Claude Opus 5.5 and GPT-6 Sol and Luna Drop Simultaneously — The Frontier Model Price War Is Here

In the same 24-hour window, Anthropic shipped Claude Opus 5.5 at 20% lower cost than its predecessor and OpenAI cut GPT-6 Sol and Luna prices in half. The timing is not coincidental. Frontier model pricing has entered a compression phase, and developers running production agentic systems should run their token math today.

What Happened

Claude Opus 5.5 arrives at $4/M input and $20/M output — down from $5/M and $25/M. Cached input drops to $0.20/M, a 60% reduction from the prior rate. Anthropic’s positioning is direct: “the intelligence of Fable 5.1 at 40% lower cost.” The model also specifically addresses “Claudish” writing patterns — the overly cautious, stilted outputs that users identified as a consistent complaint with the Opus 5.x lineage.

GPT-6 Sol lands at $2/M input and $10/M output — half the prior price. GPT-6 Luna goes further, at $0.10/M input and $0.50/M output, placing it in OpenAI’s cheapest tier alongside GPT-4.1 Nano and GPT-5 Nano. Both cuts target the inference-at-scale use case where token volume makes pricing the dominant variable.

One documented issue with Opus 5.5: the “max” thinking setting hit a critical failure mode in early testing — the model exhausted its 128,000-token output limit while reasoning through a simple SVG prompt, requiring 20 minutes and costing $2.56 per failed attempt. Maximum reasoning mode is not yet reliable for production workloads.

Why It Matters

This is the first wave of coordinated frontier model price compression. The mechanism is predictable: as training costs stabilize and inference hardware becomes more efficient, the major labs shift from capability competition to volume competition. OpenAI cutting prices in half while Anthropic simultaneously cuts 20% signals that neither lab expects to maintain current price floors.

For developers, the practical implication is that the optimal model selection from three months ago may not be the optimal selection today. GPT-6 Luna at $0.10/M input is now price-competitive with dedicated fine-tuned models for many classification and extraction tasks where ceiling performance doesn’t matter. Opus 5.5 at $4/M input makes the Fable 5.1 performance tier accessible to teams that previously routed those tasks to cheaper models.

What This Means for Builders

Two decisions to revisit now. First, any agentic pipeline routing high-volume tasks to sub-frontier models on cost grounds should benchmark against GPT-6 Luna — $0.10/M input is a different regime than anything available 60 days ago. Second, teams that avoided Opus-class models because of the “Claudish” output pattern have a reason to re-evaluate: the writing quality fix is explicit and targeted.

The Opus 5.5 max-thinking failure mode is the one concrete constraint to carry forward. If your pipeline uses extended reasoning, test the specific prompts that trigger it before production deployment — the 128K output limit under max thinking is a hard failure, not a quality degradation.

Source: Simon Willison's Weblog

Content created with AI assistance and reviewed for accuracy.