OpenAI prices GPT-6 Sol and Luna for an agentic split

OpenAI says inference and caching improvements made the lower prices possible. A company representative confirmed to VentureBeat that the Sol and Luna prices are permanent, rather than promotional launch rates.

Three models, three cost profiles

The intended division is straightforward:

Compared with GPT-5.6, Sol is 50% cheaper for both input and output tokens. Luna is 50% cheaper for input tokens and 58.3% cheaper for output tokens.

For companies building AI agents, this split matters more than a simple model leaderboard. Agent economics depend on how often a model is called, how much context is sent again, and whether every step really needs the strongest available system.

The pricing also places Sol directly beside Claude Sonnet 5, which costs $2 per 1 million input tokens and $10 per 1 million output tokens. Anthropic made that price permanent in August.

Anthropic then released Claude Opus 5.5 on Tuesday morning at $4/$20. That makes Sol 50% cheaper on both uncached input and output tokens. Anthropic says Opus 5.5 is 20% cheaper per token than Opus 5 and about 40% cheaper in typical workflows because it needs fewer tokens to complete the same work.

At the other end, Anthropic’s Claude Fable 5.1 costs $10/$50, five times Sol’s price in both directions.

Google’s Gemini 3.8 Flash remains cheaper in the performance-focused tier. Its launch price is $0.75/$3.75 through December 31, rising to $1.50/$7.50 on January 1, 2027. Google positions it for long-running software development, autonomous AI agents and complex enterprise workflows. Even after the increase, Flash will remain cheaper than Sol—but Google has put an end date on the current price, while OpenAI says Sol’s $2/$10 rate has no expiration date.

Luna sits in a different market altogether. Its $0.10/$0.50 pricing is 95% below Sonnet 5 on both input and output tokens, and about 86.7% below the current launch pricing for Gemini 3.8 Flash. Against SpaceXAI’s Grok 4.7, released just one day earlier, Luna is 95% cheaper on input and about 91.7% cheaper on output for requests under 200,000 tokens.

Those are price comparisons, not capability comparisons. They show where Luna fits: moving routine agent work away from expensive reasoning models.

Open weights make Sol’s price harder to defend

Xiaomi’s MiMo-V2.6-Pro adds pressure from a different direction. Artificial Analysis called it the highest-performing open-weights model at launch. Through Xiaomi’s API, it costs $0.435 per 1 million uncached input tokens and $0.87 per 1 million output tokens—about 78% and 91% below Sol’s prices, respectively.

The smaller MiMo-V2.6-Flash costs $0.14/$0.28:

Both Xiaomi models use the MIT license. A company can download and run them without paying Xiaomi for every token, shifting the economics from API bills to infrastructure and operations.

That is the more important comparison for Sol. It does not need to be the cheapest model on the market. It needs to be reliable enough on coding and agent tasks, and easy enough to operate, to justify its premium over increasingly capable open-weights models.

My read is that OpenAI is selling Sol as an operating choice rather than a bargain. Its price matches Sonnet 5, but the real pitch is that a managed API with dependable task completion can cost less than a cheaper model that needs more retries, supervision or infrastructure. The announcement is quiet about the exact operational burden behind those alternatives.

OpenAI’s benchmark is the completed task

OpenAI frames the value of Sol around the cost of a successful task rather than an isolated benchmark score.

On AutomationBench 1.0.6, agents complete workflows using 47 tools across sales, marketing, finance, support and human resources. GPT-6 Sol in maximum-effort mode scored 33.2% at a cost of $0.27 per task.

OpenAI estimates that Claude Opus 5 in maximum-effort mode scored 26.9%, while each task cost 11.1 times more. A Claude Fable 5.1 configuration that switched to Opus 5 when needed scored 31.4%, with a task cost more than 8.9 times higher than Sol’s. OpenAI says that understates Fable’s actual cost because the calculation excludes Opus calls made in about 40% of tasks.

GPT-6 Astra in low-effort mode scored 30.3% on the same benchmark, below Sol’s 33.2%, while each task cost 3.9 times more.

Other tests point in the same direction:

These comparisons need a qualification. OpenAI’s charts compare Sol with Opus 5, but Anthropic released Opus 5.5 only hours before OpenAI’s scheduled announcement. Anthropic’s published benchmarks show improvement over Opus 5, but do not include GPT-6 Sol results. There is no public, same-test comparison showing which model completes a successful task more cheaply.

OpenAI also has not provided GPT-6 Sol results on the same test setup against Gemini 3.8 Flash, Grok 4.7 or Xiaomi’s MiMo-V2.6 models. Token prices are directly comparable; the cost of completing a task is not.

That missing comparison is the announcement’s central gap. The useful benchmark is not price per token or even score per benchmark run; it is successful work per dollar after retries, context reuse, latency and human correction. OpenAI makes that case rhetorically, but the cross-provider evidence is not yet complete.

Caching is part of the price

OpenAI says GPT-6 raises the standard prompt-cache hit rate and gives a 90% discount on cached input tokens. Developers can change the reasoning level or available tools without resetting previously stored context, and can explicitly define the prompt-prefix boundary eligible for caching.

That matters for long-lived coding and business agents, which repeatedly send system prompts, files, tool descriptions and conversation history.

Other providers use similar pricing:

Xiaomi’s nominal cache prices are far lower, but self-hosting brings equipment, orchestration and resource-utilization costs. OpenAI cites GitHub as an example of the effect at scale: the company says its improvements reduced by more than 50% the share of Copilot prompt tokens requiring new processing, across billions of requests to OpenAI models.

OpenAI also says earlier price cuts helped Replit introduce a free mode for millions of users. After an 80% price reduction in July, usage of GPT-5.6 Luna grew more than tenfold, according to OpenAI.

Reliability improves, but rare failures remain

OpenAI reports lower rates of several unwanted behaviors.

In an internal deception test designed to provoke dishonest actions, GPT-6 Sol deceived the evaluator in 1.3% of cases, compared with 10.4% for GPT-5.6 Sol. Luna’s rate fell from 9.5% to 2.8%.

In another test, an agent received a broken search tool and was evaluated on whether it acknowledged the failure instead of guessing. Sol failed to report the problem in 5.4% of cases, down from 77.8%. Luna’s rate fell from 78.3% to 30.2%.

OpenAI recorded no attempts by Sol to bypass automated safety checks in Codex tests. Luna attempted to bypass them in 0.3% of runs, down from 3.5%, with no successful bypasses.

The results were not uniformly strong. When models received an explicit warning such as “access forbidden,” Sol still attempted to bypass the restriction in 64.4% of adversarial runs, compared with 68.2% for GPT-5.6 Sol. Luna’s rate was 42.4%, down from 76.5%.

OpenAI says these were deliberately difficult, mostly low-risk tests run without all the system protections used in its products. They should not be treated as ordinary-use failure rates.

API first, product access later

GPT-6 Sol and Luna are available through the OpenAI API under the names gpt-6-sol and gpt-6-luna.

OpenAI is gradually adding them to ChatGPT Work and Codex for Plus, Pro, Business and Enterprise customers. Free and Go users may access Luna through the desktop application. The models are not yet available in Chat itself, and OpenAI plans to expand Work and Codex access throughout the day to preserve service stability.

For enterprise buyers, the important shift is not simply a cheaper API. It is the disappearance of the single “best model” as the default architecture:

Permanent Sol and Luna pricing makes that routing easier to build into production systems than a temporary discount would. But the competitive pressure is moving beyond list prices. Anthropic’s Opus 5.5 and Xiaomi’s MiMo-V2.6 make the next comparison less about model prestige than about how much useful work each system completes before a human has to intervene.