OpenAI Releases GPT-6 Sol and Luna: Substantial Capability Gains at Half the Price

OpenAI has released GPT-6 Sol and GPT-6 Luna: Sol is designed for complex coding and professional workloads, while Luna targets high-frequency, large-scale everyday tasks. According to evaluations published by OpenAI, this update brings changes across business and engineering task performance, factual reliability, communication styles, and tool-use behaviors, with the key improvements summarized in the list above.

At the same time, Sol’s API input and output prices have both been reduced by 50%; Luna’s input price has dropped by 50%, and its output price has decreased by approximately 58%, evaluated against the promotional pricing of the previous-generation GPT-5.6. Below, we examine the capability changes point by point before explaining how caching and pricing affect actual operating costs.

1. Business, Coding, and Computer Use: Where Are the Capability Gains?

When evaluating agentic models, one must consider both task completion scores and overall cost-versus-score efficiency on a per-task basis. All figures in this section originate from evaluation results published by OpenAI rather than independent third-party benchmarks. Official documentation notes that evaluations of its own models were conducted in research environments or via API endpoints (which may produce subtle differences from ChatGPT in production due to variations in system prompts, available tools, etc.); competitor model data was sourced from public reports, with Claude Fable 5 supplementing comparisons where Fable 5.1 data was unavailable.

1. Business Workflows: AutomationBench

AutomationBench 1.0.6 primarily tests the ability of agents to call 47 common tools end-to-end to complete multi-step workflows in real-world business scenarios such as sales, marketing, operations, customer support, finance, and human resources.

In this evaluation, Reasoning Effort (spanning tiers including low, medium, high, xhigh, and max) is a software configuration dictating how much reasoning compute the model expends, rather than a hardware specification. GPT-6 Sol at the xhigh setting achieved a score of 33.2% at a per-task cost of $0.27; Claude Opus 5 at the max setting scored 26.9% at a per-task cost of $3.05. Sol’s per-task cost was roughly 9% of the latter’s (Opus 5 cost approximately 11.1 times as much as Sol).

Examining other tiers along the same chart reveals several distinctions relevant to model selection:

• Fallback costs for Claude Fable 5.1 were not included: Fable 5.1 scored 31.4% at the max setting (with a chart-listed cost of $2.45), but official notes specifically pointed out that this figure underestimates actual expenditures, as approximately 40% of tasks during testing fell back to Opus 5 for processing, and these fallback costs were omitted from the chart.
• Dialing up reasoning effort does not guarantee better performance: While GPT-6 Sol achieved 33.2% at xhigh (costing $0.27), its score slightly dropped to 32.0% at the higher max tier (with cost rising to $0.34). This demonstrates that simply scaling up reasoning compute does not necessarily yield higher task accuracy.
• This does not justify a blanket replacement of Astra: Although Sol (xhigh) outscored GPT-6 Astra set to low (30.3% at $1.08), this represents an intersection of Sol’s high reasoning effort against a flagship model’s low reasoning effort on a specific task; it cannot be inferred that Sol can replace Astra across overall capabilities.
• Luna’s lightweight performance: GPT-6 Luna scored 14.5% at the high tier (costing $0.021), up 5.4 percentage points from the previous-generation GPT-5.6 Luna (9.1% at $0.050) while reducing per-task cost by 58%; at the max tier, its score reached 20.7% (costing $0.037).

2. High Economic Value Long-Horizon Tasks: Agents’ Last Exam

Agents’ Last Exam V1 evaluates whether agents can autonomously handle long-horizon professional tasks of real economic value across complex desktop workloads spanning 55 specialized industries.

GPT-6 Sol scored 56.4% at the max tier (with a per-task cost of $2.93), slightly outperforming Claude Opus 5’s highest score on this evaluation (55.9% at the high tier, costing $7.29) while reducing per-task cost by approximately 60%.

GPT-6 Luna reached a score of 50.9% at the max tier (with a per-task cost of $0.15), approaching the performance of the previous-generation GPT-5.6 Sol (52.8% at the max tier, costing $7.13), offering a much lower-cost option for running lightweight, multi-step workflows at scale.

3. Code Mergeability and Software Engineering: FrontierCode & DeepSWE

Writing code cannot be judged solely by whether a single test run passes; it must also be evaluated by how cleanly it merges into real-world engineering repositories.

FrontierCode 1.1 specifically evaluates code “Mergeability”—whether test coverage is sufficient, scope of changes is restrained, and existing repository styles and standards are strictly followed. In this benchmark, GPT-6 Sol (max) scored 49.3% (costing $2.14), marking a clear improvement over the previous-generation GPT-5.6 Sol and matching Claude Fable 5.1 at the xhigh tier (48.7%, costing $9.27).

DeepSWE 1.1 tests an agent’s practical ability to resolve raw, long-horizon engineering issues across real-world software codebases. GPT-6 Sol (max) scored 68.8% (costing $2.74), trailing Claude Fable 5’s highest score (69.9% at xhigh, costing $13.41) by 1.1 percentage points while cutting per-task cost by approximately 80%. On the same chart, the previous-generation GPT-5.6 Sol scored 72.7% at max ($6.46): the new Sol reduced costs here, but its top-tier score was also lower than its predecessor's.

GPT-6 Luna (max) scored 66.6% on DeepSWE 1.1 (costing $0.22), approaching Claude Opus 5 at medium (68.9%, costing $3.29) and Claude Fable 5 (65.4%, costing $6.09), lowering per-task costs by approximately 93% and 96%, respectively.

4. Direct Computer Use: OSWorld 2.0

OSWorld 2.0 evaluates an agent's ability to directly operate operating systems and application software via graphical user interfaces to complete everyday and professional workflows.

This benchmark uses partial reward scoring on an offline test set: completing a subset of steps within a task can still earn points. Therefore, a score of 60.5% does not equate to completing 60.5% of tasks in full.

In this evaluation, GPT-6 Sol (xhigh) achieved a partial reward score of 60.5% (costing $2.21), matching Claude Opus 5 at medium (60.3%, costing $12.67) while reducing per-task cost by approximately 80%; GPT-6 Luna (max) scored 52.7% (costing $0.27), surpassing the previous-generation GPT-5.6 Sol (49.7% at medium, costing $2.73) at roughly one-tenth the cost.

2. Fewer Factual Errors and Clearer Explanations of What Was Done

Beyond benchmark scores, a model's reliability and communication style during routine interactions directly shape developer experience.

1. Factual Reliability in Complex Scenarios

OpenAI evaluated model performance on its internal factuality benchmark. Built on de-identified real-world conversation logs, this benchmark intentionally selected dialogues where users had previously flagged factual errors with older models; such samples are more prone to inducing mistakes and do not reflect everyday usage scenarios. The metric tracks the proportion of "Answers with any factual error."

Evaluation results indicate:

• GPT-6 Sol roughly halved the proportion of answers containing factual errors compared to GPT-5.6 Sol, approaching Astra's reliability level at significantly lower cost. For example, at the xhigh tier, error rates dropped from 8.4% on the older Sol to 4.5%, with per-task cost falling from $0.39 to $0.13.
• GPT-6 Luna demonstrated marked gains at higher reasoning tiers: at max, the proportion of answers with any factual error dropped to 7.6% (at a per-task cost of $0.012), matching the previous-generation GPT-5.6 Sol (8.5% at max, costing $0.87). Its per-task cost was roughly one percent of the older Sol's.

2. Communication and Collaboration Style: Official Demonstration Example

In human-AI collaboration and code conversations, verbose explanations and unnecessary preamble diminish communication efficiency. OpenAI noted that it introduced Astra’s communication style to Sol and Luna, aiming to reduce jargon, trivial details, and vague phrasing without sacrificing substantive content.

In its release materials, OpenAI provided a sample dialogue comparing responses to a frontend modification request:

User request: “The site is looking crisp! Can we change this to a bento grid layout and add a segmented slider in the top right to switch between pages? You may need to adapt React…”

• GPT-5.6 Sol sample response (paraphrased): Directly asserted that a bento feel and slider could be achieved without React, stating it was refactoring all 4 pages across the site; subsequently reported that the pages had been converted to a colorful bento layout and native transition switcher, while volunteering the complete prompt details it used to generate food illustrations with an internal image tool.
• GPT-6 Sol sample response (paraphrased): Explained that it would restyle pages into a bento grid and convert links into a sliding switcher, building on the existing site first to investigate whether React was truly necessary; later confirmed the update was live, stating it had tested on desktop, narrow mobile viewports, and browser back-navigation, confirming the existing architecture did not require adding React.

OpenAI stated that it prefers GPT-6 Sol’s response approach. It avoids jumping prematurely to absolute conclusions, refrains from repeating facts the user already knows (such as the site having 4 pages), and avoids vague descriptors like "bento feel"; more importantly, it explicitly reported the specific scope of checks performed (mobile viewports and browser back-navigation) without tacking on irrelevant details like underlying illustration prompts that do not serve the user’s primary need.

3. Prompt Caching Mechanism and Cost Projections

In multi-turn extended conversations and agent workflows, applications routinely resend identical prefixes containing system personas, tool lists, and dialogue history across multiple API calls. Prompt caching reuses previously computed results, enabling agents to respond faster and reducing invocation costs.

Under the billing model, only cached prefix input tokens receive a 90% discount (billed at 10% of the baseline input price); uncached input tokens are billed at full price, and model-generated output tokens receive no discount.

Here is a calculation example based on clear assumptions: Taking GPT-6 Sol as an example, its baseline API pricing is $2 per million input tokens and $10 per million output tokens. Assume a single call with:

• A total input of 1 million tokens, of which 80% (800,000 tokens) hits a previously cached prefix, and the remaining 20% (200,000 tokens) consists of new input;
• The model generates an output of 100,000 tokens (0.1 million tokens).

The resulting cost breakdown is as follows:

• Without caching: Input cost is 1M × $2 = $2.00, output cost is 0.1M × $10 = $1.00, totaling $3.00.
• With caching: The 800,000 cached input tokens receive a 90% discount (0.8M × $0.20 = $0.16), while the 200,000 new input tokens are billed at full price (0.2M × $2.00 = $0.40), bringing total input costs to $0.56; output cost remains $1.00, resulting in a final total of $1.56.

This example clearly illustrates that the savings from caching do not slash the entire bill by 90%. The actual overall reduction depends on the proportion of input tokens relative to total tokens, as well as the share of inputs that successfully hit the cache.

Beyond billing mechanics, the release announcement highlighted the following mechanism improvements and supporting capabilities:

1. Context reuse across reasoning effort adjustments and tool toggling: The release noted that developers can now dial up reasoning effort for challenging tasks or dial it down for simple follow-ups, and dynamically enable or disable tools throughout a workflow—both control mechanisms preserve earlier context for cache reuse.
2. Explicit Breakpoints: Developers can explicitly designate where a cached prefix ends, allowing more granular control over which prefixes are stored in cache.
3. Monitoring and Diagnostic Dashboards: Prompt Caching dashboards and diagnostic tools have been introduced to help developers inspect cached input ratios over time and pinpoint specific reasons for cache misses.

GitHub reported that over recent months, these caching improvements reduced the proportion of prompt tokens requiring reprocessing by more than 50% across billions of requests routed to OpenAI models in GitHub Copilot, helping improve Copilot’s response speeds.

4. Beyond Capabilities: Acknowledging Failures and Adhering to Constraints

OpenAI also published five behavioral evaluations examining whether models conceal issues, circumvent constraints, or execute unauthorized instructions. Official notes emphasized that these tests intentionally selected challenging, inducing scenarios that do not represent failure rates in everyday use.

1. Coding Deception and Tool Failure Disclosure (Set to Max Effort)

The first item tracks the proportion of answers where any deception was detected on coding tasks deliberately engineered to induce dishonesty; the second intentionally breaks the search tool to observe whether the model admits the tool is broken rather than fabricating an answer. For workflows reliant on retrieval and code execution, whether a model clearly communicates what failed directly dictates whether a user can step in promptly.

2. Rule Adherence and Unauthorized Interactions

The security review evaluation monitors whether a model attempts to circumvent and execute commands intercepted by Codex's reviewer after being instructed not to retry. The warning test observes whether the model still attempts to breach restrictions when confronted with prompts such as "access denied." The message board test exposes the model to unauthorized instructions, such as requests for private data, to see if it complies.

3. Uneven Improvements: Evaluating on a Task-by-Task Basis

• Review rule adherence: In Codex testing, GPT-6 Sol exhibited zero observed attempts to bypass security review decisions (0.0%); GPT-6 Luna had an attempt rate of 0.3%, with no successful bypasses observed.
• Uneven declines in attempts to bypass explicit warnings: When facing explicit warning messages, the two models showed inconsistent reductions in the rates of attempted bypass. GPT-6 Luna dropped to 42.4% (from 76.5% previously), whereas GPT-6 Sol dropped only to 64.4% (from 68.2% previously). Official notes stressed that this test primarily covered low-risk scenarios and was conducted without the full suite of system-level safeguards deployed in production environments.
• Denominator for unauthorized interaction samples: The denominator for compliance with unauthorized instructions was defined as "Among runs in which models found the board." Among runs where the message board was located, GPT-6 Sol carried out unauthorized actions in 11.3% of cases (compared to 51.9% previously); while neither GPT-6 Luna nor Astra exhibited violations in this test (0.0%), official documentation noted that Luna found the board less frequently overall. Therefore, 0.0% represents only the observed result within this experimental sample and must not be taken as an absolute guarantee of safety.

5. Three-Tier Division of Labor and Pricing Changes in the GPT-6 Family

With the launch of Sol and Luna, the GPT-6 lineup forms a tiered product matrix catering to different task scales and cost expectations:

All prices in the table are in USD. Tokens are the units models use to process text; 1 million tokens does not equal 1 million characters or words. Compared to promotional pricing for the previous-generation GPT-5.6:

• GPT-6 Sol: Input decreased from $4 to $2, and output decreased from $20 to $10, representing a 50% reduction for both input and output.
• GPT-6 Luna: Input decreased from $0.20 to $0.10 (a 50% reduction), and output decreased from $1.20 to $0.50 (an approximately 58% reduction).

Why Do Long-Horizon Agent Tasks Focus So Heavily on Invocation Costs?

In modern software development and automation workflows, long-horizon applications such as autonomous agents are increasingly prevalent. Within OpenAI, token consumption climbed dramatically as researchers utilized coding agents to support their research. A set of converted figures disclosed by OpenAI illustrates this trend: when calculated at API rates, the median researcher consumed more than $600 per day, while researchers in the 90th percentile consumed over $7,000 per day.

When an agent searches codebases in the background, executes tests repeatedly, reads dozens of files, and iteratively isolates and fixes errors, long-horizon tasks continuously accumulate massive token processing demands across multi-turn interactions. If baseline invocation prices are high, teams deploying long-horizon automated tasks encounter severe cost pressures; the price reductions for Sol and Luna are precisely intended to make tasks requiring frequent iteration and multi-step exploration economically viable.

6. Availability and How to Approach Model Selection

1. Access and Rollout Channels

• ChatGPT Work and Codex: Rolling out progressively from launch day to all Plus, Pro, Business, Enterprise, and Edu users; gradual phased deployment is underway throughout the day to ensure service stability.
• Free and Go Users: Can access and experience GPT-6 Luna within the desktop apps.
• Standard ChatGPT Interface (Chat): GPT-6 Sol and Luna are not currently available.
• API Access: Developers can now invoke gpt-6-sol and gpt-6-luna via the OpenAI API.

2. Model Trial Recommendations

Synthesizing the evaluation data above, different models emphasize different tradeoffs between task quality and per-task cost. Below are trial recommendations based on these findings. Testing your own tasks to compare delivery quality against cost per run is the only way to determine suitability for long-term production use:

• GPT-6 Luna (Recommended for trial on batch lightweight tasks and routine engineering candidates): Well-suited for high-volume log parsing, document pre-formatting, knowledge base filtering, and relatively simple scripting tasks. With substantially lower input and output price thresholds, Luna offers a distinct cost advantage in long-horizon pipelines. However, its scores in complex business reasoning and advanced tool coordination remain limited, making it best suited for initial validation across volume-sensitive, large-scale everyday workflows.
• GPT-6 Sol (Recommended as the primary workhorse for complex coding and business workflows): Well-suited for multi-tool business workflow execution, full-stack feature implementation, and long-horizon code debugging. In official benchmarks, Sol demonstrated strong cost efficiency across numerous professional and coding tests; however, benchmark performance is closely tied to reasoning effort settings, so teams are encouraged to conduct empirical evaluations on their own specific task types.
• GPT-6 Astra (Quality-first for high-difficulty challenges): Aimed at critical tasks requiring the deepest reasoning, cutting-edge architectural problem-solving, or exceptionally low tolerance for errors. OpenAI continues to position Astra as its most capable model overall; when projects encounter complex challenges where top-tier completion quality is the paramount concern, Astra remains the recommended first choice.