Claude Opus 5.5 and GPT-6 Sol/Luna Shift AI Competition Toward Cost per Enterprise Task

Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Sol and GPT-6 Luna launches point to a more practical phase of the AI race. Frontier capability still matters, but enterprise buyers increasingly need to know a different answer: what will it cost to complete a reliable workflow at scale?

That measure is broader than a posted API rate. It includes input and output token prices, how many tokens a model needs to finish the job, the share of context that can be reused from cache, response time and—most importantly—whether the task succeeds without costly retries or human intervention.

The price cuts were paired with efficiency claims

Anthropic priced Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, 20% below Claude Opus 5. The company says Opus 5.5 should cost about 40% less than Opus 5 on typical token-billed workloads because it requires less compute to serve and uses tokens more efficiently. Anthropic also cut cache-read pricing by 60%, to $0.20 per million tokens. 21

OpenAI made a similar case with GPT-6 Sol and GPT-6 Luna. It said caching and inference improvements enabled a 50% API-price reduction from GPT-5.6 promotional pricing. GPT-6 Sol is priced at $2 input / $10 output per million tokens, while GPT-6 Luna is priced at $0.10 input / $0.50 output per million tokens. 24

The comparison is not a simple declaration that one model is cheaper than another. Models differ in output volume, capability, latency, context behavior and task reliability. But the common message is clear: vendors are selling more useful work per dollar, not merely lower per-token rates.

Why prompt caching has become a product feature

Prompt caching is especially consequential for long-running coding and agentic workflows. These systems may repeatedly send a substantial system prompt, repository context, policy document, tool definitions or conversation history on every step.

When that repeated material can be read from cache at a steep discount, the effective cost of the workflow can fall sharply. Anthropic says cache reads represent a large portion of long-running agentic-work costs, which helps explain why its Opus 5.5 cache-read cut is more significant than a conventional input-price reduction alone. 21

OpenAI likewise attributed the Sol and Luna price reductions to improvements in caching and inference. 24 For buyers, this means a pricing evaluation should test realistic prompts and multi-step tasks—not just a single short benchmark query.

The competitive pressure is real

Cheaper open-weight and Chinese models have made raw price gaps much harder for U.S. API providers to ignore. Reuters reported that DeepSeek V4-Flash was about 105 times cheaper to run than Anthropic’s Claude Fable 5 in one research firm’s benchmark testing; the reported API prices were $0.14 per million input tokens and $0.28 per million output tokens. 1

Price is not the only buying criterion. Enterprises may weigh quality, safety practices, support, data controls, integration and reliability differently by use case. Yet large gaps in operating cost create a credible incentive to route simpler, high-volume work to lower-cost alternatives.

Moonshot’s Kimi K3 illustrates the more nuanced challenge. Moonshot said its model trailed Anthropic’s and OpenAI’s leading systems overall, while outperforming some near-frontier predecessors on selected coding and agent benchmarks. 6 That combination—competitive performance on some work and a lower-cost proposition—pushes closed-model providers to show why their premium is justified on each workload.

From a single flagship to a routing ladder

The launches also reinforce a portfolio approach to enterprise AI.

• Premium frontier models remain appropriate when a task is unusually difficult, high stakes or expensive to get wrong. In these cases, stronger reasoning, coding performance or reliability can justify a higher bill.
• Capable production models such as Opus 5.5 and GPT-6 Sol are aimed at making advanced AI economical across a much wider set of recurring workflows.
• Low-cost models such as GPT-6 Luna are suited to large-volume work where the value of each individual request is modest and unit economics dominate: classification, extraction, routing, summarization and other routine automation.

The practical outcome is model routing rather than a winner-take-all choice. A company can reserve premium capacity for difficult exceptions while assigning routine work to less expensive systems. This approach can reduce overall spend without requiring every task to accept the quality trade-offs of the cheapest available model.

What enterprises should measure instead of headline token prices

A useful evaluation asks for the cost of an outcome, not the cost of a million tokens. Teams should measure:

1. Task success rate: How often does the model complete the workflow correctly?
2. Tokens per successful task: Does one model reach an acceptable answer with less generated reasoning or fewer retries?
3. Cache-hit rate: How much repeated context can be reused at discounted rates?
4. Human-review cost: Does a lower-cost model create more corrections, escalations or rework?
5. Latency and throughput: Can the system meet operational requirements at the needed volume?

That is the strategic signal from these releases. Frontier models have not become irrelevant; they remain the high-end tier. But to win broad enterprise deployment, AI providers increasingly have to pair strong capability with credible workflow economics—efficient inference, lower token use, effective caching and a clear path for customers to use the right model at the right price. 2124