A million output tokens from GPT-6 Luna now costs $0.50. That is the rate OpenAI put on its pricing page when it shipped the model on 22 September, and it undercuts Claude Haiku 4.5 by a factor of ten on both input and output. Claude Haiku 4.5 bills $1 per million input tokens and $5 per million output. Luna bills $0.10 and $0.50.
Nobody needed to go that cheap for marketing reasons. Luna is OpenAI's replacement for GPT-5.4 Nano, which the company deprecated on 1 October and will pull from the API on 1 April 2027. Nano costs $0.20/$1.25. On a workload that Compareai.today models at 50 million input tokens and 5 million output tokens a month, the forced migration off Nano drops the bill from $122.50 to $55. That is a 55% saving for doing nothing but updating a model string.
The mid-tier is where the real fight sits, because that is where most production traffic runs. Anthropic released Claude Sonnet 5.5 on 28 September at $2 per million input tokens and $10 per million output. OpenAI released GPT-6.1 Sol the next day at exactly $2 and $10. Both halve to $1/$5 through their batch APIs. Two of the biggest labs in the industry landed on the same number, one day apart, by coincidence or not.
The Tie at $2/$10 Barely Is One
List price is the part of the bill customers can see from their office. The part they pay every month is different, and the two models diverge there.
Cached input is the first split. GPT-6.1 Sol bills a cache hit at $0.10 per million tokens, which is 5% of its input rate. Sonnet 5.5 charges $0.20 for the same token, and so did GPT-6 Sol, the model GPT-6.1 Sol replaced seven days after its own launch. Compareai.today ran a workload with 80% of input served from cache and put the difference at about $24 a month: $644 on GPT-6.1 Sol against $668 on Sonnet 5.5, before cache writes.
That gap sounds small. It is. Output tokens dominate the bill on almost every real workload, and output costs the same on both.
The bigger divergence is long prompts, and it flips the comparison outright. Sonnet 5.5 bills its full 1M-token context window at standard rates. GPT-6.1 Sol carries a 1.05M-token window, but above 272K input tokens its rate rises to $4/$15 for the entire request, not just the excess. A retrieval pipeline that routinely sends 400K tokens of context gets hit with double input and 150% output on GPT-6.1 Sol while Sonnet 5.5 stays flat. For that team the cheap model is the one with the higher list price.
Neither lab is releasing a new cheap tier on the same schedule. Anthropic has not shipped a Haiku since 4.5. So Haiku's pitch now rests on output quality, not price, and Luna has taken the floor out from under it. Gemini 2.5 Flash-Lite matches Luna at $0.10 input, but Google closed Gemini 2.5 Pro, Flash and Flash-Lite to new developers on 18 September without deprecating them or changing the price. Existing users keep access. Nobody starting a project does.
Cache Reads Are the Quiet Price War
Three of September's launches moved the cached-input line and nothing else. Claude Fable 5.1 kept $10/$50 and cut cache reads to $0.25, or 0.025 times the input price. GPT-6.1 Sol kept $2/$10 and cut cache reads from $0.20 to $0.10. Claude Opus 5.5 cut its cache reads 60%, from $0.50 to $0.20, on the way to its 20% input and output reduction.
Anthropic says cache reads make up the majority of agentic and coding costs. That is a self-serving claim from a company whose Opus 5.5 cache price is double GPT-6.1 Sol's, and it is also the best explanation for why the labs keep editing that line instead of the headline number. A coding agent re-reads the same file ten times before it edits it. A chat assistant does not.
Opus 5.5 landed at $4/$20, which is exactly where GPT-5.6 Sol's promotional rate sits. Only one of those prices is permanent. Anthropic's is the list price. OpenAI says the Sol promotion runs at least through 21 November 2026, and calls that a floor rather than an end date, which means Sol can go back to $5/$30 after it. Teams on Sol's promotional rate should be testing GPT-6.1 Sol before 21 November rather than after, because GPT-6.1 Sol is already half the price and OpenAI describes it as near-Astra capability for complex work.
xAI moved in the other direction, or rather didn't move at all. Grok 4.7 replaced Grok 4.6 as the flagship on 21 September at an unchanged $2/$6, with cached input at $0.50. That cache rate is now five times GPT-6.1 Sol's. Grok 4.7 has no batch API at all, and requests that reach 200K prompt tokens are billed at the higher rate for every token in the request, same as OpenAI's long-context threshold.
What Breaks on 23 October
OpenAI shuts down the o4-mini API on 23 October. Azure retires its o4-mini deployments on 19 November. The named replacement is GPT-5.6 Terra.
This is the part of the story that should worry people, because it is where the falling prices stop helping. OpenAI names GPT-5.6 Terra as o4-mini's successor. On Compareai.today's standard workload, that move takes the bill from $550 to $1,200. GPT-6 Luna would cost $55 for the same work. A company that follows the migration guide without reading the rate card more than doubles what it pays.
The deprecations are not finished. OpenAI deprecated GPT-5.4 Nano, GPT-5.1 and GPT-5.3-Codex on 1 October. Nano's API closes on 1 April 2027 and the GPT-5 nano snapshot closes on 11 December. Anthropic retires Claude Sonnet 4.5 from its API on 30 November, and Haiku 4.5's retirement commitment runs only to "not sooner than" 15 October, so a notice could land any time with the company's usual 60 days of warning.
Perplexity changed the shape of its pricing rather than the number. It ended support for the Sonar Chat Completions API on 27 September and moved Sonar to the Agent API, where input fell from $1 to $0.25, output rose to $2.50, and web searches are billed separately at $2.50 per 1,000 calls, or $1 with Fast Search. The per-request fee is gone. Whether a given application got cheaper depends entirely on how many searches it was making before, and there is no way to know that without running both bills.
Mistral held the middle. Mistral Large 3 costs $0.50/$1.50, which is more expensive than Luna on input and three times on output, but cheaper than GPT-6.1 Sol on both. DeepSeek's deepseek-flash runs $0.30/$1.20 at peak and half that off-peak, and the company extended its off-peak window to Chinese public holidays, which means a batch job scheduled for a holiday now runs all day at the low rate.
The Spread Is Now 1 to 100
ActuIA surveyed eighteen public offers on 6 October between 00:00 and 01:00 UTC, reading the rate cards of OpenAI, Anthropic, Google, Mistral, DeepSeek and xAI plus Amazon Bedrock and Google Cloud. A million output tokens ranges from $0.50 on GPT-6 Luna to $50 on GPT-6 Astra and Claude Fable 5.1. That is a 1-to-100 spread on the same page, from the same vendors, on the same day. The survey is published as a full comparison table with euro conversions at the ECB reference rate.
Two of GPT-6's three tiers launched at half the price, or less, of the GPT-5.6 models sitting beside them. The obvious reading is that the labs are buying volume by cutting cost. The other reading is that the cheap tier is a funnel into the expensive one, and that Luna's $0.50 output rate exists to get a startup's traffic pattern onto OpenAI's infrastructure before it needs something that costs $50.
It's not clear either reading holds for long. Prices this low get used for things nobody budgeted for, and the counter-pressure shows up in the fine print: rate limits, data-use terms, context thresholds, promotions with floors instead of end dates. The rate card is the marketing page now. The bill is somewhere else.
Watch two dates. GPT-5.6 Sol's promotional floor passes on 21 November, and whether OpenAI extends it, ends it, or lets it lapse quietly will say more about this price war than any launch. Google's Gemini 3.8 Flash introductory pricing doubles to $1.50/$7.50 on 1 January 2027.
More AI pricing and model coverage is in our AI desk archive.