OpenAI GPT-6 Sol & Luna Pricing: What It Means (2026)

OpenAI launched GPT-6 Sol and GPT-6 Luna on September 22, 2026, extending the GPT-6 family below Astra. Both models are available now through the API as gpt-6-sol and gpt-6-luna; OpenAI also began a gradual rollout in ChatGPT Work and Codex.

The pricing move is unusually direct: the new models cost less than the GPT-5.6 tiers they replace while adding the GPT-6 generation’s coding, computer-use, factuality, tool, and alignment improvements. The live table and chart above pull today’s rates from our maintained dataset.

What changed

GPT-6 Sol is OpenAI’s balanced tier for complex coding and agentic workflows. GPT-6 Luna is the efficiency tier for focused, repeatable, high-volume work. Astra remains the maximum-capability option.

Both new models accept text and images, produce text, and expose a 1.05-million-token context window with up to 922,000 input tokens and 128,000 output tokens. They support the Responses API, Chat Completions, Batch, prompt caching, structured outputs, function calling, file search, web search, code execution, computer use, MCP, and other hosted tools. Fine-tuning, Realtime, audio, and video endpoints are not supported.

Reasoning effort runs from none through max. Chat Completions function calling requires none; use Responses for built-in tools and reasoning workflows.

GPT-6 Sol and Luna pricing impact

The live comparison above is the old-versus-new pricing table. Standard short-context input, cached-input, and cache-write rates are half the GPT-5.6 equivalents. Sol output is also halved, while Luna output receives a larger reduction.

Requests above 272,000 input tokens reprice the full request: input and cache categories double, while output rises by half. Batch and Flex cost half the Standard rate; Fast mode costs twice Standard. Eligible regional processing adds an uplift, and EU data residency is available only with Standard processing.

Prompt-cache reads cost one-tenth of uncached input. That matters for agents carrying stable instructions, tool definitions, repository maps, or long conversation prefixes. Track cache hit rate in OpenAI’s dashboard rather than assuming every repeated prompt qualifies.

Use the token calculator for workload-specific math and the OpenAI pricing page for the complete current model table.

How the capability claims compare

OpenAI reports that Sol at xhigh reasoning scored 33.2% on AutomationBench, ahead of low-effort Astra and the cited Claude results at a lower provider-calculated cost per task. On DeepSWE, Sol at max reached 68.8%, while Luna reached 66.6%. On OSWorld, Sol at xhigh scored 60.5%.

Those are vendor-reported evaluations, not a universal buying verdict. Reasoning level, tool harness, retries, and stopping rules all change task cost. OpenAI also says Luna can match GPT-5.6 Sol on selected factuality and computer-use settings at a fraction of the cost, but buyers should reproduce that result on their own failure modes.

For an outside baseline, compare current Anthropic pricing and the broader AI API pricing comparison.

Who benefits—and who should wait

Luna winners: support routing, extraction, classification, RAG, code review triage, and other high-volume work where outputs are easy to verify. Its lower output rate is especially useful for response-heavy workflows.

Sol winners: coding agents, browser and computer-use automation, multi-tool research, and professional workflows where a stronger first attempt can avoid retries or reviewer time.

Who should wait: teams with a validated GPT-5.6 production baseline and no spare evaluation capacity. Lower list prices do not justify an untested model swap when behavior, latency, or tool use could regress. Astra users should also keep Astra for tasks where Sol does not reproduce the accepted-result rate.

What developers should do now

1. Add the new model IDs to allowlists, observability, budget controls, and eval reports without silently repointing old labels.
2. Replay a representative task set with identical prompts, tools, reasoning effort, and stopping rules.
3. Measure accepted results, retries, latency, cache hits, output length, tool charges, and reviewer minutes—not token rates alone.
4. Test short and long context separately because crossing the threshold changes the whole request’s rate.
5. Start routine traffic on Luna, escalate failed or uncertain work to Sol, and keep Astra as the measured premium route.
6. Verify regional-processing and product entitlements before moving regulated workloads.

For a managed open-model control, benchmark Novita’s OpenAI-compatible routes on the same accepted-result rubric.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect this analysis.

Independent Labs result

Both exact OpenRouter routes completed AI Pricing Guru’s same 49 deterministic tasks with zero transport or model errors. GPT-6 Luna scored 49/49 at $0.0013686 total ($0.00002793 per correct answer), while GPT-6 Sol scored 49/49 at $0.015702 total ($0.00032045 per correct answer).

These narrow machine-graded tasks establish route availability and a repeatable cost baseline; they do not reproduce OpenAI’s coding, computer-use, or agent benchmarks. See the current Labs leaderboard and coverage decision for the measured scope.

Bottom line

GPT-6 Sol and Luna reset OpenAI’s price-performance ladder: the replacement tiers are cheaper than GPT-5.6 while inheriting the GPT-6 tool and capability stack. Luna is the first test for defined high-volume work; Sol is the balanced coding and agent route; Astra remains the escalation tier.

The practical migration is evaluation-first routing, not a global model-string replacement. Move traffic only where the new tier lowers cost per accepted result.

Sources: OpenAI’s official GPT-6 Sol and Luna announcement, GPT-6 Sol model documentation, GPT-6 Luna model documentation, and API pricing. Pricing, availability, context limits, endpoints, and features verified September 22, 2026 at 19:18 UTC.