Price / 1M tokens |
|---|---|
| Input | $4 |
| Output | $20 |
| Cache read | $0.20 (~60% cheaper than Opus 5) |
| Batch (input/output) | $2 / $10 |
GPT-6 Sol & Luna: bringing Astra down to two cheaper tiers
Almost simultaneously, OpenAI expanded GPT-6 with two models: Sol (balancing capability and cost) and Luna (prioritizing speed, high volume and low cost). Both inherit from GPT-6 Astra with improved caching/inference; prices are about 50% below the corresponding GPT-5.6 generation — and are permanent, not promotional.
| Model | Input / 1M | Output / 1M |
|---|---|---|
| GPT-6 Sol | $2 | $10 |
| GPT-6 Luna | $0.10 | $0.50 |
The gap is huge: for the same 1M input tokens, Luna costs about 1/20 of Sol. That enables tiered AI deployment instead of using the strongest model for every request — e.g. classification/extraction/routing on Luna, research/coding/reasoning on Sol, and especially hard frontier tasks on Astra.
The shift: from 'strongest model' to 'intelligence per dollar'
Two near-simultaneous launches show the next AI generation competing not just on benchmarks, but on how much money, how many tokens and how much time a task takes. This matters more as AI moves from one-shot chatbots to agents running dozens or hundreds of steps — a model that costs more per token but takes fewer steps may not be more expensive over the whole task.
MCPA: MCP gets its first official certification
The Agentic AI Foundation (under the Linux Foundation) introduced the Model Context Protocol Associate (MCPA) — the first official certification for MCP. It covers five domains: MCP fundamentals, architecture & components, interactions & execution, security & governance, and use cases & ecosystem; aligned with the MCP specification dated 2026-07-28. It signals MCP moving from a pioneer-team protocol to a standardizable skill — and since agents read data and call tools, MCPA emphasizes security and governance too.
Frequently asked questions
- How is Opus 5.5 cheaper than Opus 5? — Similar per-token Opus pricing, but ~40% lower typical-workload cost via fewer tokens/steps and ~60% cheaper cache reads.
- Sol or Luna? — Luna for simple, high-volume work (classification, extraction); Sol for reasoning/coding; only the hardest tasks need Astra.
- What is MCPA? — The first MCP certification (Agentic AI Foundation/Linux Foundation), testing both architecture and security & governance.
- The big trend? — Competition shifts from 'strongest model' to 'intelligence per dollar' + agents + tool connectivity.
About this article: a sourced summary by DETOON. Prices and specs are as announced by each vendor at launch and may change.
