GPT-5.6 Luna
Overview
GPT-5.6 Luna is OpenAI's cost-oriented GPT-5.6 tier, made generally available on July 9, 2026. In the API, its model ID is gpt-5.6-luna, and OpenAI positions it for efficient, high-volume workloads.
Key Features
- Cost-Oriented GPT-5.6 Tier: Designed for high-volume workloads where lower token cost and responsiveness are important.
- Expanded Reasoning Controls: Supports
none,low,medium,high,xhigh, andmaxreasoning effort; GPT-5.6 also supports Pro mode throughreasoning.mode: "pro". - Agentic Tooling: Supports Responses API workflows, functions, web search, file search, computer use, Programmatic Tool Calling, persisted reasoning, and multi-agent beta.
- Large Working Context: Supports a 1.05M-token context window and up to 128K output tokens, with a February 16, 2026 knowledge cutoff.
Best Use Cases
- High-Volume Assistance: Summarization, rewriting, classification, extraction, routing, and lightweight analysis.
- Responsive Product Features: User-facing assistants, workflow copilots, and routine automation with predictable cost constraints.
- First-Pass Technical Work: Initial issue triage, draft code review, test-case generation, and document analysis before escalating difficult cases to Terra or Sol.
Capabilities and Limitations
| Capability | Description |
|---|---|
| Reasoning | Supports none, low, medium, high, xhigh, and max; API Pro mode is enabled with reasoning.mode: "pro" rather than a separate Pro model slug. |
| Coding | Official benchmark examples include SWE-Bench Pro 62.7%, DeepSWE v1.1 67.2%, and Terminal-Bench 2.1 84.7%. |
| Creative Writing | General text generation is supported. |
| Multimodal | Text and image input with text output; GPT-5.6 preserves original image dimensions for original or auto detail. |
| Response Speed | Public API latency depends on reasoning effort, tools, and processing tier; Luna is positioned for efficient, high-volume use. |
| Context Window | 1.05M tokens. |
| Max Output | 128K tokens. |
| Tool Use | Functions, web search, file search, computer use, Programmatic Tool Calling, persisted reasoning, and multi-agent beta. |
| Multilingual | OpenAI API model documentation lists multilingual support across current models. |
Known Limitations
- Luna is positioned below Terra and Sol for capability, so complex reasoning, sensitive technical work, and quality-first workflows may need escalation.
- Higher reasoning efforts, Pro mode, multi-agent runs, and long-context use can increase latency and cost.
- Cyber and biology safeguards may refuse, block, or pause sensitive requests, including some legitimate dual-use work.
- Prompt cache writes are billed on GPT-5.6 models, so cache breakpoints should be monitored with
cache_write_tokensandcached_tokens.
Standard Pricing
The token prices below are shown in USD per 1 million tokens; web search is billed in USD per use.
| Context | Input (USD / 1M Tokens) | 30m Cache Write (USD / 1M Tokens) | 12h Cache Write (USD / 1M Tokens) | Cache Read (USD / 1M Tokens) | Output (USD / 1M Tokens) | Web Search (USD / use) |
|---|---|---|---|---|---|---|
| Up to 272K input tokens | $0.20 | $0.25 | $0.50 | $0.02 | $1.20 | $0.01 |
| More than 272K input tokens | $0.40 | $0.50 | $1.00 | $0.04 | $1.80 | $0.01 |
- Above 272K input tokens, input and cache-write/read rates are 2x the base tier and output rates are 1.5x. The higher rates apply to the entire request, not only the portion above the threshold.
- Web search costs
10,000 Credits($0.01) per use and is not affected by the context tier.
Credits settlement: B.AI converts charges at 1 USD = 1,000,000 Credits and deducts Credits from the account balance.
Pricing note
Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through account benefits. Specific prices and account benefits are subject to the platform display and final billing records.
- Cache-write rates replace the ordinary input charge for those tokens, rather than adding to it. A 30-minute cache write costs 1.25x the applicable input rate; a 12-hour write costs 2.5x. The default minimum lifetime is 30 minutes. Both lifetimes have the same cache-read rate, 0.1x the applicable input rate.