September 25, 2026. The frontier labs spent this week fighting on price, and for once the businesses that buy AI by the token are the winners. On Monday September 22 OpenAI released GPT-6 Sol and Luna at half the per-token price of the GPT-5.6 models they replace. Hours later Anthropic released Claude Opus 5.5 with a 20% cut to input and output prices and a 60% cut to cache reads, which it says nets out to 40% less on typical workloads. The same day, VentureBeat pointed out that Sol is now half the price of Opus 5.5 on both input and output. For an agency that runs an AI receptionist, a lead qualifier or a document pipeline for clients, the model line item just fell for the second time this year, and the only lab moving the other way is the one that was supposed to be the cheap option.
Who moved, and by how much
- OpenAI, September 22. GPT-6 Sol went from $4 to $2 per million input tokens and from $20 to $10 output. GPT-6 Luna went from $0.20 to $0.10 and from $1.20 to $0.50. OpenAI's launch page labels both 50% cheaper. GPT-6 Astra, the flagship released September 3, stays at $10 and $50.
- Anthropic, September 22. Opus 5.5 is $4 input and $20 output against $5 and $25 for Opus 5. Cache reads fell from $0.50 to $0.20 per million, and since Anthropic says cache reads make up the majority of agentic and coding costs, that is where the 40% figure comes from. Fast mode is $8 and $40. Sonnet 5.5 and Haiku 5.5 follow in the coming weeks.
- DeepSeek, the other direction. The Information, via PYMNTS, reports that DeepSeek raised API prices in August by 2.3 to 4.5 times depending on the model, told investors the increase did not shrink its customer base, and has more than doubled its annualized revenue run rate to $1 billion while raising about $7.5 billion at a roughly $75 billion valuation ahead of a Shanghai listing.
- The pricing edges did not move. OpenAI still reprices any request over 272K input tokens at 2x input and 1.5x output for the whole request. Anthropic still bills its full 1M context at standard rates. Cache reads are now $0.20 per million on Sol, Luna and Opus 5.5 alike, which makes the cached prefix the cheapest token on either platform.
What a task costs now
Price a standard agent turn of 50,000 input tokens, 40,000 of them cached, and 5,000 output. GPT-6 Sol: $0.02 fresh input, $0.008 cached, $0.05 output, about $0.08. Opus 5.5: $0.04, $0.008, $0.10, about $0.15. Astra: $0.39. Luna: under a cent. A voice agent that handles 3,000 calls a month at that shape runs about $234 a month on Sol and $444 on Opus 5.5 in model fees, before telephony, and those numbers were roughly double a week ago. Our cached input pricing analysis walks through why the cache multiplier, not the headline rate, decides the bill on long-running agents.
Why DeepSeek can raise prices in a price war
The reported figures say DeepSeek's customers are mostly developers on the API, that prices stayed among the lowest for major models even after the rise, and that the run rate doubled after the increase. The lesson for buyers is uncomfortable: the cheapest model on the market is the one whose price you control least, because there is no discount left to protect. Frontier labs are cutting because they can afford to, and because every cut pulls workloads off the low-cost providers. If a client stack was built around a cheap model to save money, this week moved the break-even.
What it means for operators
Reprice every automation you sell, this week, on the new tables. A quote built on GPT-5.6 Sol or Opus 5 rates is now double the real model cost, and a client who reads the news will notice. Where a task runs on Sol, keep the request under 272K input tokens or the saving disappears; where it runs on Opus 5.5, use the one-hour cache for prefixes that repeat across a shift. Then look at the two places where a lower model price changes what you can sell rather than what you charge: always-on voice, where a per-minute model cost that has halved makes an AI calling agent for a home services company viable at the volumes a plumbing or HVAC firm actually sees, and per-lead enrichment, where a Luna call now costs a fraction of a cent. The price war is a margin gift for the businesses we serve, provided someone actually passes it through, and that repricing is a service an AI automation agency should be doing for every client this month.
Frequently Asked Questions
OpenAI's GPT-6 Sol ($4 to $2 input, $20 to $10 output per million tokens) and GPT-6 Luna ($0.20 to $0.10, $1.20 to $0.50), both released September 22, and Anthropic's Claude Opus 5.5 ($4 and $20 against $5 and $25 for Opus 5, with cache reads down from $0.50 to $0.20), released the same day.
Per token, yes: Sol is $2 input and $10 output against Opus 5.5's $4 and $20, so half the price on both sides. Cache reads are $0.20 per million on both. Sol reprices requests over 272K input tokens at double, while Opus 5.5 charges standard rates across its full 1M context.
According to The Information's reporting, DeepSeek raised API prices by 2.3 to 4.5 times in August, told investors the increase did not reduce its customer base, and its prices remained among the lowest for major models. Its annualized revenue run rate has since reached $1 billion.
Reprice first. If your workload already runs on GPT-5.6 Sol or Opus 5, the successor models cut the same task's model cost by roughly half at similar or better quality. Test the same set of real tasks on the new models and compare cost per completed task before migrating.