September 23, 2026: Competition between leading artificial-intelligence companies is increasingly moving beyond raw model capability and into another battleground: how much intelligence customers can get for every dollar they spend.
OpenAI and Anthropic released new models within hours of each other on September 22, highlighting the growing importance of inference costs as developers and businesses seek to control the expense of deploying AI at scale.
OpenAI introduced GPT-6 Sol and GPT-6 Luna, adding lower-priced options alongside its premium GPT-6 Astra model.
Anthropic, meanwhile, launched Claude Opus 5.5, which it says delivers performance comparable to its higher-end Claude Fable 5.1 on most work while costing significantly less to operate than the previous Opus generation.
The near-simultaneous releases underline how cost efficiency has become a major competitive factor in the AI market.
OpenAI Launches GPT-6 Sol and GPT-6 Luna
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, according to the company's API changelog.
Both are reasoning models capable of accepting text and image inputs and generating text through OpenAI's Responses and Chat Completions APIs.
The models occupy different positions within OpenAI's expanding model portfolio.
GPT-6 Sol provides a more capable option for demanding workloads while being priced considerably below premium GPT-6 Astra. GPT-6 Luna targets applications where developers prioritise low costs and high-volume usage.
For standard API processing with prompts of up to 272,000 input tokens, OpenAI lists GPT-6 Sol at:
$2 per million input tokens
$0.20 per million cached input tokens
$10 per million output tokens
GPT-6 Luna is substantially cheaper:
$0.10 per million input tokens
$0.01 per million cached input tokens
$0.50 per million output tokens
Those prices contrast sharply with OpenAI's premium GPT-6 Astra, which carries standard short-context rates of $10 per million input tokens and $50 per million output tokens.
The pricing structure allows OpenAI to compete across different levels of the market rather than relying on one flagship model for every workload.
Anthropic Unveils Claude Opus 5.5
Anthropic responded with its own efficiency-focused release on the same day.
The company introduced Claude Opus 5.5, the first model in its new Claude 5.5 family.
Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work, while costing approximately 40% less to run than Claude Opus 5.
Claude Opus 5.5 is priced at:
$4 per million input tokens
$20 per million output tokens
Anthropic has positioned the model around a combination of high-end capability and lower operating costs rather than simply pursuing the highest possible benchmark performance regardless of expense.
The company says early testers reported significant performance improvements on complex tasks, including software engineering.
Anthropic Says Efficiency Savings Are Being Passed to Customers
The lower pricing is not being presented merely as a promotional discount.
Anthropic says improvements in model efficiency have allowed it to reduce the underlying cost of operating Claude.
The company said it is passing those efficiency gains to customers through lower prices and higher rate limits.
That distinction matters for developers.
If model providers can achieve comparable or better performance using less computation, they may be able to reduce inference prices without accepting the same pressure on margins that would come from straightforward discounting.
Claude Opus 5.5 Starts a New Model Family
Opus 5.5 is only the beginning of Anthropic's latest model generation.
The company says Claude Sonnet 5.5 and Claude Haiku 5.5 are expected to arrive over the coming weeks.
That could extend the pricing competition beyond premium models into mid-range and lower-cost workloads.
Anthropic's model strategy increasingly gives customers multiple performance and pricing tiers, similar to the approach taken by OpenAI.
The result is a market in which customers can choose models based not only on benchmark performance but also on latency, workload complexity and total inference cost.
The AI Competition Is Shifting Toward Performance Per Dollar
For the first several years of the generative-AI boom, competition between leading labs largely centred on which company could build the most capable frontier model.
That competition has not disappeared.
But enterprises deploying AI at scale face a different economic problem.
A small difference in token pricing can translate into substantial expenditure when a company processes billions of tokens across customer-service systems, coding agents, document analysis, research tools or automated workflows.
That makes performance per dollar increasingly important.
A model does not necessarily need to be the most capable available to be economically attractive. For many workloads, businesses may prefer a model that completes a task reliably at a fraction of the cost.
OpenAI Has Already Been Cutting Prices
The latest GPT-6 releases follow earlier price reductions across OpenAI's model portfolio.
On July 30, 2026, OpenAI reduced the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%.
Then, on August 21, OpenAI cut API and credit pricing for GPT-5.6 Sol by more than 20% for a three-month period.
Those reductions, followed by the arrival of GPT-6 Luna at extremely low token prices, point to a broader effort to make advanced AI economical for high-volume applications.
Rather than forcing every customer toward the company's most expensive model, OpenAI is building a wider range of options spanning inexpensive workloads through premium frontier intelligence.
Anthropic Faces the Same Enterprise Cost Pressure
Cost efficiency is particularly important for Anthropic because businesses form a significant part of its customer base.
Financial Times reporting described the latest launch as part of Anthropic's effort to compete with OpenAI and lower-cost alternatives as increasingly price-sensitive customers evaluate AI spending.
The challenge extends beyond the two US companies.
Open-weight AI models, including increasingly competitive alternatives developed outside the largest American AI labs, give businesses additional options and put pressure on proprietary model providers to justify their pricing.
That means OpenAI and Anthropic are competing not only against each other but also against an expanding ecosystem of models that can sometimes be deployed at considerably lower cost.
Price Alone Does Not Determine the Cheapest AI Model
Token prices provide an easy comparison, but they do not reveal the complete cost of running an AI application.
A model that charges less per token can still become more expensive if it requires significantly more tokens, retries or additional processing to complete the same task.
Conversely, a more expensive model can sometimes deliver a lower total task cost if it reaches the desired result faster and with fewer attempts.
Businesses therefore increasingly need to evaluate models using measures such as:
cost per successfully completed task, token efficiency, latency, reliability and required human intervention.
That is a more meaningful comparison than simply asking which company has the lowest headline API price.
Safety Remains Part of Anthropic's Opus 5.5 Pitch
Anthropic's latest release is also notable because it follows CEO Dario Amodei's recent call for slowing the development of increasingly powerful frontier AI systems.
Anthropic says Opus 5.5 underwent external evaluation before release, including testing by Frontier Design and METR.
The company also says the model incorporates safeguards developed for its most capable systems and performed strongly in its automated behavioural audits.
These are Anthropic's reported evaluation results and should not be interpreted as independent proof that one model is universally safer than competing systems.
The company is attempting to demonstrate that improvements in efficiency do not necessarily require reducing investment in safety evaluation.
Why Cheaper Models Matter for Businesses
Lower inference prices could have significant consequences for AI adoption.
A company experimenting with an AI assistant may generate relatively modest usage. But once AI agents begin performing continuous tasks across thousands or millions of customers, token consumption can increase dramatically.
Cheaper models can make previously uneconomical applications viable.
Customer support, document processing, software development, research, translation, data extraction and agentic workflows are all areas where relatively small improvements in cost per task can become important at scale.
The emergence of models such as GPT-6 Luna also gives developers another option: use inexpensive models for routine tasks and reserve premium models for problems requiring significantly greater reasoning capability.
AI's Next Battle May Be About Economics as Much as Intelligence
OpenAI and Anthropic are still racing to build increasingly capable AI systems.
But the releases of GPT-6 Sol, GPT-6 Luna and Claude Opus 5.5 demonstrate that capability alone is no longer the only competitive metric.
For developers and businesses, the question is increasingly becoming: How much useful work can a model perform for a given amount of money?
OpenAI is addressing that question by widening its model portfolio from the inexpensive GPT-6 Luna through Sol to premium Astra.
Anthropic is pursuing a similar strategy by improving the efficiency of its Claude family, beginning with Opus 5.5 and planning additional Sonnet and Haiku releases.
As models become more capable and businesses move from experimentation to large-scale deployment, the competition between AI labs is increasingly becoming a race not simply for the smartest model — but for the most economically useful intelligence.
