OpenAI GPT-5.6 Pricing and Speed Overhaul and Model Selection Strategy

OpenAI has linked improvements in the execution efficiency of the GPT-5.6 family to lower API prices and faster processing. The key is not to apply the single most powerful model to every task, but to combine Sol, Terra, and Luna according to each task’s risk, complexity, latency, and verifiability to reduce the cost per result.

Pricing and performance figures are based on information announced by OpenAI on July 30, 2026. Customer benchmarks and efficiency improvement figures are measurements from OpenAI or companies cited in the announcement, so they should not be assumed to be reproducible in every environment.

July 30, 2026 API Pricing Changes

The revised API prices for GPT-5.6 Luna and Terra are as follows.

Paid subscription prices and total allocation budgets for ChatGPT and Codex are not changing as a result of this announcement. However, fewer credits will be deducted when using Terra and Luna. OpenAI stated that pricing changes through AWS would be rolled out gradually beginning later on July 30.

API Cost Calculation Example

Token costs can be calculated as follows.

Total cost = Number of input tokens ÷ 1,000,000 × input price + Number of output tokens ÷ 1,000,000 × output price

For example, if 10 million input tokens and 2 million output tokens are processed, the cost based solely on token prices is as follows.

Actual charges may be affected by service-specific conditions, including whether caching is applied, the selected processing method, and the tool-calling structure.

Roles Within the GPT-5.6 Family

GPT-5.6 is not a single model but a tiered family with different cost and performance characteristics.

GPT-5.6 Sol

Sol handles the highest level of reasoning and complex problem-solving. It is suitable for ambiguous requirements, decisions with a high cost of failure, long-term planning, and reviews of important results.

GPT-5.6 Terra

Terra aims to balance performance, cost, and response speed. It is suitable for tasks that require greater judgment than Luna but do not always require Sol, such as internal organizational Q&A, bounded agent work, and general analysis and coding.

GPT-5.6 Luna

Luna is the fastest and least expensive tier. Because it supports tool calls and multistep workflows in addition to generating simple, short text, it can serve as an execution model for repeating clearly defined tasks at scale.

Representative applications include:

• Classifying documents and customer inquiries at scale
• Extracting structured data
• Repetitive code modifications
• Writing and running tests
• Generating documents under clear rules
• Reviewing content at scale
• Background agent automation
• Repetitive research assistance

OpenAI claims that Luna delivers performance similar to that of models considered state-of-the-art a year ago at approximately 6% of the estimated cost per task and nearly 9 times the speed. Here, 6% is not a direct comparison of token prices, but a comparison of the estimated cost required to achieve the same task result.

Features of Sol Fast Mode

Fast mode for the API has been introduced for GPT-5.6 Sol. It replaces the existing Priority Processing and corresponds to Codex’s /fast feature.

The phrase “up to 2.5 times faster” does not guarantee that latency for every request will decrease by exactly the same ratio. Actual perceived speed may vary depending on prompt length, output length, service load, and the number of tool calls.

When Fast Mode Is Suitable

• Real-time services where users wait for responses
• Development environments that rapidly iterate on code changes and verification
• Tasks where Sol calls determine the latency of the entire agent
• Situations where even a delay of a few minutes matters, such as incident response or outage analysis
• Work where the cost caused by processing delays exceeds the additional API cost

Standard processing may be more economical for tasks without urgent completion deadlines, such as background batches, overnight analysis, and asynchronous processing.

Outcome-Oriented Model Selection Criteria

Model selection is not a matter of choosing the model at the top of a leaderboard. The following questions should be reviewed for each task.

Clear tasks that can be verified automatically are likely to be suitable for Luna. Terra can be considered when a consistent level of judgment is needed. Sol is suitable for resolving ambiguity, making high-risk judgments, or conducting final reviews of failure outcomes.

Planning With Sol and Executing With Luna

Different models can be assigned to different stages even within a single agent task. For example, a coding agent can be structured as follows.

1. Sol identifies ambiguities in the requirements and organizes questions.
2. Sol determines the implementation plan, scope of changes, and risk factors.
3. Luna implements the clarified changes in code.
4. Luna writes and runs tests.
5. Luna or Terra evaluates the test results and code differences.
6. Sol reviews only conclusions that are important or have a high likelihood of failure.

This structure can reduce costs compared with using Sol at every stage while concentrating strong reasoning capabilities on important judgments. However, splitting work across models requires separate designs for routing rules, error handling, log tracing, and evaluation systems.

Three Efficiency Layers Behind the Price Reductions

OpenAI explains that the cost reductions resulted not merely from pricing policy, but from technical improvements across three layers.

1. Token efficiency of the model itself
2. Hardware efficiency of the inference system
3. Efficiency of the agent harness connecting models, tools, and context

This explanation is based on technical materials released by OpenAI, and the full details of the cost structure have not been disclosed publicly. It is therefore difficult for outsiders to determine how much of the price reduction came from technical efficiency versus strategic pricing policy.

Model and Inference System Optimization

Improving Work Per Token

According to OpenAI, GPT-5.6 was trained to optimize not only task success rates but also processing efficiency. This means it was designed to reduce unnecessarily long reasoning and repetition and reach the required result with fewer tokens, increasing intelligence and work performed per token.

Load Balancing and Request Routing

The inference system distributes requests across data centers and clusters based on region, available capacity, and accelerator type. Within a cluster, it selects model instances by considering factors such as current load, input context length, cache availability, and request characteristics.

OpenAI stated that it used GPT-5.6 Sol and Codex to analyze production traffic, investigate the causes of load imbalances, test routing strategies, and adjust heuristics.

GPU Kernel Optimization

GPU kernels are the core code that executes a model’s mathematical operations on hardware. Processing costs can vary even on the same GPU depending on memory movement, synchronization, data layout, and parallelization methods.

According to OpenAI, GPT-5.6 Sol participated in writing and optimizing production kernels using Triton and Gluon within the Codex environment. The company stated that kernel improvements and related optimizations reduced the end-to-end cost of serving the model by 20%.

End-to-end cost refers not to a single operation, but to the cost of processing an actual request in its entirety, including routing, data movement, model execution, and output generation.

Verifying Kernel Accuracy

Even a fast kernel cannot be used if it produces numerically incorrect results. OpenAI explained that it invested in verification tools, including FpSan, to check the accuracy of AI-written kernels. FpSan stands for Floating-Point Sanitizer and is an open-source tool that detects errors related to floating-point operations.

This demonstrates that when AI generates production infrastructure code, numerical verification, regression testing, and failure recovery procedures are required alongside performance benchmarks.

Speculative Decoding and KV Cache

Speculative Decoding

Speculative decoding is a method in which a small draft model first proposes upcoming tokens and a larger primary model verifies multiple candidates in parallel. If the proposals are accepted, the number of expensive sequential computations performed by the primary model can be reduced.

OpenAI stated that GPT-5.6 Sol designed and ran hundreds of experiments that varied the size and architecture of the draft model and monitored training. The company explained that this improved token generation efficiency by at least 15%.

KV Cache and Workload-Specific Configurations

As a model processes input, it creates a Key-Value cache, or KV cache. The optimal configuration varies according to input and output length, batch size, cache hit rate, the number of concurrent requests, memory capacity, and model sharding method.

OpenAI explains that it used Sol and Codex to analyze actual workloads, evaluate candidate configurations, and finely tune engine configurations for each type of task. The goal is to process more requests on the same hardware.

Agent Harnesses and Context Costs

An agent calls models and tools multiple times to resolve a single user request. If a task requires 30 model calls and each call incurs 1 second of unnecessary delay, the total latency can increase by approximately 30 seconds.

OpenAI describes the Rust-based orchestration layer connecting models, tools, and user environments as an agent harness.

Deferring Exposure of Tools Until Needed

Including information about every tool, plugin, skill, and MCP integration in the prompt from the outset increases input tokens and latency. The harness uses deferred discovery, exposing relevant tool information only when needed. OpenAI stated that tool output is limited to 10,000 tokens by default unless the model requests a separate limit.

Exact Prefix Preservation and Prompt Caching

Prompt caching reduces redundant computation by reusing an identical prefix from previously processed input. To reuse the cache, the prompt prefix must match exactly.

OpenAI’s harness manages history in an append-only structure, adding new messages and tool results at the end. Tools are presented in a deterministic order, while execution settings such as approval policies are applied at runtime without changing the tool definitions themselves. This approach helps increase cache hit rates in repetitive agent loops.

How to Interpret Enterprise Case Study Figures

OpenAI’s official announcement included evaluations from Replit, Notion, Ramp, Blitzy, Cognition, Dust, and others.

• Notion stated that in its own evaluation, Terra delivered quality similar to GPT-5.5 at half the cost per task and in 60% less time.
• Blitzy explained that after adopting Luna, its prompt cache reuse rate increased from 24% to 90%, while costs were 87% lower than with its previous default model.
• Dust stated that for the same agent work, Luna was 40% faster and 40% less expensive than its previous default model.
• Ramp stated that it uses Luna as the default model for background agent automation.

These figures are case studies measured using each company’s internal work, prompts, evaluation criteria, and system architecture. Because they are not independent common benchmarks, they should not be applied directly to the work of other organizations. Before adoption, organizations need their own evaluations reflecting actual data and the cost of errors.

Pre-Adoption Validation Checklist

1. Prepare a representative sample of tasks and correct answers or evaluation criteria.
2. Compare the success rates and retry rates of Luna, Terra, and Sol under the same conditions.
3. Include not only token costs but also tool calls, review personnel, and failure recovery costs.
4. Measure 95th- or 99th-percentile latency in addition to average latency.
5. Classify stages that support automated testing or schema validation as candidates for lower-cost models.
6. Apply separate approval and logging policies to work involving personal data, security, or regulatory matters.
7. Define routing criteria for escalating low-confidence results to Terra or Sol.
8. Use actual traffic to verify whether the value of reduced latency exceeds the additional cost of Fast mode.

Significance and Limitations

This restructuring shows that the basis of AI model competition is shifting from a single top score to cost per result. By assigning Luna to large-scale repetitive work, Terra to everyday knowledge work, and Sol to ambiguous and important judgments, organizations can combine intelligence, speed, and cost according to each task.

However, publicly available information does not make it possible to separate how much of the 80% price reduction came from improvements in technical efficiency and how much came from market strategy. The economics of lower-cost models are also not determined by token prices alone. If higher error rates lead to more retries and human review, the total cost of the work may increase.

Therefore, the key metric is not the price of a single model call, but the total cost of producing one result that passes validation. Success rates by model, the number of retries, latency, and review costs must be measured together to determine whether GPT-5.6’s price reductions translate into actual business value.