OpenAI has launched GPT-6 Sol and GPT-6 Luna with a pricing change that matters particularly for search agents: the company has cut standard API token prices by 50% compared with the promotional pricing of their GPT-5.6 predecessors. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens under OpenAI’s standard short-context pricing, while GPT-6 Luna costs $0.10 and $0.50 respectively.
The new models were announced in OpenAI’s official September 22 launch. OpenAI says Sol and Luna inherit advances from GPT-6 Astra while targeting different points on the cost-capability curve: Sol is positioned for demanding professional, coding, computer-use and agentic work, while Luna provides a much cheaper option for high-volume workloads where speed and efficiency matter more than maximum model depth.
For web research and retrieval systems, the headline token reduction is only part of the economics. OpenAI has also improved prompt caching for GPT-6, a change specifically designed to help agents and long conversations reuse more context instead of paying to process the same information repeatedly. Together, lower token prices and more effective caching can materially change how aggressively developers can scale persistent research workflows.
GPT-6 Sol costs $2 in and $10 out per million tokens
OpenAI’s API changelog confirms that GPT-6 Sol and Luna were released through the Responses and Chat Completions APIs on September 22. For prompts up to 272,000 input tokens under Standard processing, GPT-6 Sol is priced at $2 per million input tokens, $0.20 per million cached input tokens and $10 per million output tokens.
GPT-6 Luna is dramatically cheaper: $0.10 per million input tokens, $0.01 per million cached input tokens and $0.50 per million output tokens under the same standard short-context conditions.
OpenAI’s launch table compares those figures directly with GPT-5.6 promotional pricing. Sol falls from $4 to $2 for input and from $20 to $10 for output. Luna falls from $0.20 to $0.10 for input, while output moves from $1.20 to $0.50. OpenAI summarizes both transitions as a 50% price reduction, reflecting the overall pricing comparison it is making across the models.
The exact bill still depends on processing tier and context length. OpenAI’s current pricing documentation lists different rates for long-context requests and alternative processing modes, so the $2/$10 and $0.10/$0.50 figures should be understood as Standard short-context rates rather than universal prices for every API configuration.
Search agents consume tokens differently from ordinary chatbots
A 50% model-price reduction can be especially meaningful for agents because an agent rarely makes one model call and stops. A research system may interpret the user’s request, decide what to search, inspect results, issue follow-up searches, compare sources, call tools, maintain state and finally synthesize an answer.
Each stage can consume input and output tokens. Long-running agents also repeatedly carry instructions, tool definitions, accumulated context and intermediate results through the workflow.
That multiplication effect means the cost of the underlying model becomes more visible as the agent becomes more capable. A single research answer may contain several hidden reasoning and retrieval cycles even when the user sees only one final response.
NetContentSEO previously examined this dynamic in a dataset where 189 user prompts generated 1,797 hidden search queries. That study did not measure GPT-6 costs, but it illustrates why agentic retrieval economics cannot be estimated from the visible user prompt alone.
Lower model prices make deeper retrieval loops easier to justify
When model inference is expensive, developers have an incentive to keep research loops shallow. The agent may run fewer searches, inspect fewer sources or delegate fewer verification tasks because every additional reasoning step increases the bill.
Cheaper Sol and Luna do not make web research free, but they reduce the model component of that marginal cost. A developer can spend more tokens deciding which sources deserve attention, comparing conflicting evidence or performing a second verification pass without paying GPT-5.6-level model rates.
This can matter more than simply reducing the price of the final answer. Research quality often depends on the work performed before synthesis: query reformulation, source selection, extraction, contradiction checks and deciding whether enough evidence has been collected.
A cheaper model can therefore be used either to lower the cost of an existing workflow or to increase the amount of work the agent performs for the same model budget.
Luna creates a new low-cost layer for high-volume retrieval tasks
GPT-6 Luna’s standard $0.10 input and $0.50 output pricing creates a particularly inexpensive option for tasks that do not require the strongest available reasoning on every step.
A sophisticated research system does not necessarily need to use the same model throughout the entire workflow. Low-cost models can classify search results, extract structured facts, filter duplicates, identify candidate documents or perform straightforward transformations before a stronger model handles difficult synthesis.
OpenAI itself positions the GPT-6 family as a cost-intelligence curve rather than a single model replacing every other tier. GPT-6 Astra remains the company’s highest-capability option, while Sol and Luna trade some of that depth for lower cost and broader scalability.
For agent architects, that encourages model routing. Luna can handle routine retrieval work, Sol can take more demanding analysis, and Astra can remain reserved for tasks where the additional capability justifies its higher price.
Sol is positioned for the harder parts of agentic work
OpenAI emphasizes GPT-6 Sol’s performance on professional workflows, coding and computer use. Those capabilities matter to research agents because modern retrieval increasingly extends beyond issuing a search query.
An agent may need to operate software, manipulate files, write code to analyze data, call business tools or navigate multi-stage workflows. OpenAI reports substantial improvements for Sol over GPT-5.6 Sol on its selected evaluations, while also claiming better cost efficiency against several competitor models.
Those benchmark results are OpenAI’s reported evaluations and should not be treated as independent proof that Sol will outperform alternatives in every production workload. The more durable point for developers is that OpenAI is positioning Sol as a capable agent model at a substantially lower token price than its previous Sol tier.
That combination is relevant when the research workflow includes both retrieval and action. The same agent that searches the web may also need to write an analysis script, inspect a spreadsheet or operate a browser-based system before producing the final report.
Prompt caching can compound the token-price reduction
OpenAI’s GPT-6 launch includes a second economic change aimed specifically at agents and long conversations: improved prompt caching. The company says GPT-6 delivers higher cache-hit rates by default and provides a 90% discount on cached input-token reads.
For GPT-6 Sol under Standard short-context pricing, cached input costs $0.20 per million tokens instead of $2 for fresh input. For Luna, cached input is $0.01 instead of $0.10.
This matters because agents repeatedly reuse large prompt prefixes. System instructions, tool schemas, project context and earlier conversation state can appear again and again across a long-running workflow.
If more of that repeated context remains cacheable, the effective cost of persistence can fall by more than the headline 50% reduction in fresh token prices. The exact savings depend on the application’s cache-hit rate and prompt structure, so developers should measure rather than assume the maximum discount.
GPT-6 lets agents change effort and tools without throwing away as much cache
OpenAI says developers can now adjust reasoning effort and enable or disable tools while preserving earlier context for cache reuse. This is a subtle but important improvement for agent architectures.
A research workflow does not need the same reasoning intensity at every stage. A simple follow-up might run at lower effort, while a difficult source conflict may justify higher effort. Likewise, an agent may need web search during retrieval but not during final formatting.
If changing those controls invalidates a large cached prefix, adaptive workflows can lose much of their cost advantage. OpenAI says GPT-6 is designed to preserve more of that earlier context when these settings change.
The company has also introduced explicit cache breakpoints and tools for monitoring caching performance, making cache optimization more visible as an engineering discipline rather than a hidden implementation detail.
The Agents API gives the cheaper models a persistent research harness
The timing of GPT-6 Sol and Luna is particularly relevant because OpenAI released its public-beta Agents API less than two weeks earlier. NetContentSEO covered that launch as OpenAI productizing the research harness for persistent search agents.
The Agents API packages durable sessions, automatic context compaction, sandboxed execution, MCP, native web search, dynamic tool discovery and parallel subagents into a managed runtime. Those features make it easier to build long-running research systems, but they also create exactly the kind of workloads where repeated model calls and accumulated context can become expensive.
GPT-6 Sol and Luna attack that problem from the model side. The harness can persist and orchestrate the work, while lower inference prices and improved caching reduce the cost of keeping the agent active across many steps.
The combination is more strategically significant than either release in isolation. OpenAI is simultaneously making complex agents easier to build and making the model layer cheaper to run.
Web-search charges are not the same thing as model-token charges
The headline “half the model price” needs an important qualification. A web research agent can incur costs beyond input and output tokens.
Search tools, external APIs, MCP services, sandbox compute, storage and application infrastructure can all contribute to the total cost of a research workflow. Cutting the model’s token price in half does not automatically cut the complete agent bill in half.
The savings will be largest in applications where model inference represents a substantial share of total cost. In workflows dominated by expensive third-party data or heavy compute, the percentage reduction in total spend will be smaller.
Developers should therefore measure cost per completed research task rather than extrapolate directly from token pricing. The economically relevant metric is what the entire workflow costs to produce an acceptable result.
Cheaper inference can change how AI visibility monitoring is built
The economics are also relevant to search and GEO platforms that query AI systems repeatedly to measure brand visibility, citations or recommendations. These products often need large prompt sets, repeated runs and retrieval-enabled responses to capture probabilistic variation.
A lower-cost model tier can make higher sampling frequencies more practical. Instead of testing one prompt once, a monitoring platform can afford more repetitions, more markets or more variations for the same inference budget.
That does not solve a separate methodological problem: API outputs may not perfectly reproduce consumer-facing ChatGPT behavior. NetContentSEO has reported research showing that chatbot interfaces and APIs can retrieve substantially different source sets. Cheaper API sampling therefore improves scale, not necessarily representativeness.
Cost efficiency should enable better experimental design rather than encourage teams to treat a larger volume of imperfect measurements as ground truth.
Lower prices make verification loops cheaper too
One of the most useful ways to spend the savings is not on producing more first drafts but on checking them. A research agent can allocate additional calls to source verification, numerical consistency, contradiction detection or independent synthesis.
Parallel subagents can investigate separate evidence branches and compare conclusions. A low-cost model can perform routine citation checks before a stronger model reviews the final result.
These patterns increase token consumption, which is why they are often reduced when teams optimize narrowly for inference cost. Lower prices give developers more room to treat verification as a normal stage of the workflow rather than an expensive luxury.
There is no guarantee that more model calls automatically produce greater accuracy. Verification only helps when the workflow is designed to seek independent evidence rather than have several agents repeat the same unsupported assumption.
GPT-6 Sol and Luna are rolling out across Work, Codex and the API
OpenAI says GPT-6 Sol and Luna became available in ChatGPT Work and Codex on September 22 for Plus, Pro, Business, Enterprise and Edu users, with a gradual rollout during the day. Free and Go users can access GPT-6 Luna through the desktop app.
Both models are also available through the OpenAI API as gpt-6-sol and gpt-6-luna. OpenAI says they are not yet available in the standard Chat experience, making Work, Codex and API integrations the initial deployment surfaces.
The positioning reinforces the models’ role in active workflows rather than only conversational Q&A. Work is designed for substantial multi-step assignments, Codex for software-engineering tasks and the API for developers building their own systems.
Those are precisely the environments where inference efficiency compounds over repeated tool calls and long-running tasks.
The price cut changes the architecture question
When frontier-quality inference is expensive, teams often ask how little model work they can get away with. GPT-6 Sol and especially Luna make a different question more practical: how much additional retrieval, checking and iteration can fit inside the same budget?
OpenAI has not made search agents free, and the strongest GPT-6 model remains Astra at a much higher price. But Sol at $2/$10 and Luna at $0.10/$0.50 materially widen the range of agentic workloads that can be run continuously or at high volume.
Improved caching amplifies that shift for persistent systems, where repeated context can otherwise dominate input costs. Combined with the new Agents API, OpenAI now has both a managed architecture for long-running research agents and lower-priced models designed to operate inside it.
For developers building web research and retrieval products, GPT-6’s most important feature may therefore be economic rather than spectacular: more searches, more verification and longer-running agents can fit inside the same model budget. At scale, that can change not only what an agent costs, but how ambitious its research process is allowed to become.