Why OpenAI and Anthropic Slashed LLM Prices by 80%

The Price War Nobody Expected

In just four months, OpenAI and Anthropic have slashed the price of their flagship large language models by as much as 80%—a pace that has surprised even the most seasoned AI market watchers. The cuts began in late June and have accelerated through September, with both companies engaging in a tit-for-tat pricing battle that has reshaped the economics of AI.

According to data from Vercel's AI model gateway, the cost-per-token has fallen to 43.3% of its April 2026 level. Silicon Data's LLM Token Expenditure Index shows prices down 54% from their May peak, while Ramp's business spending index reports a 48% decline since February. This isn't a minor adjustment—it's a structural shift.

The Timeline of Cuts

The price war started on June 26 when OpenAI released GPT-5.6 Sol at $5/$30 per million tokens, alongside the smaller Luna model at $1/$6. Anthropic countered with Claude Sonnet 5 at $2/$10 on June 30. But the real shock came on July 30, when OpenAI cut Luna's price by a staggering 80% to $0.20/$1.20.

Anthropic responded by making Sonnet 5's temporary discount permanent on August 10, then released Claude Opus 5.5 on September 22 with pricing 40-60% lower than its predecessor. OpenAI fired back the same day with GPT-6 Sol at $2/$10, and followed up a week later with GPT-6.1 Sol, which also halved cache read costs to match Anthropic's 5% rate.

The Open-Weight Threat

The primary driver is competition from open-weight models. DeepSeek's V4 Flash, released in April at just $0.14/$0.28, went viral and forced OpenAI's hand. The model is now hosted by 29 distinct providers on OpenRouter, creating a marketplace where switching costs are nearly zero.

Moonshot AI's Kimi K3.1 Pro, announced in late October, matches Claude Opus 5.5's performance at half the price. With weights dropping November 4, dozens of providers will likely host it at even lower costs. As Max Woolf notes, "Open-weight models are catching up fast."

Cache Economics: The Hidden Lever

A critical factor is prompt caching. For agentic coding tools like Codex and Claude Code, cache reads account for 96-99% of token usage. Anthropic's decision to cut cache read costs by 75% on Fable 5.1, and the subsequent matching by OpenAI, created effective cost reductions of 70-75% for heavy users.

Anthropic offers the steepest cache discount at 90% off, while OpenAI provides 50%. These reductions matter more than headline input/output prices for real-world workloads.

IPO Timing and Strategic Bets

The cuts come as both companies prepare for public offerings—OpenAI rumored for 2027, Anthropic potentially within months. Cutting primary revenue by 80% seems counterintuitive pre-IPO, but reports suggest inference profit margins remain healthy, possibly as high as 80% per token.

Woolf notes his own behavior shifted: after GPT-6.1 Sol's effective cost dropped 75%, he cut his Codex usage quota to one-fifth, keeping spend flat. OpenAI simultaneously reduced the $200/month Codex tier's allowance by 50%, citing "increasingly efficient models."

The Infrastructure Efficiency Story

Anthropic explicitly stated that Opus 5.5 "requires less compute to serve than Opus 5," suggesting genuine infrastructure improvements. Better serving efficiency, improved quantization, and optimized inference stacks are enabling lower costs without quality loss.

This creates a virtuous cycle: cheaper models drive more usage, which drives more optimization, which drives further price cuts. As long as inference providers can profit at these levels, prices will keep falling.

What It Means for Enterprises

For businesses, this is transformative. AI coding assistants, customer service automation, and data analysis tools are now viable at scale. The cost of running an AI agent for a full day of work has dropped from hundreds of dollars to tens of dollars.

The old fear that token prices would rise once companies hooked users has proven unfounded. With 29+ providers competing on OpenRouter and open-weight models matching proprietary performance, the market has found a natural floor near marginal cost.

The Bottom Line

This price war is not a temporary promotion—it's the new normal. OpenAI and Anthropic are building moats through enterprise commitments and personal agent products like OpenAI Dots and Claude Duet, not through pricing power. The era of expensive AI tokens is over, and the era of AI as a commodity has begun.