The LLM price war has officially gone nuclear. Between July and October 2026, OpenAI and Anthropic executed a series of aggressive price cuts, reducing costs for key models by up to 80 percent. These aren't just minor adjustments; they represent a fundamental shift in the economics of AI inference. Developers relying on pricing data from before September are operating with outdated numbers, likely overpaying significantly for their API calls.

The New Pricing Landscape

OpenAI’s GPT-5.6 Luna saw an 80 percent drop overnight on July 30, falling from $1/$6 per million tokens to $0.20/$1.20. More critically, the introduction of GPT-6.1 Sol on September 29 set a new benchmark at $2/$10 per million tokens, a 60 percent reduction in input costs compared to GPT-5.6 Sol. Anthropic countered with Claude Haiku 5.5, which launched at $0.10/$0.50 per million tokens in October, representing a roughly 90 percent cost reduction for prompts under 100K tokens compared to Haiku 4.5. These moves suggest that model efficiency and infrastructure optimization are now the primary battlegrounds, not just raw capability.

Why the Sudden Drop?

While OpenAI remains tight-lipped about the specific drivers, Anthropic explicitly stated that Claude Opus 5.5 requires less compute to serve than its predecessor, a claim supported by a 40 percent estimated cost reduction for common workloads. Max Woolf, whose analysis sparked much of the current discourse, argues that the near-interchangeability of models has lowered switching costs. With open-weight models like DeepSeek Flash available across 29 providers on OpenRouter, a price floor has been established. Additionally, speculation mounts that both companies are preparing for IPOs, making current market share and usage volume more valuable than short-term margin preservation.

Hidden Costs and Cache Traps

Developers must beware of hidden levers that can inflate bills despite lower per-token rates. Cache write operations are expensive; Anthropic charges 1.25x the base input rate for 5-minute caches and 2x for 1-hour caches. OpenAI follows a similar structure, charging 1.25x for 30-minute caches and 2.5x for 12-hour caches. If you write to cache but rarely read from it, you are paying a premium for nothing. Furthermore, reasoning tokens are billed as output tokens. A model that 'thinks' extensively will drain your budget rapidly, even if the visible output is short. Long-context penalties also apply: prompts exceeding 272K tokens on GPT-6.1 Sol are billed at double the input rate, a trap for careless agent implementations.

Key Takeaways

  • Recalculate all API budgets immediately; pre-September figures are obsolete.
  • Monitor cache write-to-read ratios to avoid paying 125-250 percent premiums on unused cache.
  • Track reasoning tokens separately, as they consume output budget without generating visible value.
  • Note that GPT-5.6 Sol promotional pricing expires November 21, 2026; migrate to GPT-6.1 Sol for stable, lower rates.

The Bottom Line

This isn't just a discount; it's a race to the bottom driven by efficiency and IPO anxieties. If your team hasn't audited their token usage and cache strategies in the last 30 days, you are leaving money on the table.