AI's Price War: OpenAI and Anthropic Slash Token Costs by Half

The price of frontier intelligence just fell off a cliff. Over a five-day stretch in late September, three of the biggest AI labs cut their API prices by 40 to 50 percent within hours of each other. xAI's Grok 4.7 opened on September 21, then Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Sol and Luna landed on September 22, ninety minutes apart. The most consequential move came from OpenAI: for the first time, a closed frontier lab priced a model below an open-weight rival on both input and output tokens.

That is a genuine inflection point. For most of the past year, the comfortable assumption across the industry was that open-weight models would always win on cost while closed labs competed on raw capability. This week broke that assumption, and it says more about where AI economics are heading than any single benchmark release.

The numbers: three labs, one week

OpenAI's GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, exactly half the promotional rate on its GPT-5.6 predecessor. GPT-6 Luna, the high-volume workhorse, dropped to $0.10 input and $0.50 output per million tokens. Anthropic's Claude Opus 5.5 landed at $4 and $20, a 20 percent list cut but roughly 40 percent cheaper to run by the company's own measure, with cache reads slashed to $0.20 per million tokens. xAI's Grok 4.7 opened at $2 and $6.

According to Fortune, the two labs are "deep into an all-out price war," and OpenAI did not hide its target. As TechCrunch reported, the company repeatedly claims Sol and Luna handle tasks substantially better than Anthropic's top models, while attributing the price cuts to improvements in caching and inference. OpenAI says GPT-6 Sol makes about half as many mistakes as its predecessor, "reaching Astra-level reliability at much lower cost."

Why now: open weights became the price floor

The pressure did not appear overnight. DeepSeek's V4.1 Flash, the current open-weight reference point, charges $0.15 input and $0.60 output per million tokens off-peak. GPT-6 Luna undercuts it on every axis, and because Luna carries no peak-time surcharge, the gap widens when DeepSeek doubles its rates during high-demand windows. The "cheapest frontier callable" label just moved from open weights to a closed provider. It is the logical next step in a trend we flagged earlier this week, when open-weight models from DeepSeek and Qwen began winning developers on price and performance.

The enterprise pushback behind this is real. According to CNBC, startups like Lindy have already moved 100 percent of their traffic off Anthropic's Claude models to DeepSeek, watching their cost curve "crash to the ground." CFOs burned by sticker shock on early AI pilots are now asking how to optimize spend for 2027. Fortune quotes Ramp's lead economist, Ara Kharazian, on the blunt mechanics: the price war is "driving down the price of AI and therefore driving down their ability to profit from it."

Cost per task is the new unit

The deeper story is not the sticker price but the switch in how AI work gets measured. OpenAI's launch does not lead with per-token rates; it leads with cost per completed task. Its own table shows GPT-6 Sol at extra-high reasoning effort finishing AutomationBench tasks at about $0.27 each, against Claude Opus 5 at more than 11 times that cost per task. Per-token pricing hides the variables that actually drive a bill: reasoning tokens, tool calls, retries, and how many turns a model needs before it stops.

Caching is where the war is really being fought. Both labs cut cache-read pricing aggressively because agentic and coding workloads spend most of their tokens re-reading the same system prompts and tool definitions on every turn. A coding-agent loop that caches over 90 percent of its input can cost a fraction of what the headline rate suggests. For builders, the unit to instrument is no longer "price per million tokens" but "what does one of my tasks cost end to end" on my own traffic.

What it means for builders

The collapse in token costs is good news with a catch. Agentic, multi-step, high-context work that was too expensive a year ago is now affordable, and gross margins on AI-native products improved again, for the fourth time this year. The catch: if your moat was "we make inference cheaper," the labs have commoditized it out from under you. The differentiation now lives in your data, your workflow design, and how you route tasks across a two-tier structure, cheap high-volume models for extraction and summarization, flagship reasoning models for complex coding.

There is a second tension worth watching. The same week these cuts landed, the cost of building the data centers that serve these models kept climbing. The labs are slashing prices to defend volume against open-weight rivals even as their infrastructure bills balloon, and none of them has disclosed gross margin at the new price points. Whether this is healthy competition or a subsidized race compressing everyone's unit economics at once is unanswered. For now, the developers building on top are the clear winners, as long as they measure the right thing.