The AI Price War Just Got a Scoreboard: Luna Down 80%, Opus 5 at Half Price, and Token Spend Is Moving

Two American labs cut prices in the same week. One Chinese research firm is now tracking what the market does about it. The result is the clearest picture yet of where AI model pricing is headed.

On July 30, OpenAI dropped the price of GPT-5.6 Luna by 80%, three weeks after launch. Anthropic answered by shipping Opus 5 at roughly half the cost of its own premium tier. Behind both moves sits the same pressure: open-weight models from Chinese vendors that match US flagships on many benchmarks while charging far less per token. source

What OpenAI Did to Luna

OpenAI priced Luna at $1 per million input tokens and $6 per million output tokens when GPT-5.6 launched. The July 30 cut takes it to $0.20 input and $1.20 output, a combined $7 down to $1.40 per million tokens. The same announcement trimmed Terra by 20%, from $2.50/$15 to $2/$12, while leaving the flagship Sol untouched at $5/$30. source source

Model	Old input/$1M	Old output/$1M	New input/$1M	New output/$1M	Change
GPT-5.6 Luna	$1.00	$6.00	$0.20	$1.20	-80%
GPT-5.6 Terra	$2.50	$15.00	$2.00	$12.00	-20%
GPT-5.6 Sol	$5.00	$30.00	unchanged	unchanged	0%

OpenAI's marketing leans hard on the per-task math. It claims Luna delivers performance comparable to frontier-class models from a year ago at roughly 6 cents on the dollar per task, and nearly nine times the speed. On Agents' Last Exam, the company says Luna beats Claude Fable 5 at an estimated cost per task close to 99% lower. source

The structural move matters more than the headline discount. At $1.40 combined per million tokens, Luna now undercuts Google's Gemini 3.5 Flash-Lite at $2.80 and sits far below Gemini 3.6 Flash at $9. It is no longer an OpenAI mid-tier model priced against OpenAI's own lineup. It is competing in the market's low-cost inference tier, against Xiaomi's MiMo, DeepSeek's flash models, and the rest of the budget crowd. source

OpenAI also added a premium Fast mode for Sol, priced at a combined $70 per million tokens. That is the mirror image of the Luna cut: charge more for latency-sensitive frontier work, charge far less for high-volume utility work. source

Anthropic's Half-Price Answer

Anthropic responded a few days later with Claude Opus 5, priced at $5 per million input tokens and $25 per million output tokens. The company markets it as "frontier intelligence" at half the price of its top-tier Mythos model. Opus 5 lands at the same sticker price as Opus 4.8, but delivers substantially more capability, which is another way of cutting the price per unit of work while leaving the list price alone. source source

Model	Input/$1M	Output/$1M	Combined	Positioning
Anthropic Mythos	$10.00	$50.00	$60.00	Top tier
Claude Opus 5	$5.00	$25.00	$30.00	Half price, near-top intelligence
GPT-5.6 Sol	$5.00	$30.00	$35.00	OpenAI flagship
GPT-5.6 Luna	$0.20	$1.20	$1.40	High-volume utility

Anthropic also added an adjustable effort setting that lets developers trade reasoning depth for speed and fewer tokens. source The subtext is the same as OpenAI's split: protect the premium tier, make the everyday tier cheap, and let customers tune the middle.

The Pattern: Flagship Stays Priced, Utility Gets Cheap

Neither lab discounted its flagship. OpenAI left Sol at $5/$30. Anthropic kept Mythos at $10/$50 and let Opus 5 undercut Fable 5 by half. source

Two readings compete in the market commentary. The bullish read: top-end intelligence is not a commodity yet, premium margins survive through enterprise agreements, and the discounts are confined to the workload layer where customers actually shop on cost. The bearish read: the repeatable-workload tier is the layer that scales fastest, and when both labs compete there, revenue per task keeps falling even as usage climbs. source

OpenAI was explicit about the mechanism that makes cheaper prices survivable. Its engineers rewrote production GPU kernels, cutting serving cost by about 20%. Speculative decoding was redesigned, improving token generation efficiency by more than 15%. Agent runtime changes reuse prompt caching across multi-step workflows instead of repeating computation. The claim is that every point of serving efficiency becomes a reason to lower API prices, not just an engineering milestone. source source

The Index That Caught the Shift

SiliconData, a US research firm, publishes a daily LLM Token Expenditure Index that blends what enterprises actually pay across frontier APIs, open-weight platforms, and dedicated instances. The index is average willingness to pay, not a sticker price. source

Enterprise inference costs hit their lowest level of the year in the first week of August, averaging $1.16 to $1.18 per million tokens across providers. That is down from $2.04 on May 31 and $1.45 in late July, according to Jefferies citing SiliconData. source

The sharper signal is the direction of spend. Per the Financial Times, SiliconData data shows inference token spend pulling away from closed models since mid-July. Spending still leans heavily toward proprietary models, but the gap has narrowed over the past month. source

There is a counter-intuitive wrinkle inside the same dataset. Effective prices paid for open models actually rose in that window, because more powerful near-frontier Chinese releases like GLM 5.2 and Moonshot's Kimi K3 are served at higher prices than the bargain DeepSeek tier they displaced. source The two families are converging on each other from opposite directions: closed models got cheaper, open models got better and slightly more expensive.

Who Actually Moved Workloads

The trend is not abstract. DoorDash, Airbnb, and Coinbase are among the large US tech companies that confirmed using Chinese AI models, often shifting production or low-intensity requests to cheaper alternatives while keeping US models for development. Model-routing services make that split trivial: send the hard prompt to a frontier model, route the easy traffic to a budget one. source

The friction is data governance. Security and compliance teams remain concerned about sending potentially sensitive data outside the country, which is one reason closed flagship spend still dominates in absolute terms. source

Token Price Is Not Task Price

The most useful lens for anyone buying inference: comparison by headline token rate overstates the savings of an open model and understates the value of an efficient closed one. Agents make dozens or hundreds of model calls to finish one task, so serving cost and prompt-caching behavior matter more than a few cents of difference in list price. source

SiliconData says as much when it stresses that token price is not task price. source A cheaper token that forces a rewrite of your routing layer is not cheaper. A pricier token that finishes in one tenth the calls often is.

What Buyers Should Do Now

The market is repricing every week, and the direction of travel is one-way. Enterprise inference cost per million tokens has roughly halved since the end of May, and the shift of spend away from closed models is being measured, not just asserted. The window where "list price times volume" ruled procurement is closing.

Buy decisions should weight the real unit of value: completed work per dollar, not tokens per dollar. Benchmark your open-weight candidate and the discounted closed model on the same agentic workload, including what caching and tool-calling actually cost. Re-check quarterly, because the index moves monthly and the gap keeps narrowing. And before sending sensitive data to a budget provider, decide whether the per-token savings survive your compliance review, because that is exactly the trade-off the biggest spenders are now making.