Opus 5.5 and GPT-6 Sol/Luna: tiered AI is cutting software costs

On September 22, 2026, Anthropic launched Claude Opus 5.5 and, hours later, OpenAI shipped GPT-6 Sol and GPT-6 Luna: two unrelated announcements, the same day, with the same underlying message — per-token prices dropping 20% to 60% (Anthropic). It lands barely ten days after executives from Anthropic, OpenAI and other labs publicly called for slowing down the pace of AI development, something we covered here. At Evicron, an AI and custom software studio based in Barcelona, what matters about this double launch isn’t who “wins” — it’s a structural shift: the major providers no longer compete with a single flagship model, they compete with three pricing tiers built to fit a real project’s budget.

What each company actually shipped

Claude Opus 5.5 matches Fable 5.1’s performance on most tasks, runs 30% faster, and costs 40% less on typical workloads: $4 per million input tokens and $20 for output (down from $5 and $25 on Opus 5), with cached reads dropping 60%, from $0.50 to $0.20 per million tokens. It’s already available on AWS, Google Cloud and Azure, and Anthropic says Sonnet 5.5 and Haiku 5.5 are coming within weeks, rounding out its own three-tier lineup.

GPT-6 Sol and GPT-6 Luna sit below flagship GPT-6 Astra, with API prices cut roughly in half versus the previous generation: Sol is aimed at development and demanding knowledge work, Luna at high-volume routine tasks like data extraction or document summarization, with up to 90% off cached input tokens. Several independent outlets, including VentureBeat and TechCrunch, have confirmed these figures.

This week’s irony

Ten days before this double launch, the same executives now competing on the lowest price had signed a joint call to slow the AI race down. The contradiction is real, but for a business running or building software with AI baked in, what matters isn’t the industry’s consistency — it’s that the cost of building with these models keeps trending down, month after month, whether or not the sector agrees to slow anything.

What “tiered AI” means for your custom software

Until recently, picking an AI provider mostly meant picking one model — the most capable one your budget allowed, used for nearly everything. With Anthropic offering Opus, Sonnet and Haiku, and OpenAI offering Astra, Sol and Luna, the decision that matters is no longer “which model” but which tier for each task inside the same product:

• Top tier (Opus 5.5, GPT-6 Astra) for tasks that genuinely need complex reasoning: reviewing long contracts, decisions with real financial stakes, generating code in critical parts of the system.
• Mid tier (Sonnet, Sol) for the bulk of everyday development and knowledge work: routine features, code review, drafting answers from internal documentation.
• Budget tier (Haiku, Luna) for high, predictable volume: triaging support tickets, extracting fields from a form, summarizing a short document.

Mixing tiers inside the same application, instead of forcing everything through the most expensive model, is what turns a bloated AI bill into a reasonable one without touching what the user actually experiences.

How to apply this to a real project

1. Map your use cases by complexity, not by screen or page. A single feature usually mixes tasks that need genuine reasoning with others that are purely mechanical; not all of them need the same tier.
2. Design the architecture so switching models doesn’t mean rewriting business logic. If your integration talks to the provider through a middle layer, moving a task from top tier to mid tier — or switching provider entirely — is a config change, not a project.
3. Re-review your tier split every quarter. What needs the most expensive model today may be handled by the mid tier within two or three generations, a pattern we’ve seen repeat with every model cycle.
4. Don’t confuse price tier with risk tier. A cheap model processing sensitive data still needs the same security and compliance guarantees as an expensive one; the savings should come from matching the task, not relaxing the controls.

What doesn’t change

As we noted when Anthropic cut its cache pricing, token price is rarely the biggest line item in a custom software project with AI built in: solution design, integrating with existing systems, and pre-launch testing usually outweigh the monthly API bill, as we break down in our guide to custom software development costs. A price cut, or a new tier of models, improves the business case for a well-designed project — it doesn’t fix a badly designed one.

How we handle this at Evicron

In our applied AI projects, we design the architecture so each task uses the tier it actually needs, mixing Claude, GPT and Gemini based on the use case rather than whichever model made headlines this week. When the question is whether your current software is overpaying by routing everything through a single model, our AI consulting for businesses audits how tasks are actually distributed and tells you, with numbers, how much you can save without touching the user experience.

Bottom line

Opus 5.5 and GPT-6 Sol/Luna confirm that major providers now compete with three-tier lineups rather than a single flagship, and the real savings come from matching each task in your software to the right tier — not from chasing this week’s release. Ten days after calling for AI to slow down, the market is moving faster than ever, and the advantage goes to whoever designs their architecture to capture it.

Want to know if your custom software is using the right AI tier for each task? Get in touch: the first consultation is free and we reply within 24 hours.