{"slug": "gpt-5-6-luna-drops-80-rethink-your-model-routing-now", "title": "GPT-5.6 Luna Drops 80%: Rethink Your Model Routing Now", "summary": "On July 30, OpenAI cut the input price of GPT-5.6 Luna by 80% from $1.00 to $0.20 per million tokens and output from $6.00 to $1.20, while Terra dropped 20% and Sol gained a Fast mode at double the price for 2.5x speed. The move follows Chinese models capturing 66.5% of OpenRouter token volume in the week of July 20–26, according to a CNBC investigation. Developers are advised to route tasks across Luna, Terra, and Sol to cut costs without sacrificing accuracy, as Luna's long-context recall score is 41.3%.", "body_md": "On July 30, OpenAI cut GPT-5.6 Luna’s input price by 80% — from $1.00 to $0.20 per million tokens, with output falling from $6.00 to $1.20. The mid-tier Terra dropped 20%. Sol held flat but gained a new Fast mode at double the price for 2.5x the speed. This is the sharpest price move OpenAI has made since ChatGPT launched, and it is not generosity: Chinese models just claimed 66.5% of all OpenRouter token volume in the week of July 20–26. OpenAI is not giving you a discount. It is fighting for your API calls.\n\n## What the Three Tiers Actually Do\n\nGPT-5.6 launched on July 9 as three distinct models, not one model with different caps. Understanding the split matters before you touch your routing config.\n\n**Sol** is the flagship. Long-horizon professional coding, complex multi-step agent workflows, computer use, cybersecurity research — Sol is where you send the hard stuff. It set a new benchmark record on [Agents’ Last Exam](https://openai.com/index/gpt-5-6/) (53.6), and it costs $5/$30 per million input/output tokens. Fast Mode raises that to $10/$60 for 2.5x throughput.\n\n**Terra** is the everyday workhorse. Planning, drafting, code generation with tool calls, most production workflows — Terra handles what Sol would overkill. Post-cut it runs at $2/$12 per million.\n\n**Luna** is the high-volume lane. Classification, triage, summarization, intent detection, translation, simple extractions, annotation pipelines. Now $0.20/$1.20 per million. Luna costs one-twenty-fifth what Sol costs on input.\n\n## What Luna Cannot Do (Read This Before Routing)\n\nThe price is right. The benchmarks are not the whole story. [Luna’s long-context recall score (MRCR) is 41.3%](https://artificialanalysis.ai/models/gpt-5-6-luna) — significantly below Terra and Sol. That 1,050,000-token context window exists on paper, but stuffing an entire codebase or log file into Luna and expecting reliable retrieval is going to burn you. Leave long-form work to Terra or above.\n\nLuna also scores below GPT-5.5 on Terminal-Bench 2.1 and BrowseComp. If you had GPT-5.5 handling terminal or browser-based tasks in your pipeline, do not swap it for Luna to save money. You will get worse results than what you already had.\n\nOne more thing: use the explicit model ID `gpt-5.6-luna`\n\n. The generic `gpt-5.6`\n\nalias routes to Sol. If you are pricing your app on Luna’s rates and accidentally sending calls to Sol, that is a 25x bill shock waiting to happen.\n\n## The Routing Playbook\n\nThe pattern consolidating in the developer community after the cut is a simple three-lane split:\n\n**Luna**— first-pass triage, summaries, labels, intent detection, simple routing decisions, high-volume annotation** Terra**— main execution: planning, drafting, code generation, research, tool-heavy tasks** Sol**— escalation: long-horizon workflows, complex reasoning, computer use, final review on critical outputs\n\nThe cost math is concrete. Take a product processing 10 million input tokens and 2 million output tokens per month:\n\n- All-Sol (old habit): ~$110/month\n- All-Luna: ~$4.40/month\n- Routed (80% Luna, 20% Terra): ~$12.32/month\n- Routed (80% Luna, 20% Sol): ~$25.52/month\n\nThe all-Luna number is misleading — you cannot run everything through Luna without accuracy regressions on anything requiring reasoning. The 80/20 Luna-Terra split delivers real savings without sacrificing output quality on the tasks that matter.\n\nTargeting Luna in the API is straightforward:\n\n```\nresponse = openai.chat.completions.create(\n    model=\"gpt-5.6-luna\",  # explicit — \"gpt-5.6\" routes to Sol\n    messages=[{\"role\": \"user\", \"content\": prompt}]\n)\n```\n\n## Why OpenAI Cut Now\n\nA [CNBC investigation published July 7](https://www.cnbc.com/2026/07/30/open-ai-price-cut-gpt.html) found that Chinese AI models had captured 46% of US enterprise token usage on OpenRouter. By the week of July 20–26, Chinese models were processing 66.5% of all OpenRouter volume — 38.6 trillion tokens in a single week. DeepSeek, Kimi, and MiniMax are not niche alternatives anymore; they are the default for a significant portion of production workloads.\n\nDeepSeek V4 launched on the same day as OpenAI’s announcement — July 30. That timing was not a coincidence. OpenAI’s Luna cut puts its input price at $0.20 per million, undercutting DeepSeek V4 Pro’s promotional rate of $0.435 per million. DeepSeek wins on output ($0.87 vs $1.20), but for classification-heavy workloads where input tokens dominate, Luna is now the cheaper option.\n\nOpenAI says the cuts were [funded by a rewrite of Sol’s inference stack](https://venturebeat.com/technology/ai-price-wars-openai-cuts-gpt-5-6-luna-prices-by-80-as-model-competition-shifts-toward-cost). Infrastructure efficiency on the flagship unlocked margin on the cheaper tiers. Whether that is the full story or a face-saving explanation for competitive necessity, the result for developers is the same: Luna at $0.20 is real.\n\n## What Comes Next\n\nThis price cut is not an endpoint. Chinese models processed more US enterprise tokens than American models for thirteen consecutive weeks before OpenAI responded. The AI API commodity race is structurally underway, and the pressure on Anthropic and Google to follow with their own efficiency tiers is now direct. Expect further cuts — or expanded context windows at current prices — within the next two quarters.\n\nFor now: audit your pipelines, identify the tasks that genuinely need reasoning versus the ones that just need fast and cheap, and [route accordingly](https://community.openai.com/t/announcing-a-major-price-drop-for-5-6-terra-and-luna-and-fast-mode-for-5-6-sol/1388484). Luna at $0.20 changes the math on use cases that were not viable six weeks ago.", "url": "https://wpnews.pro/news/gpt-5-6-luna-drops-80-rethink-your-model-routing-now", "canonical_source": "https://byteiota.com/gpt56-luna-price-cut-model-routing/", "published_at": "2026-08-03 18:09:09+00:00", "updated_at": "2026-08-03 18:23:42.904387+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products"], "entities": ["OpenAI", "GPT-5.6 Luna", "GPT-5.6 Terra", "GPT-5.6 Sol", "OpenRouter", "CNBC", "DeepSeek", "Kimi"], "alternates": {"html": "https://wpnews.pro/news/gpt-5-6-luna-drops-80-rethink-your-model-routing-now", "markdown": "https://wpnews.pro/news/gpt-5-6-luna-drops-80-rethink-your-model-routing-now.md", "text": "https://wpnews.pro/news/gpt-5-6-luna-drops-80-rethink-your-model-routing-now.txt", "jsonld": "https://wpnews.pro/news/gpt-5-6-luna-drops-80-rethink-your-model-routing-now.jsonld"}}