OpenAI slipper GPT-6 Sol og Luna til API med lavere publiserte tokenpriser

OpenAI released GPT-6 Sol and GPT-6 Luna to the API on September 22, adding two new GPT-6 models for developers using the Responses API and Chat Completions API. The launch matters less as a simple model-list update than as a pricing and routing event: OpenAI is positioning Sol as a stronger general-purpose model and Luna as a much cheaper efficiency tier, while cloud and gateway partners made both models available on the same day.

OpenAI lists GPT-6 Sol under the model ID gpt-6-sol and GPT-6 Luna under gpt-6-luna. For standard pricing on prompts up to 272,000 input tokens, GPT-6 Sol is listed at $2 per 1 million input tokens, $0.20 per 1 million cached input tokens, and $10 per 1 million output tokens. GPT-6 Luna is listed at $0.10 per 1 million input tokens, $0.01 per 1 million cached input tokens, and $0.50 per 1 million output tokens.

That gap is large enough to change day-to-day architecture decisions. A team that treats “the latest OpenAI model” as a single default now has a stronger reason to separate workloads by risk, complexity, latency sensitivity, and output length. The cheaper model may be sufficient for classification, extraction, formatting, routing, summarization, and routine assistant steps. The higher-priced model may be reserved for complex reasoning, difficult coding, high-value customer interactions, or escalation paths where accuracy carries more business weight.

What changed

The immediate change is availability. OpenAI says both models are live through its API surfaces, including the Responses API and Chat Completions API. AWS says GPT-6 Sol and GPT-6 Luna are generally available on Amazon Bedrock as of the same date. Microsoft says both joined the generally available model lineup in Microsoft Foundry. Vercel and Netlify also list same-day gateway availability.

Same-day distribution across OpenAI, AWS, Microsoft, and developer gateways reduces the friction usually associated with new model adoption. Teams that already centralize procurement, access, audit logs, or networking through Bedrock or Foundry can test the models without opening a direct OpenAI integration. Teams that use OpenAI-compatible gateways can often add models through routing configuration rather than building a new provider adapter from scratch.

The second change is the explicit price spread between Sol and Luna. Luna’s listed input and output prices are one-twentieth of Sol’s standard prices, before considering cache behavior. That does not mean Luna is the right model for every low-cost task, or that Sol is automatically better for every complex one. It does mean cost controls can no longer be limited to aggregate monthly spend. Model choice becomes an application-level design variable.

Why businesses should care

For engineering leaders, the release creates a new evaluation queue. GPT-6 Sol and Luna need to be tested not only against GPT-5.6-era OpenAI models, but also against Anthropic, Google, xAI, Alibaba, Z.ai, and other models already exposed through enterprise routing layers. The relevant question is not “which model is best?” It is which model is good enough for each workflow, at what latency, under what governance policy, and with what failure mode.

For finance and platform teams, Luna’s price point puts pressure on model selection discipline. If developers manually choose Sol for every request because it is the newer flagship option, output-heavy workloads can become expensive quickly. If policy routes too aggressively to Luna, quality regressions may surface in support automation, code generation, agent planning, or executive-facing summaries. The practical answer is usually tiered routing backed by evaluation data, budget limits, and per-team attribution.

This is where a well-run enterprise LLM API layer becomes useful. A central control plane can expose both models while still enforcing organization-level permissions, project budgets, model allowlists, fallback rules, and usage analytics. For companies running B2B LLM or another multi-B2B LLMway pattern, GPT-6 Sol and Luna are exactly the kind of release that tests whether model access is governed centrally or scattered across individual application teams.

Developer consequences

Developers should expect three immediate work items. First, benchmark both models on real prompts rather than relying on vendor positioning. OpenAI and AWS make performance and workload-fit claims, including comparisons to prior models, but those are vendor claims until teams validate them against their own tasks.

Second, update routing policies. The price difference between Sol and Luna supports more granular rules: Luna for first-pass extraction, metadata enrichment, lightweight chat, or internal draft generation; Sol for escalations, complex agent steps, difficult tool-use sequences, or tasks with high cost of error. The exact split will vary by application, but the release makes a single-model default harder to justify.

Third, revisit caching strategy. OpenAI’s listed cached-input prices are materially lower than uncached input prices for both models. Teams with repeated system prompts, retrieval templates, policy instructions, or stable context blocks should measure whether prompt caching changes total cost enough to influence model choice. Long-context applications are especially exposed because repeated input can dominate spend even when output is controlled.

What remains uncertain

The public facts establish availability and listed pricing, but not independent quality. OpenAI and AWS describe model capabilities and performance improvements, yet businesses still need their own evaluations for accuracy, latency, tool-use reliability, refusal behavior, and cost per successful task. The most important metric is rarely token price alone. A cheaper model that requires more retries, more human review, or more compensating logic can be more expensive operationally than it looks on a rate card.

Availability through Bedrock, Foundry, Vercel, and Netlify also does not mean every enterprise deployment constraint is identical across channels. Regional availability, logging behavior, data-processing terms, rate limits, context limits, and support paths can differ by platform. Procurement teams should verify those details before moving production traffic, especially in regulated environments or customer-facing automation.

The broader signal is clear: frontier-model launches are becoming multi-cloud, gateway-ready events on day one. That makes adoption faster, but it also raises the bar for governance. The teams best positioned to benefit from GPT-6 Sol and Luna will be the ones that can compare models quickly, route by workload, attribute cost by project, and change policy without rewriting applications.