“Token Cost Management Takes Center Stage in the LLM Market” From OpenAI to SpaceXI, Race for “Lower-Cost, Higher-Efficiency” Models Intensifies

Three-Model Lineup Spanning Flagship Sol to Lightweight Luna
Routing, Caching and Batch Processing Built Into GPT-5.6
Intensifying Price Competition Among Providers Shifts Battleground to Deployment Capabilities

Price competition encompassing both model performance and inference costs is intensifying across the global artificial intelligence (AI) market. With Google, SpaceXAI, Meta and Anthropic promoting lightweight models alongside caching and batch-processing discounts, OpenAI has joined the fray with the launch of its cost-efficient GPT-5.6 product family. As enterprise customers scrutinize AI spending and return on investment more closely, token consumption and the actual cost per task have emerged as key criteria in model selection. The market’s competitive battleground is consequently expected to shift toward deployment capabilities encompassing data quality, workflow design and distribution networks.

Inference Costs Halved From Previous Generation

According to The Wall Street Journal (WSJ) and other foreign media outlets on the 12th, OpenAI released the GPT-5.6 product family to the public on the 9th after making it available exclusively to selected institutions for the preceding two weeks at the request of the U.S. government. GPT-5.6 comprises Sol, its highest-performing flagship model; Terra, its second-tier model; and Luna, a cost-efficient lightweight model. Users can move away from deploying the most powerful model for every request and instead assign models according to task complexity and budget. OpenAI said the new models not only achieved industry-leading performance in coding, knowledge work and scientific research, but could also complete the same tasks using fewer tokens and at a lower cost than rival flagship models, thereby improving “performance per dollar.” Tokens are the fundamental units AI models use to process information and generate responses.

OpenAI Chief Executive Officer Sam Altman said in a recent CNBC interview that “GPT-5.6 Sol improved token efficiency for agentic coding tasks by 54%.” He emphasized GPT-5.6’s superior cost efficiency relative to performance, saying, “Companies care about what value they are receiving in return for the money they spend on AI.” Just a year ago, executives confidently asserted that they could continue raising corporate subscription fees on the strength of high-performance AI models’ competitive advantages. More recently, however, they have increasingly stressed the need to secure price competitiveness. Altman had previously said at the Allen & Company Sun Valley Conference in Idaho on the 7th that “AI spending has emerged as a major concern for the first time this year.” His remarks suggest executives have begun focusing more closely on expenditure levels and return on investment than on AI adoption itself.

GPT-5.6’s performance metrics were also presented in a manner reflecting this market shift. According to AI model evaluator Artificial Analysis, GPT-5.6 Sol’s maximum-reasoning mode scored 59 on the Intelligence Index. Although one point below Anthropic’s Claude Fable 5, its cost per evaluation task was just $1.04, roughly one-third that of Fable 5. The figure illustrates how model competition is moving away from a singular focus on achieving the highest score toward an approach that measures intelligence and cost together. GPT-5.6 Sol’s advantage was even more pronounced in agentic coding. Sol scored 80 on a coding agent index combining DeepSWE, Terminal-Bench 2.1 and SWE-Atlas-QnA, the highest result among all models evaluated. Its cost per task was approximately 40% lower than Claude Fable 5 and around 10% lower than Claude Opus 4.8.

Token consumption also declined. GPT-5.6 Sol used approximately 15,000 output tokens per comprehensive intelligence evaluation task, delivering a higher score with fewer resources than GPT-5.5, which used 16,000. OpenAI also expanded the scope of cost controls by introducing more granular reasoning levels. GPT-5.6 adds a “max” reasoning tier that allocates greater computing resources to complex problems, along with an “ultra” mode in which multiple sub-agents divide the workload. By allowing users to adjust reasoning budgets according to a task’s importance, the system is designed to curb excessive token consumption at the platform level.

Actual Costs Hinge on Utilization Rates

These cost-control features incorporate optimization techniques that enterprise customers and developers have previously implemented on their own. Companies using generative AI have traditionally reduced costs by combining model routing, prompt compression, caching and batch processing. Common approaches include assigning standardized queries to lower-cost models while routing only complex analytical tasks to high-performance models, or storing recurring context in a cache to reduce input-token consumption.

In GPT-5.6, these techniques—previously dependent on developers’ operational expertise—have been embedded directly into the product family and application programming interface (API). The tiered structure of Sol, Terra and Luna simplifies model routing. Sol handles sophisticated tasks such as coding, science, cybersecurity and long-horizon analysis, while Terra is assigned to general knowledge work and Luna to workloads where repetition and throughput are paramount. Rather than invoking the same model for every request, companies can allocate computing resources according to performance requirements and expected returns.

Pricing has also been differentiated according to each model’s intended use and performance. GPT-5.6 Sol’s API pricing is set at $5 per million input tokens and $30 per million output tokens, while Terra costs $2.50 and $15, respectively, and Luna costs $1 and $6. In Artificial Analysis evaluations, Terra and Luna recorded per-task costs of $0.55 and $0.21, respectively—approximately 50% and 80% lower than Sol.

Alongside its differentiated pricing structure, OpenAI introduced explicit prompt-cache controls and reasoning-state retention features to reduce token costs incurred during API operation. Retrieving repeated inputs from the cache provides a 90% discount, while retaining previous reasoning results across multistage tasks reduces unnecessary recomputation. Tool-calling procedures can also be configured programmatically, cutting the number of exchanges between the model and external tools as well as the volume of intermediate output tokens.

A separate pricing principle applies to cache storage. GPT-5.6 is the first OpenAI model to charge 1.25 times the standard input price for cache writes. The measure reflects the underlying cost of stored tokens occupying memory resources. Indiscriminately storing even infrequently reused context can diminish the cost-saving benefits, increasing the need for companies to manage cache utilization rates precisely.

Enterprise customers are already using multi-model operations as a cost-control mechanism. Coinbase CEO Brian Armstrong said the company was reviewing an approach under which engineers would “test inexpensive Chinese AI models as their default models and route requests to the most suitable model depending on the request type.” Cloud platform Vercel likewise concluded that companies must combine models from OpenAI, Anthropic, Google and Chinese AI developers according to specific workloads to improve investment efficiency.

High-Performance AI No Exception to Price Competition as Downward Pressure on Rates Intensifies

As cost efficiency assumes greater importance in corporate model-selection criteria, price competition among AI providers has intensified further. Google unveiled Gemini 3.1 Flash-Lite in March, targeting high-volume agentic tasks, translation and data processing, and began full-scale distribution in May. API pricing is set at $0.25 per million input tokens and $1.50 per million output tokens. Batch pricing, which groups requests for asynchronous processing, falls to $0.125 and $0.75, respectively. Cached input pricing for retrieving recurring context is set at $0.025 per million tokens, easing the cost burden for companies operating large-scale services.

Elon Musk-led SpaceXAI launched Grok 4.5, a model designed for coding and agentic workloads, on the 8th, one day before GPT-5.6’s public release. Its API pricing is $2 per million input tokens and $6 per million output tokens, substantially below the $5 and $25 charged for Anthropic’s Claude Opus 4.8. Musk said Grok 4.5 delivers Opus-class performance at greater speed and with lower token costs.

Meta also entered the price competition on the 9th with the release of Muse Spark 1.1, a model targeting coding and AI agent workloads. Meta set its API pricing at $1.25 per million input tokens and $4.25 per million output tokens. Meta CEO Mark Zuckerberg emphasized in an interview with Bloomberg that there remains considerable scope to provide high-performance intelligence at a substantially lower cost. Meta Chief AI Officer Alexandr Wang likewise said the company had established a price point capable of supporting large-scale coding workloads.

Anthropic is also responding to cost-sensitive demand through differentiated product pricing and caching and batch-processing discounts. API pricing for its flagship Claude Fable 5 model is $10 per million input tokens and $50 per million output tokens, but batch processing cuts those rates in half to $5 and $25. The lightweight Claude Haiku 4.5 costs $1 and $5, respectively, while cached input is priced at $0.10 after a 90% discount.

Experts broadly agree that as price differentials narrow, AI companies’ competitiveness is increasingly likely to be determined by their deployment systems rather than model usage fees. Business performance will hinge on the ability to refine operational data and connect it to models, automatically route tasks to the appropriate model, establish governance procedures for verifying errors and secure distribution networks with direct customer access. Even when organizations use the same models, productivity disparities may widen depending on the quality of their data and workflow design.