Opus 5.5 and GPT-6 Luna slash prices instead of breaking new ground

The race for the smartest model has collided with economic reality. Companies began experimenting with routing queries to smaller, cheaper open-weight models because deploying frontier models for everyday tasks proved too expensive. Two key announcements this week show how the major players are trying to retain enterprise customers.

Cutting costs instead of dazzling benchmarks

Anthropic introduced Opus 5.5, the latest version of its workhorse for complex knowledge tasks and coding. OpenAI counters with GPT-6 Sol and Luna, more efficient models focused on speed. The common denominator isn't a massive leap in capabilities, but a drop in price.

Steep discounts rewrite the math for long-running agents

While benchmarks show Opus 5.5 modestly beating the recently released GPT-6 Astra, the real draw is the pricing sheet. Anthropic cut the price of input and output tokens to $4 and $20 per million, respectively, representing a 20% drop from Opus 5.

The crucial news is the 60% reduction in cache read costs to $0.20 per million tokens. Because these operations constitute the bulk of costs for long-running agentic tasks, Opus 5.5 now allows more headroom for iterative problem-solving that would previously have devoured budgets.

A cheaper frontier model won't solve vendor lock-in

The catch is that cheaper APIs from top providers may only temporarily mask the advantages of owning infrastructure. Customers continue to build on closed interfaces and lack the flexibility to shift workloads to another model without rewriting prompts.

Although Opus 5.5 promises 30% faster generation, real-world enterprise operations will reveal how well this number holds up during peak loads and complex queries.

Adoption outside the vendor's narrative will dictate the shift

We will see the true impact once companies react to these discounts. If customers start abandoning complex systems with model routers and return to blindly using a single vendor due to lower costs, then Anthropic and OpenAI's gamble paid off. Otherwise, we are just watching a delaying tactic before inevitable commoditization.

Lilith's verdict

Instead of the search for AGI, we got a price war in the lower-middle class. The winner is whoever first convinces the CFO that cheaper versions no longer require a dedicated prompt optimization team.