GPT-6 Sol, Luna, and Claude Opus 5.5 Make Routing the New Model Moat

The most important AI releases of September 22 were not simply two new model families. OpenAI introduced GPT-6 Sol and Luna as faster, cheaper relatives of GPT-6 Astra. Anthropic launched Claude Opus 5.5 with lower token prices and fewer tokens used per task than its predecessor. A day later, Stripe and Strands offered the other half of the story: serious agent systems increasingly treat the model as one component selected by a runtime.

At our September 23 check, the Hacker News discussions had reached 1,655 points and 795 comments for Sol and Luna, and 1,672 points and 1,022 comments for Opus 5.5. That is exceptional developer attention. It does not tell buyers which model will be cheaper on their own workload.

Four releases, one architectural shift

OpenAI now presents a tiered GPT-6 family. Astra is positioned for the hardest work, Sol for demanding professional tasks at lower cost and latency, and Luna for high-volume work. Anthropic positions Opus 5.5 as a frontier model that can vary its effort, use a faster premium mode, and lower total task cost through improved token efficiency.

Strands makes provider choice a configuration option inside a reusable harness. Stripe’s Kai chooses among more than 1,000 skills and tools inside governed sessions while sharing a common execution layer. Together, these systems point toward a routing stack with three decisions: which capability to load, which model to call, and how much inference effort to spend.

The product moat therefore moves upward. If several models can complete a task, the winning system is the one that recognizes the task, supplies the right evidence, constrains the action space, and verifies the result at the lowest acceptable total cost.

Why per-token price is the wrong headline

Token prices are easy to compare because they are published. Cost per verified outcome is what matters.

A cheaper model can become expensive if it needs longer prompts, repeats tool calls, produces work that fails review, or forces a human to redo the task. A frontier model can be economical if it finishes in one pass, uses fewer tokens, and avoids an incident. Caching and compaction can change the equation again.

Anthropic explicitly argues this point. It reports that Opus 5.5 costs less per token than Opus 5 and uses fewer tokens on the same work, producing a 40% total cost reduction in its tests. OpenAI frames Sol and Luna as extending Astra-derived capabilities across different speed and budget envelopes. Both companies are selling an efficiency frontier, not a single intelligence score.

Vendor comparisons remain difficult. Benchmark harnesses, safeguards, effort levels, retries, and tool access can differ. Anthropic’s own launch table notes cases where safeguard fallbacks affected results. A percentage without the run policy is not a portable fact.

The routing policy becomes the product

A useful production router needs more than a label such as “easy” or “hard.” It should consider:

1. Consequence: Can an error be reversed, and who is affected?
2. Evidence shape: Does the task require retrieval, code execution, images, long documents, or a private data source?
3. Verification: Is there a deterministic test, a second-model check, or only human judgment?
4. Latency: Is the user waiting interactively, or can the job run asynchronously?
5. Context size: Will a smaller model lose critical information after compression?
6. Privacy and policy: Which providers, regions, retention modes, and tools are permitted?
7. Fallback cost: What happens if the first attempt fails or a safeguard redirects the request?

The same user request may move between tiers. A fast model can classify documents and draft a plan. A stronger model can handle an ambiguous exception. A deterministic validator can decide whether the artifact passes. Routing is a sequence, not a one-time model dropdown.

What Stripe and Strands add to the model debate

Stripe’s Kai shows why routing cannot be separated from authorization. A model should not load a customer-data skill merely because the employee could access it in another context. The declared task and session boundary must limit retrieval before intelligence enters the picture.

Strands shows how context management changes model economics. Offloading long tool output, caching stable prompts, and compacting before overflow can reduce tokens without changing the underlying model. But aggressive truncation can also remove the one paragraph that determines whether an answer is correct.

These are runtime choices. A benchmark that compares model APIs without the surrounding harness may miss the cost and failure modes users actually experience. Conversely, a harness benchmark can accidentally favor the prompt and tool conventions for which one model was tuned.

A practical routing blueprint

Start with three service classes rather than dozens of model names.

Routine lane: classification, extraction, formatting, low-risk summaries, and repetitive tool use. Optimize for latency and unit cost. Require schema validation and abstention.

Professional lane: research, code changes, financial models, multi-document synthesis, and longer agent loops. Optimize for cost per accepted artifact. Preserve source provenance and run domain-specific checks.

Frontier lane: ambiguous, high-impact, novel, or failure-sensitive work. Spend more inference and add an independent reviewer, deterministic gate, or human approval. Do not equate stronger reasoning with authorization to act.

Log the route, model version, effort, tool set, retrieved sources, cache behavior, retries, validator result, and human corrections. Without that ledger, a team cannot learn whether a cheaper route genuinely worked.

Limitations and unresolved questions

All four primary sources are produced by organizations promoting their own systems. OpenAI and Anthropic report internal and partner benchmarks. Stripe reports internal adoption and business correlations. Strands reports harness-level efficiency before publishing its full methods paper.

Model releases can also age faster than procurement cycles. A carefully tuned router may face new prices, context limits, safeguards, or model versions within weeks. Portability requires evaluation fixtures and versioned policies, not just a common API.

There is also a governance problem. Dynamic routing can make behavior harder to explain: two similar requests may reach different models or fallbacks. Users need to know when the system changed lanes, especially when privacy, safety, or quality guarantees differ.

Final verdict

GPT-6 Sol, Luna, and Claude Opus 5.5 intensify model competition, but they weaken model identity as the center of an AI product. The durable advantage is a control plane that selects capability, context, model, effort, and verification for each task.

The next benchmark that matters is not “which model wins?” It is “which routed system delivers the most verified work per dollar without crossing its authority boundary?”