{"id":343,"date":"2026-09-24T00:25:49","date_gmt":"2026-09-23T16:25:49","guid":{"rendered":"https:\/\/qoraapi.com\/blog\/ai-api-cost-index-2026-2\/"},"modified":"2026-10-27T18:10:08","modified_gmt":"2026-10-27T10:10:08","slug":"ai-api-cost-index-2026","status":"publish","type":"page","link":"https:\/\/qoraapi.com\/blog\/ai-api-cost-index-2026\/","title":{"rendered":"The 2026 AI API Cost Index: What 22 Models Really Cost Per Workload"},"content":{"rendered":"\n<style>\n.qapi-table{overflow-x:auto;margin:18px 0 22px;border:1px solid #e5e7eb;border-radius:12px}\n.qapi-table table{width:100%;border-collapse:collapse;font-size:13.5px;font-family:-apple-system,BlinkMacSystemFont,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif}\n.qapi-table th{text-align:left;padding:9px 10px;background:#f9fafb;font-size:10.5px;font-weight:700;text-transform:uppercase;letter-spacing:.03em;color:#6b7280;border-bottom:1px solid #e5e7eb;white-space:nowrap}\n.qapi-table th.num{text-align:right}\n.qapi-table td{padding:8px 10px;border-bottom:1px solid #e5e7eb;color:#111827;vertical-align:middle;font-size:13px}\n.qapi-table tr:last-child td{border-bottom:0}\n.qapi-table tbody tr:first-child td{background:#ecfdf5}\n.qapi-table td.num{text-align:right;font-variant-numeric:tabular-nums;white-space:nowrap}\n.qapi-table td:first-child{font-weight:600}\n.qapi-table td:first-child .qapi-tag{color:#4f46e5}\n.qapi-tag{display:inline-block;font-size:10px;font-weight:700;padding:1px 6px;border-radius:999px;background:#eef2ff;color:#4f46e5;margin-left:6px;vertical-align:1px}\n.qapi-inline{width:100%;border-collapse:collapse;margin:18px 0 22px;border:1px solid #e5e7eb;border-radius:12px;font-family:-apple-system,BlinkMacSystemFont,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif;font-size:14px}\n.qapi-inline td{padding:11px 14px;border-bottom:1px solid #e5e7eb;color:#374151}\n.qapi-inline tr:last-child td{border-bottom:0}\n.qapi-inline td:last-child{text-align:right;font-variant-numeric:tabular-nums;white-space:nowrap}\n.qapi-inline strong{color:#111827}\n.qapi-cap{font-size:12.5px;color:#6b7280;margin:-12px 0 22px}\n.qapi-meta{font-size:13.5px;color:#4b5563;background:#f9fafb;border-left:3px solid #4f46e5;padding:10px 14px;border-radius:0 8px 8px 0;margin:0 0 16px}\n.qapi-findings{margin:26px 0;padding:22px 24px;border:1px solid #e5e7eb;border-radius:14px;background:#f9fafb;font-family:-apple-system,BlinkMacSystemFont,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif}\n.qapi-findings-h{margin:0 0 14px;font-size:16px;font-weight:800;color:#111827;letter-spacing:-.01em}\n.qapi-findings ol{margin:0;padding-left:22px}\n.qapi-findings li{margin:0 0 12px;font-size:14.5px;line-height:1.65;color:#374151}\n.qapi-findings li:last-child{margin-bottom:0}\n.qapi-findings strong{color:#111827}\n.qapi-dl{margin:24px 0;padding:20px 22px;border:1px solid #e5e7eb;border-radius:12px;background:#fff;font-family:-apple-system,BlinkMacSystemFont,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif}\n.qapi-dl b{display:block;font-size:15px;font-weight:800;color:#111827;margin-bottom:5px}\n.qapi-dl span{display:block;font-size:13.5px;color:#6b7280;margin-bottom:12px}\n.qapi-dl a{display:inline-block;background:#4f46e5;color:#fff;font-weight:700;font-size:13.5px;text-decoration:none;padding:10px 18px;border-radius:8px}\n.qapi-cta{margin:26px 0;padding:22px;border-radius:14px;background:linear-gradient(135deg,#4f46e5,#7c3aed);color:#fff;font-family:-apple-system,BlinkMacSystemFont,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif}\n.qapi-cta strong{display:block;font-size:19px;font-weight:800;margin-bottom:7px;letter-spacing:-.01em}\n.qapi-cta p{margin:0 0 14px;font-size:14.5px;color:rgba(255,255,255,.93);line-height:1.6}\n.qapi-cta p:last-child{margin:0}\n.qapi-cta a{display:inline-block;background:#fff;color:#4f46e5;font-weight:700;font-size:14px;text-decoration:none;padding:11px 22px;border-radius:9px}\n.qapi-cta a:hover{background:#f5f3ff}\n@media (max-width:620px){.qapi-table table{font-size:12px}.qapi-table th,.qapi-table td{padding:7px 8px}.qapi-table .hide-m{display:none}.qapi-findings{padding:18px}.qapi-cta{padding:18px}}\n\/* qapi-cite:start *\/.qapi-cite{margin:26px 0;padding:18px 20px;border:1px dashed #c7d2fe;border-radius:12px;background:#f8f9ff;font-family:-apple-system,BlinkMacSystemFont,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif}.qapi-cite b{display:block;font-size:14px;font-weight:800;color:#111827;margin-bottom:6px}.qapi-cite p{margin:0 0 10px;font-size:13.5px;color:#4b5563;line-height:1.6}.qapi-cite p:last-child{margin:0}.qapi-cite-q{font-family:ui-monospace,SFMono-Regular,Menlo,Consolas,monospace;font-size:12.5px;color:#111827;background:#fff;border:1px solid #e5e7eb;border-radius:8px;padding:10px 12px;overflow-wrap:anywhere;line-height:1.55}.qapi-cite-m{font-size:12.5px;color:#6b7280}\/* qapi-cite:end *\/<\/style>\n\n<p>Almost every AI API pricing page answers the wrong question. It tells you the cost of a million tokens, when what you actually need to know is the cost of <strong>your<\/strong> workload \u2014 one support reply, one code review, one document summary \u2014 multiplied by how many of them you run.<\/p>\n\n<p>This index closes that gap. It takes the published list rates of <strong>22 models across four providers<\/strong>, verified against each vendor&#8217;s official pricing page in September 2026, and prices six realistic production workloads against all of them. Every number is reproducible from the methodology at the bottom of this page, and the raw dataset is available to download.<\/p>\n\n\n<div class=\"qapi-findings\">\n<p class=\"qapi-findings-h\">Five findings that per-token price lists hide<\/p>\n<ol>\n<li><strong>The same workload costs up to 200\u00d7 more depending on the model.<\/strong>\nEvery one of the six workloads below spans more than 100\u00d7 from cheapest to most expensive\nmodel. Per-token pricing makes that gap invisible because nobody buys a million tokens in the\nabstract \u2014 they buy an answer.<\/li>\n<li><strong>Crossing a long-context threshold adds about 98%, not the 2\u00d7 you would expect\nfrom the input rate.<\/strong> Six of the models here switch to long-context pricing above a\nthreshold, and the multiplier applies to the <em>entire<\/em> request \u2014 input at 2\u00d7\nand output at 1.5\u00d7.<\/li>\n<li><strong>One extra input token can cost $5,540 per month.<\/strong> At 2,000 calls a month,\nGPT-6 Astra costs $5,640 at 272,000 input tokens and $11,180 at 272,001. The cliff is a\ndiscontinuity, not a slope.<\/li>\n<li><strong>Splitting a long prompt in two halves the bill.<\/strong> Two 150K-token calls cost\n50% less than one 300K-token call on every cliff-affected model, because each request stays on\nthe cheap side of the threshold.<\/li>\n<li><strong>Prompt caching pays for itself on the second request.<\/strong> Cache writes cost\n1.25\u00d7 normal input and reads cost 0.025\u20130.1\u00d7, so break-even lands between 1.28\nand 1.39 reads on every provider that publishes both rates.<\/li>\n<\/ol>\n<\/div>\n\n\n<h2>Cost per workload<\/h2>\n\n<p>Each workload below states its token assumptions explicitly, because the assumptions matter more than the rates. Change the input-to-output ratio and the ranking can move: output tokens are priced 3\u20136\u00d7 higher than input almost everywhere, so a workload that emits a lot is far more sensitive to the output rate than to the input rate.<\/p>\n\n<p>All figures are standard-tier list prices. Column meanings: <strong>Monthly<\/strong> is the no-cache, standard-tier bill; <strong>Per 1k req<\/strong> is the same figure divided by thousands of requests; <strong>Cached 70%<\/strong> assumes a 70% prompt-cache hit rate; <strong>Batch<\/strong> applies the 50% batch discount. The cheapest row in each table is highlighted.<\/p>\n\n\n<h3>Classification \/ routing<\/h3>\n<p>Ticket triage, intent labels, spam checks. Short in, very short out.<\/p>\n<p class=\"qapi-meta\">1,500 input + 30 output tokens per request \u00d7 500,000 requests\/month\n= <strong>750.0M input and 15.0M output tokens per month<\/strong>.\nCheapest: <strong>Gemini 2.5 Flash-Lite<\/strong> at $81.0. Most expensive: <strong>GPT-5.6 Cyber<\/strong> at $10,500.<\/p>\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"num\">Monthly<\/th><th class=\"num hide-m\">Per 1k req<\/th><th class=\"num\">Cached 70%<\/th><th class=\"num hide-m\">Batch<\/th><\/tr><\/thead><tbody><tr><td>Gemini 2.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$81.0<\/td><td class=\"num hide-m\">$0.16<\/td><td class=\"num\">$33.8<\/td><td class=\"num hide-m\">$40.5<\/td><\/tr><tr><td>GPT-6 Luna <span class=\"qapi-tag\">OpenAI<\/span><\/td><td class=\"num\">$82.5<\/td><td class=\"num hide-m\">$0.17<\/td><td class=\"num\">$35.2<\/td><td class=\"num hide-m\">$41.2<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$122<\/td><td class=\"num hide-m\">$0.24<\/td><td class=\"num\">$44.3<\/td><td class=\"num hide-m\">$60.8<\/td><\/tr><tr><td>Gemini 3.1 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$210<\/td><td class=\"num hide-m\">$0.42<\/td><td class=\"num\">$91.9<\/td><td class=\"num hide-m\">$105<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$243<\/td><td class=\"num hide-m\">$0.49<\/td><td class=\"num\">$88.7<\/td><td class=\"num hide-m\">$122<\/td><\/tr><tr><td>Gemini 3.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$262<\/td><td class=\"num hide-m\">$0.53<\/td><td class=\"num\">$121<\/td><td class=\"num hide-m\">$131<\/td><\/tr><tr><td>Gemini 2.5 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$262<\/td><td class=\"num hide-m\">$0.53<\/td><td class=\"num\">$121<\/td><td class=\"num hide-m\">$131<\/td><\/tr><tr><td>DeepSeek V4 Pro (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$525<\/td><td class=\"num hide-m\">$1.05<\/td><td class=\"num\">$190<\/td><td class=\"num hide-m\">$262<\/td><\/tr><tr><td>Gemini 3.8 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$619<\/td><td class=\"num hide-m\">$1.24<\/td><td class=\"num\">$264<\/td><td class=\"num hide-m\">$309<\/td><\/tr><tr><td>Haiku 4.5 <span class=\"qapi-tag\">Anthropic<\/span><\/td><td class=\"num\">$825<\/td><td class=\"num hide-m\">$1.65<\/td><td class=\"num\">$352<\/td><td class=\"num hide-m\">$412<\/td><\/tr><\/tbody><\/table><\/div>\n<p class=\"qapi-cap\">Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the <a href=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/ai-api-cost-index-2026.csv\">downloadable dataset<\/a>.<\/p>\n\n\n<h3>Customer support reply<\/h3>\n<p>System prompt + retrieved policy + short answer.<\/p>\n<p class=\"qapi-meta\">3,000 input + 400 output tokens per request \u00d7 50,000 requests\/month\n= <strong>150.0M input and 20.0M output tokens per month<\/strong>.\nCheapest: <strong>Gemini 2.5 Flash-Lite<\/strong> at $23.0. Most expensive: <strong>GPT-5.6 Cyber<\/strong> at $3,375.<\/p>\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"num\">Monthly<\/th><th class=\"num hide-m\">Per 1k req<\/th><th class=\"num\">Cached 70%<\/th><th class=\"num hide-m\">Batch<\/th><\/tr><\/thead><tbody><tr><td>Gemini 2.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$23.0<\/td><td class=\"num hide-m\">$0.46<\/td><td class=\"num\">$13.6<\/td><td class=\"num hide-m\">$11.5<\/td><\/tr><tr><td>GPT-6 Luna <span class=\"qapi-tag\">OpenAI<\/span><\/td><td class=\"num\">$25.0<\/td><td class=\"num hide-m\">$0.50<\/td><td class=\"num\">$15.6<\/td><td class=\"num hide-m\">$12.5<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$34.5<\/td><td class=\"num hide-m\">$0.69<\/td><td class=\"num\">$19.1<\/td><td class=\"num hide-m\">$17.2<\/td><\/tr><tr><td>Gemini 3.1 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$67.5<\/td><td class=\"num hide-m\">$1.35<\/td><td class=\"num\">$43.9<\/td><td class=\"num hide-m\">$33.8<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$69.0<\/td><td class=\"num hide-m\">$1.38<\/td><td class=\"num\">$38.1<\/td><td class=\"num hide-m\">$34.5<\/td><\/tr><tr><td>Gemini 3.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$95.0<\/td><td class=\"num hide-m\">$1.90<\/td><td class=\"num\">$66.7<\/td><td class=\"num hide-m\">$47.5<\/td><\/tr><tr><td>Gemini 2.5 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$95.0<\/td><td class=\"num hide-m\">$1.90<\/td><td class=\"num\">$66.7<\/td><td class=\"num hide-m\">$47.5<\/td><\/tr><tr><td>DeepSeek V4 Pro (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$139<\/td><td class=\"num hide-m\">$2.77<\/td><td class=\"num\">$71.6<\/td><td class=\"num hide-m\">$69.3<\/td><\/tr><tr><td>Gemini 3.8 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$188<\/td><td class=\"num hide-m\">$3.75<\/td><td class=\"num\">$117<\/td><td class=\"num hide-m\">$93.8<\/td><\/tr><tr><td>Haiku 4.5 <span class=\"qapi-tag\">Anthropic<\/span><\/td><td class=\"num\">$250<\/td><td class=\"num hide-m\">$5.00<\/td><td class=\"num\">$156<\/td><td class=\"num hide-m\">$125<\/td><\/tr><\/tbody><\/table><\/div>\n<p class=\"qapi-cap\">Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the <a href=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/ai-api-cost-index-2026.csv\">downloadable dataset<\/a>.<\/p>\n\n\n<h3>RAG answer with context<\/h3>\n<p>Retrieved passages plus instructions.<\/p>\n<p class=\"qapi-meta\">12,000 input + 600 output tokens per request \u00d7 30,000 requests\/month\n= <strong>360.0M input and 18.0M output tokens per month<\/strong>.\nCheapest: <strong>Gemini 2.5 Flash-Lite<\/strong> at $43.2. Most expensive: <strong>GPT-5.6 Cyber<\/strong> at $5,850.<\/p>\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"num\">Monthly<\/th><th class=\"num hide-m\">Per 1k req<\/th><th class=\"num\">Cached 70%<\/th><th class=\"num hide-m\">Batch<\/th><\/tr><\/thead><tbody><tr><td>Gemini 2.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$43.2<\/td><td class=\"num hide-m\">$1.44<\/td><td class=\"num\">$20.5<\/td><td class=\"num hide-m\">$21.6<\/td><\/tr><tr><td>GPT-6 Luna <span class=\"qapi-tag\">OpenAI<\/span><\/td><td class=\"num\">$45.0<\/td><td class=\"num hide-m\">$1.50<\/td><td class=\"num\">$22.3<\/td><td class=\"num hide-m\">$22.5<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$64.8<\/td><td class=\"num hide-m\">$2.16<\/td><td class=\"num\">$27.8<\/td><td class=\"num hide-m\">$32.4<\/td><\/tr><tr><td>Gemini 3.1 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$117<\/td><td class=\"num hide-m\">$3.90<\/td><td class=\"num\">$60.3<\/td><td class=\"num hide-m\">$58.5<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$130<\/td><td class=\"num hide-m\">$4.32<\/td><td class=\"num\">$55.5<\/td><td class=\"num hide-m\">$64.8<\/td><\/tr><tr><td>Gemini 3.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$153<\/td><td class=\"num hide-m\">$5.10<\/td><td class=\"num\">$85.0<\/td><td class=\"num hide-m\">$76.5<\/td><\/tr><tr><td>Gemini 2.5 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$153<\/td><td class=\"num hide-m\">$5.10<\/td><td class=\"num\">$85.0<\/td><td class=\"num hide-m\">$76.5<\/td><\/tr><tr><td>DeepSeek V4 Pro (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$273<\/td><td class=\"num hide-m\">$9.11<\/td><td class=\"num\">$112<\/td><td class=\"num hide-m\">$137<\/td><\/tr><tr><td>Gemini 3.8 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$338<\/td><td class=\"num hide-m\">$11.2<\/td><td class=\"num\">$167<\/td><td class=\"num hide-m\">$169<\/td><\/tr><tr><td>Haiku 4.5 <span class=\"qapi-tag\">Anthropic<\/span><\/td><td class=\"num\">$450<\/td><td class=\"num hide-m\">$15.0<\/td><td class=\"num\">$223<\/td><td class=\"num hide-m\">$225<\/td><\/tr><\/tbody><\/table><\/div>\n<p class=\"qapi-cap\">Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the <a href=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/ai-api-cost-index-2026.csv\">downloadable dataset<\/a>.<\/p>\n\n\n<h3>Coding agent turn<\/h3>\n<p>Repo context, tool results, patch output. Heavy output.<\/p>\n<p class=\"qapi-meta\">60,000 input + 3,000 output tokens per request \u00d7 20,000 requests\/month\n= <strong>1200.0M input and 60.0M output tokens per month<\/strong>.\nCheapest: <strong>Gemini 2.5 Flash-Lite<\/strong> at $144. Most expensive: <strong>GPT-5.6 Cyber<\/strong> at $19,500.<\/p>\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"num\">Monthly<\/th><th class=\"num hide-m\">Per 1k req<\/th><th class=\"num\">Cached 70%<\/th><th class=\"num hide-m\">Batch<\/th><\/tr><\/thead><tbody><tr><td>Gemini 2.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$144<\/td><td class=\"num hide-m\">$7.20<\/td><td class=\"num\">$68.4<\/td><td class=\"num hide-m\">$72.0<\/td><\/tr><tr><td>GPT-6 Luna <span class=\"qapi-tag\">OpenAI<\/span><\/td><td class=\"num\">$150<\/td><td class=\"num hide-m\">$7.50<\/td><td class=\"num\">$74.4<\/td><td class=\"num hide-m\">$75.0<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$216<\/td><td class=\"num hide-m\">$10.8<\/td><td class=\"num\">$92.5<\/td><td class=\"num hide-m\">$108<\/td><\/tr><tr><td>Gemini 3.1 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$390<\/td><td class=\"num hide-m\">$19.5<\/td><td class=\"num\">$201<\/td><td class=\"num hide-m\">$195<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$432<\/td><td class=\"num hide-m\">$21.6<\/td><td class=\"num\">$185<\/td><td class=\"num hide-m\">$216<\/td><\/tr><tr><td>Gemini 3.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$510<\/td><td class=\"num hide-m\">$25.5<\/td><td class=\"num\">$283<\/td><td class=\"num hide-m\">$255<\/td><\/tr><tr><td>Gemini 2.5 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$510<\/td><td class=\"num hide-m\">$25.5<\/td><td class=\"num\">$283<\/td><td class=\"num hide-m\">$255<\/td><\/tr><tr><td>DeepSeek V4 Pro (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$911<\/td><td class=\"num hide-m\">$45.5<\/td><td class=\"num\">$375<\/td><td class=\"num hide-m\">$455<\/td><\/tr><tr><td>Gemini 3.8 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$1,125<\/td><td class=\"num hide-m\">$56.2<\/td><td class=\"num\">$558<\/td><td class=\"num hide-m\">$562<\/td><\/tr><tr><td>Haiku 4.5 <span class=\"qapi-tag\">Anthropic<\/span><\/td><td class=\"num\">$1,500<\/td><td class=\"num hide-m\">$75.0<\/td><td class=\"num\">$744<\/td><td class=\"num hide-m\">$750<\/td><\/tr><\/tbody><\/table><\/div>\n<p class=\"qapi-cap\">Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the <a href=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/ai-api-cost-index-2026.csv\">downloadable dataset<\/a>.<\/p>\n\n\n<h3>Bulk document summarisation<\/h3>\n<p>Nightly pipeline over a document backlog.<\/p>\n<p class=\"qapi-meta\">8,000 input + 800 output tokens per request \u00d7 200,000 requests\/month\n= <strong>1600.0M input and 160.0M output tokens per month<\/strong>.\nCheapest: <strong>Gemini 2.5 Flash-Lite<\/strong> at $224. Most expensive: <strong>GPT-5.6 Cyber<\/strong> at $32,000.<\/p>\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"num\">Monthly<\/th><th class=\"num hide-m\">Per 1k req<\/th><th class=\"num\">Cached 70%<\/th><th class=\"num hide-m\">Batch<\/th><\/tr><\/thead><tbody><tr><td>Gemini 2.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$224<\/td><td class=\"num hide-m\">$1.12<\/td><td class=\"num\">$123<\/td><td class=\"num hide-m\">$112<\/td><\/tr><tr><td>GPT-6 Luna <span class=\"qapi-tag\">OpenAI<\/span><\/td><td class=\"num\">$240<\/td><td class=\"num hide-m\">$1.20<\/td><td class=\"num\">$139<\/td><td class=\"num hide-m\">$120<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$336<\/td><td class=\"num hide-m\">$1.68<\/td><td class=\"num\">$171<\/td><td class=\"num hide-m\">$168<\/td><\/tr><tr><td>Gemini 3.1 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$640<\/td><td class=\"num hide-m\">$3.20<\/td><td class=\"num\">$388<\/td><td class=\"num hide-m\">$320<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$672<\/td><td class=\"num hide-m\">$3.36<\/td><td class=\"num\">$343<\/td><td class=\"num hide-m\">$336<\/td><\/tr><tr><td>Gemini 3.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$880<\/td><td class=\"num hide-m\">$4.40<\/td><td class=\"num\">$578<\/td><td class=\"num hide-m\">$440<\/td><\/tr><tr><td>Gemini 2.5 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$880<\/td><td class=\"num hide-m\">$4.40<\/td><td class=\"num\">$578<\/td><td class=\"num hide-m\">$440<\/td><\/tr><tr><td>DeepSeek V4 Pro (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$1,373<\/td><td class=\"num hide-m\">$6.86<\/td><td class=\"num\">$658<\/td><td class=\"num hide-m\">$686<\/td><\/tr><tr><td>Gemini 3.8 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$1,800<\/td><td class=\"num hide-m\">$9.00<\/td><td class=\"num\">$1,044<\/td><td class=\"num hide-m\">$900<\/td><\/tr><tr><td>Haiku 4.5 <span class=\"qapi-tag\">Anthropic<\/span><\/td><td class=\"num\">$2,400<\/td><td class=\"num hide-m\">$12.0<\/td><td class=\"num\">$1,392<\/td><td class=\"num hide-m\">$1,200<\/td><\/tr><\/tbody><\/table><\/div>\n<p class=\"qapi-cap\">Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the <a href=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/ai-api-cost-index-2026.csv\">downloadable dataset<\/a>.<\/p>\n\n\n<h3>Long-document analysis (past the cliff)<\/h3>\n<p>Whole-report analysis. Past 200K\/272K every affected model charges long-context rates for the ENTIRE request.<\/p>\n<p class=\"qapi-meta\">300,000 input + 2,000 output tokens per request \u00d7 2,000 requests\/month\n= <strong>600.0M input and 4.0M output tokens per month<\/strong>.\nCheapest: <strong>Gemini 2.5 Flash-Lite<\/strong> at $61.6. Most expensive: <strong>GPT-6 Astra<\/strong> at $12,300.<\/p>\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"num\">Monthly<\/th><th class=\"num hide-m\">Per 1k req<\/th><th class=\"num\">Cached 70%<\/th><th class=\"num hide-m\">Batch<\/th><\/tr><\/thead><tbody><tr><td>Gemini 2.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$61.6<\/td><td class=\"num hide-m\">$30.8<\/td><td class=\"num\">$23.8<\/td><td class=\"num hide-m\">$30.8<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$92.4<\/td><td class=\"num hide-m\">$46.2<\/td><td class=\"num\">$30.7<\/td><td class=\"num hide-m\">$46.2<\/td><\/tr><tr><td>GPT-6 Luna <span class=\"qapi-tag\">OpenAI<\/span><\/td><td class=\"num\">$123<\/td><td class=\"num hide-m\">$61.5<\/td><td class=\"num\">$43.2<\/td><td class=\"num hide-m\">$61.5<\/td><\/tr><tr><td>Gemini 3.1 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$156<\/td><td class=\"num hide-m\">$78.0<\/td><td class=\"num\">$61.5<\/td><td class=\"num hide-m\">$78.0<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$185<\/td><td class=\"num hide-m\">$92.4<\/td><td class=\"num\">$61.3<\/td><td class=\"num hide-m\">$92.4<\/td><\/tr><tr><td>Gemini 3.5 Flash-Lite <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$190<\/td><td class=\"num hide-m\">$95.0<\/td><td class=\"num\">$76.6<\/td><td class=\"num hide-m\">$95.0<\/td><\/tr><tr><td>Gemini 2.5 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$190<\/td><td class=\"num hide-m\">$95.0<\/td><td class=\"num\">$76.6<\/td><td class=\"num hide-m\">$95.0<\/td><\/tr><tr><td>DeepSeek V4 Pro (off-peak) <span class=\"qapi-tag\">DeepSeek<\/span><\/td><td class=\"num\">$404<\/td><td class=\"num hide-m\">$202<\/td><td class=\"num\">$136<\/td><td class=\"num hide-m\">$202<\/td><\/tr><tr><td>Gemini 3.8 Flash <span class=\"qapi-tag\">Google<\/span><\/td><td class=\"num\">$465<\/td><td class=\"num hide-m\">$232<\/td><td class=\"num\">$182<\/td><td class=\"num hide-m\">$232<\/td><\/tr><tr><td>Haiku 4.5 <span class=\"qapi-tag\">Anthropic<\/span><\/td><td class=\"num\">$620<\/td><td class=\"num hide-m\">$310<\/td><td class=\"num\">$242<\/td><td class=\"num hide-m\">$310<\/td><\/tr><\/tbody><\/table><\/div>\n<p class=\"qapi-cap\">Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the <a href=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/ai-api-cost-index-2026.csv\">downloadable dataset<\/a>.<\/p>\n\n\n<h2>The context cliff: why one extra token can cost thousands<\/h2>\n\n<p>Six of the models in this index change their unit price once a request crosses a size threshold. The mechanism is the same everywhere: above the threshold, the input rate doubles and the output rate rises by half \u2014 and critically, that multiplier applies to <strong>the entire request<\/strong>, not just the portion above the threshold.<\/p>\n\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"hide-m\">Provider<\/th><th class=\"num\">Threshold (tokens)<\/th><th class=\"num hide-m\">Input rate above<\/th><th class=\"num hide-m\">Output rate above<\/th><th class=\"num\">Bill impact<\/th><\/tr><\/thead><tbody><tr><td>GPT-6 Astra<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">272,000<\/td><td class=\"num hide-m\">2.0\u00d7<\/td><td class=\"num hide-m\">1.5\u00d7<\/td><td class=\"num\">+98%<\/td><\/tr><tr><td>GPT-6 Sol<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">272,000<\/td><td class=\"num hide-m\">2.0\u00d7<\/td><td class=\"num hide-m\">1.5\u00d7<\/td><td class=\"num\">+98%<\/td><\/tr><tr><td>GPT-6 Luna<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">272,000<\/td><td class=\"num hide-m\">2.0\u00d7<\/td><td class=\"num hide-m\">1.5\u00d7<\/td><td class=\"num\">+98%<\/td><\/tr><tr><td>GPT-5.6 Sol<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">272,000<\/td><td class=\"num hide-m\">2.0\u00d7<\/td><td class=\"num hide-m\">1.5\u00d7<\/td><td class=\"num\">+98%<\/td><\/tr><tr><td>Gemini 3.1 Pro<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">200,000<\/td><td class=\"num hide-m\">2.0\u00d7<\/td><td class=\"num hide-m\">1.5\u00d7<\/td><td class=\"num\">+98%<\/td><\/tr><tr><td>Gemini 2.5 Pro<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">200,000<\/td><td class=\"num hide-m\">2.0\u00d7<\/td><td class=\"num hide-m\">1.5\u00d7<\/td><td class=\"num\">+98%<\/td><\/tr><\/tbody><\/table><\/div>\n\n<p>That produces a genuine discontinuity. Take GPT-6 Astra running 2,000 calls a month with a 2,000-token answer:<\/p>\n\n<table class=\"qapi-inline\">\n<tbody>\n<tr><td>272,000 input tokens per call<\/td><td><strong>$5,640<\/strong> \/ month<\/td><\/tr>\n<tr><td>272,001 input tokens per call<\/td><td><strong>$11,180<\/strong> \/ month<\/td><\/tr>\n<\/tbody>\n<\/table>\n\n<p>One additional token in the prompt \u2014 a slightly longer retrieved document, one more sentence of instruction \u2014 costs <strong>$5,540 a month<\/strong>. This is not a rounding error and it does not appear on any pricing page as a warning.<\/p>\n\n<h3>The workaround: chunk before you cross<\/h3>\n\n<p>Because the multiplier applies per request, splitting a long prompt into pieces that each stay under the threshold avoids it entirely. Two 150K-token calls cost exactly half of one 300K-token call:<\/p>\n\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"num\">One 300K call<\/th><th class=\"num\">Two 150K calls<\/th><th class=\"num hide-m\">Saving<\/th><\/tr><\/thead><tbody><tr><td>GPT-6 Astra<\/td><td class=\"num\">$12,300<\/td><td class=\"num\">$6,200<\/td><td class=\"num hide-m\">50%<\/td><\/tr><tr><td>GPT-6 Sol<\/td><td class=\"num\">$2,460<\/td><td class=\"num\">$1,240<\/td><td class=\"num hide-m\">50%<\/td><\/tr><tr><td>GPT-6 Luna<\/td><td class=\"num\">$123<\/td><td class=\"num\">$62.0<\/td><td class=\"num hide-m\">50%<\/td><\/tr><tr><td>Gemini 3.1 Pro<\/td><td class=\"num\">$2,472<\/td><td class=\"num\">$1,248<\/td><td class=\"num hide-m\">50%<\/td><\/tr><tr><td>Gemini 2.5 Pro<\/td><td class=\"num\">$1,560<\/td><td class=\"num\">$790<\/td><td class=\"num hide-m\">49%<\/td><\/tr><\/tbody><\/table><\/div>\n\n<p>The trade-off is real and worth stating plainly: chunking removes the model&#8217;s ability to reason across the whole document at once. It is the right move for extraction, summarisation and per-section analysis. It is the wrong move when the task depends on relationships between distant parts of the input. If you need both, run the cheap pass first and escalate only the cases that need the full context.<\/p>\n\n<h2>Prompt caching: break-even on the second request<\/h2>\n\n<p>Caching has an entry cost. Writing a prefix to the cache is billed at 1.25\u00d7 the normal input rate on both OpenAI and Anthropic, so caching is only worth it if you read the prefix back often enough to recover that premium.<\/p>\n\n<p>The arithmetic is simple. If the write costs <code>W<\/code> and each read costs <code>R<\/code> (both relative to uncached input), caching beats not caching after <code>W \/ (1 \u2212 R)<\/code> reads:<\/p>\n\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"hide-m\">Provider<\/th><th class=\"num\">Input \/1M<\/th><th class=\"num\">Cached \/1M<\/th><th class=\"num hide-m\">Reads are cheaper by<\/th><th class=\"num\">Break-even<\/th><\/tr><\/thead><tbody><tr><td>GPT-6 Astra<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$10.00<\/td><td class=\"num\">$1.000<\/td><td class=\"num hide-m\">10.0\u00d7<\/td><td class=\"num\">1.39<\/td><\/tr><tr><td>GPT-6 Sol<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$2.00<\/td><td class=\"num\">$0.200<\/td><td class=\"num hide-m\">10.0\u00d7<\/td><td class=\"num\">1.39<\/td><\/tr><tr><td>GPT-6 Luna<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$0.10<\/td><td class=\"num\">$0.010<\/td><td class=\"num hide-m\">10.0\u00d7<\/td><td class=\"num\">1.39<\/td><\/tr><tr><td>GPT-5.6 Sol<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$4.00<\/td><td class=\"num\">$0.400<\/td><td class=\"num hide-m\">10.0\u00d7<\/td><td class=\"num\">1.39<\/td><\/tr><tr><td>GPT-5.6 Cyber<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$12.50<\/td><td class=\"num\">$1.250<\/td><td class=\"num hide-m\">10.0\u00d7<\/td><td class=\"num\">1.39<\/td><\/tr><tr><td>GPT-5.3 Codex<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$1.75<\/td><td class=\"num\">$0.175<\/td><td class=\"num hide-m\">10.0\u00d7<\/td><td class=\"num\">1.39<\/td><\/tr><tr><td>Fable 5.1<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">$10.00<\/td><td class=\"num\">$0.250<\/td><td class=\"num hide-m\">40.0\u00d7<\/td><td class=\"num\">1.28<\/td><\/tr><tr><td>Opus 5.5<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">$4.00<\/td><td class=\"num\">$0.200<\/td><td class=\"num hide-m\">20.0\u00d7<\/td><td class=\"num\">1.32<\/td><\/tr><tr><td>Sonnet 5<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">$2.00<\/td><td class=\"num\">$0.200<\/td><td class=\"num hide-m\">10.0\u00d7<\/td><td class=\"num\">1.39<\/td><\/tr><tr><td>Haiku 4.5<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">$1.00<\/td><td class=\"num\">$0.100<\/td><td class=\"num hide-m\">10.0\u00d7<\/td><td class=\"num\">1.39<\/td><\/tr><\/tbody><\/table><\/div>\n\n<p>Every provider that publishes both rates lands between <strong>1.28 and 1.39 reads<\/strong>. In practice that means caching is profitable from the <strong>second<\/strong> request that shares a prefix \u2014 a very low bar for any workload with a fixed system prompt, a stable tool definition set, or a document that gets queried repeatedly.<\/p>\n\n<h2>The reasoning-token tax<\/h2>\n\n<p>Reasoning tokens are billed as output tokens. That matters because two of the four Anthropic models here have thinking <strong>always on<\/strong> \u2014 you cannot switch it off \u2014 while OpenAI&#8217;s GPT-6 family lets you set reasoning effort to <code>none<\/code>.<\/p>\n\n<p>For a high-volume, low-difficulty task such as classification or routing, that difference shows up directly in the bill. A model with mandatory reasoning will emit thinking tokens for a task that needed none, and every one of them is charged at the output rate. Before committing to a per-token price, check whether you can actually turn reasoning off on that model \u2014 the column below is the one most comparisons omit.<\/p>\n\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"hide-m\">Provider<\/th><th class=\"num\">Context window<\/th><th class=\"num\">Max output<\/th><th>Thinking control<\/th><\/tr><\/thead><tbody><tr><td>GPT-6 Astra<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">1,050,000<\/td><td class=\"num\">128,000<\/td><td>low..max<\/td><\/tr><tr><td>GPT-6 Sol<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">1,050,000<\/td><td class=\"num\">128,000<\/td><td>none..max<\/td><\/tr><tr><td>GPT-6 Luna<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">1,050,000<\/td><td class=\"num\">128,000<\/td><td>none..max<\/td><\/tr><tr><td>Fable 5.1<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">1,000,000<\/td><td class=\"num\">128,000<\/td><td>always on<\/td><\/tr><tr><td>Opus 5.5<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">1,000,000<\/td><td class=\"num\">128,000<\/td><td>always on<\/td><\/tr><tr><td>Sonnet 5<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">1,000,000<\/td><td class=\"num\">128,000<\/td><td>adaptive<\/td><\/tr><tr><td>Haiku 4.5<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">200,000<\/td><td class=\"num\">64,000<\/td><td>opt-in<\/td><\/tr><tr><td>Gemini 3.8 Flash<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">1,048,576<\/td><td class=\"num\">65,536<\/td><td>low\/med\/high<\/td><\/tr><tr><td>Gemini 3.1 Pro<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">1,048,576<\/td><td class=\"num\">65,536<\/td><td>yes<\/td><\/tr><tr><td>Gemini 2.5 Pro<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">1,048,576<\/td><td class=\"num\">65,536<\/td><td>yes<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (peak)<\/td><td class=\"hide-m\">DeepSeek<\/td><td class=\"num\">1,000,000<\/td><td class=\"num\">384,000<\/td><td>both modes<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (off-peak)<\/td><td class=\"hide-m\">DeepSeek<\/td><td class=\"num\">1,000,000<\/td><td class=\"num\">384,000<\/td><td>both modes<\/td><\/tr><tr><td>DeepSeek V4 Pro (peak)<\/td><td class=\"hide-m\">DeepSeek<\/td><td class=\"num\">1,000,000<\/td><td class=\"num\">384,000<\/td><td>both modes<\/td><\/tr><tr><td>DeepSeek V4 Pro (off-peak)<\/td><td class=\"hide-m\">DeepSeek<\/td><td class=\"num\">1,000,000<\/td><td class=\"num\">384,000<\/td><td>both modes<\/td><\/tr><\/tbody><\/table><\/div>\n\n<h2>Full price list<\/h2>\n\n<p>List rates per 1M tokens, standard tier, verified September 2026. Google&#8217;s Gemini 3 Flash family is on promotional pricing through 31 December 2026 and doubles on 1 January 2027, so treat those rows as temporary.<\/p>\n\n<div class=\"qapi-table\"><table><thead><tr><th>Model<\/th><th class=\"hide-m\">Provider<\/th><th class=\"num\">Input \/1M<\/th><th class=\"num\">Cached \/1M<\/th><th class=\"num\">Output \/1M<\/th><\/tr><\/thead><tbody><tr><td>Haiku 4.5<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">$1.00<\/td><td class=\"num\">$0.100<\/td><td class=\"num\">$5.00<\/td><\/tr><tr><td>Sonnet 5<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">$2.00<\/td><td class=\"num\">$0.200<\/td><td class=\"num\">$10.00<\/td><\/tr><tr><td>Opus 5.5<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">$4.00<\/td><td class=\"num\">$0.200<\/td><td class=\"num\">$20.00<\/td><\/tr><tr><td>Fable 5.1<\/td><td class=\"hide-m\">Anthropic<\/td><td class=\"num\">$10.00<\/td><td class=\"num\">$0.250<\/td><td class=\"num\">$50.00<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (off-peak)<\/td><td class=\"hide-m\">DeepSeek<\/td><td class=\"num\">$0.15<\/td><td class=\"num\">$0.003<\/td><td class=\"num\">$0.60<\/td><\/tr><tr><td>DeepSeek V4.1 Flash (peak)<\/td><td class=\"hide-m\">DeepSeek<\/td><td class=\"num\">$0.30<\/td><td class=\"num\">$0.006<\/td><td class=\"num\">$1.20<\/td><\/tr><tr><td>DeepSeek V4 Pro (off-peak)<\/td><td class=\"hide-m\">DeepSeek<\/td><td class=\"num\">$0.66<\/td><td class=\"num\">$0.022<\/td><td class=\"num\">$1.98<\/td><\/tr><tr><td>DeepSeek V4 Pro (peak)<\/td><td class=\"hide-m\">DeepSeek<\/td><td class=\"num\">$1.32<\/td><td class=\"num\">$0.044<\/td><td class=\"num\">$3.96<\/td><\/tr><tr><td>Gemini 2.5 Flash-Lite<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">$0.10<\/td><td class=\"num\">$0.010<\/td><td class=\"num\">$0.40<\/td><\/tr><tr><td>Gemini 3.1 Flash-Lite<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">$0.25<\/td><td class=\"num\">$0.025<\/td><td class=\"num\">$1.50<\/td><\/tr><tr><td>Gemini 3.5 Flash-Lite<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">$0.30<\/td><td class=\"num\">$0.030<\/td><td class=\"num\">$2.50<\/td><\/tr><tr><td>Gemini 2.5 Flash<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">$0.30<\/td><td class=\"num\">$0.030<\/td><td class=\"num\">$2.50<\/td><\/tr><tr><td>Gemini 3.8 Flash<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">$0.75<\/td><td class=\"num\">$0.075<\/td><td class=\"num\">$3.75<\/td><\/tr><tr><td>Gemini 2.5 Pro<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">$1.25<\/td><td class=\"num\">$0.125<\/td><td class=\"num\">$10.00<\/td><\/tr><tr><td>Gemini 3.5 Flash<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">$1.50<\/td><td class=\"num\">$0.150<\/td><td class=\"num\">$9.00<\/td><\/tr><tr><td>Gemini 3.1 Pro<\/td><td class=\"hide-m\">Google<\/td><td class=\"num\">$2.00<\/td><td class=\"num\">$0.200<\/td><td class=\"num\">$12.00<\/td><\/tr><tr><td>GPT-6 Luna<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$0.10<\/td><td class=\"num\">$0.010<\/td><td class=\"num\">$0.50<\/td><\/tr><tr><td>GPT-5.3 Codex<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$1.75<\/td><td class=\"num\">$0.175<\/td><td class=\"num\">$14.00<\/td><\/tr><tr><td>GPT-6 Sol<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$2.00<\/td><td class=\"num\">$0.200<\/td><td class=\"num\">$10.00<\/td><\/tr><tr><td>GPT-5.6 Sol<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$4.00<\/td><td class=\"num\">$0.400<\/td><td class=\"num\">$20.00<\/td><\/tr><tr><td>GPT-6 Astra<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$10.00<\/td><td class=\"num\">$1.000<\/td><td class=\"num\">$50.00<\/td><\/tr><tr><td>GPT-5.6 Cyber<\/td><td class=\"hide-m\">OpenAI<\/td><td class=\"num\">$12.50<\/td><td class=\"num\">$1.250<\/td><td class=\"num\">$75.00<\/td><\/tr><\/tbody><\/table><\/div>\n\n<div class=\"qapi-dl\">\n<b>Download the dataset<\/b>\n<span>All 22 models with input, cached-input and output rates, cache read ratios, cache break-even points, long-context thresholds and multipliers, context windows and max output \u2014 as a single CSV.<\/span>\n<a href=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/ai-api-cost-index-2026.csv\" rel=\"noopener\">ai-api-cost-index-2026.csv \u2193<\/a>\n<\/div>\n\n<h2>Methodology<\/h2>\n\n<p>So that the numbers can be checked rather than trusted:<\/p>\n\n<ol>\n<li><strong>Rates.<\/strong> Taken from each provider&#8217;s official pricing page in September 2026: <a href=\"https:\/\/platform.openai.com\/docs\/pricing\" target=\"_blank\" rel=\"noopener\">platform.openai.com\/docs\/pricing<\/a>, <a href=\"https:\/\/anthropic.com\/pricing\" target=\"_blank\" rel=\"noopener\">anthropic.com\/pricing<\/a>, <a href=\"https:\/\/ai.google.dev\/gemini-api\/docs\/pricing\" target=\"_blank\" rel=\"noopener\">ai.google.dev\/gemini-api\/docs\/pricing<\/a>, <a href=\"https:\/\/api-docs.deepseek.com\/quick_start\/pricing\" target=\"_blank\" rel=\"noopener\">api-docs.deepseek.com\/quick_start\/pricing<\/a>. Standard service tier only. Batch, flex, priority and fast tiers are excluded except where the tables say otherwise.<\/li>\n<li><strong>Cost formula.<\/strong> <code>cost = input_M \u00d7 (1 \u2212 hit) \u00d7 input_price + input_M \u00d7 hit \u00d7 cached_price + output_M \u00d7 output_price<\/code>. Cache rates apply to input tokens only.<\/li>\n<li><strong>Long-context handling.<\/strong> Where a provider documents a threshold, the multiplier is applied to the whole request \u2014 input and output both \u2014 not just to the tokens above the threshold. That matches the vendors&#8217; published wording.<\/li>\n<li><strong>Cache break-even.<\/strong> <code>reads = cache_write_multiplier \/ (1 \u2212 cache_read_ratio)<\/code>, using the 1.25\u00d7 write multiplier that OpenAI and Anthropic both publish. Providers that do not publish a write rate are omitted from that table rather than estimated.<\/li>\n<li><strong>Workload volumes.<\/strong> The six workloads are illustrative, not benchmarks. Their token counts are stated so you can substitute your own. Nothing here is a measurement of speed, quality or reliability \u2014 this index is about price only.<\/li>\n<\/ol>\n\n<p>Two caveats worth repeating. List prices change without notice, and Google&#8217;s promotional Flash pricing expires on 1 January 2027. And a cheap model that fails the task is more expensive than an expensive model that succeeds \u2014 the point of this index is to show where the price differences are large enough to be worth testing, not to declare a winner.<\/p>\n\n<h2>Frequently asked questions<\/h2>\n\n<h3>Why do the same token counts cost different amounts on different models?<\/h3>\n<p>Because providers set their own rates, and the input and output rates move independently. Output is where the divergence is widest: at the time of writing, output rates across these 22 models span from $0.40 to $75 per million tokens \u2014 a 187\u00d7 range \u2014 while input rates span 125\u00d7.<\/p>\n\n<h3>Should I just pick the cheapest model?<\/h3>\n<p>No. The right conclusion is that the spread is large enough to justify routing different traffic to different models. Classification and extraction rarely need a frontier model; multi-step reasoning and code generation often do. The saving comes from matching, not from minimising.<\/p>\n\n<h3>How much does caching actually save?<\/h3>\n<p>On the workloads above, a 70% cache hit rate cuts 40\u201360% off the input portion of the bill. Since input is usually the larger token count but the cheaper rate, the total saving typically lands between 20% and 40%. The tables show the exact figure for each workload.<\/p>\n\n<h3>Is chunking a long document always safe?<\/h3>\n<p>No. It is safe when each chunk can be understood independently \u2014 extraction, summarisation, per-section classification. It breaks tasks that require cross-document reasoning, comparison, or a global view. Test on your own evaluation set before shipping it.<\/p>\n\n<h3>How often should I re-run this analysis?<\/h3>\n<p>Quarterly at minimum, and immediately after any provider announces a price change. Two of the models here carry promotional pricing with a known expiry date, and providers routinely retire models on short notice.<\/p>\n\n<div class=\"qapi-cite\">\n<b>Cite this data<\/b>\n<p>These figures are free to reuse with attribution. Suggested citation:<\/p>\n<p class=\"qapi-cite-q\">Qora API. (2026). The 2026 AI API Cost Index: What 22 Models Really Cost Per Workload. https:\/\/qoraapi.com\/blog\/ai-api-cost-index-2026\/<\/p>\n<p class=\"qapi-cite-m\">Prices verified against provider documentation in September 2026 \u00b7 dataset: <a href=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/ai-api-cost-index-2026.csv\">ai-api-cost-index-2026.csv<\/a> (22 models \u00d7 14 fields) \u00b7 licensed CC BY 4.0<\/p>\n<\/div>\n\n<div class=\"qapi-cta\">\n<strong>Acting on this without rewriting your code<\/strong>\n<p>Every technique above \u2014 routing by task, capping output, chasing the cheapest capable model \u2014 requires you to change which model a request goes to. If your app talks to four providers through four SDKs, that change means four code paths and four retry strategies, which is why most teams never get to it.<\/p>\n<p>QoraAPI exposes all of these models through one OpenAI-compatible endpoint, so routing, failover and budget limits become configuration rather than application code.<\/p>\n<p><a href=\"https:\/\/qoraapi.com\">Explore QoraAPI \u2192<\/a><\/p>\n<\/div>\n<script type=\"application\/ld+json\">{\"@context\":\"https:\/\/schema.org\",\"@type\":\"Dataset\",\"@id\":\"https:\/\/qoraapi.com\/blog\/ai-api-cost-index-2026\/#dataset\",\"name\":\"The 2026 AI API Cost Index: What 22 Models Really Cost Per Workload\",\"alternateName\":\"AI API Cost Index 2026\",\"description\":\"Cost per workload for 22 frontier AI models across six fixed production workloads (classification, customer-support reply, RAG answer, coding agent turn, bulk summarisation, and long-document analysis past the context cliff). Derived by applying each provider's published token rates - input, cached input, output, and long-context multipliers - to identical token profiles. Includes cache break-even read counts and per-provider long-context thresholds. Prices verified against provider documentation in September 2026.\",\"url\":\"https:\/\/qoraapi.com\/blog\/ai-api-cost-index-2026\/\",\"creator\":{\"@type\":\"Organization\",\"name\":\"Qora API\",\"url\":\"https:\/\/qoraapi.com\/\"},\"publisher\":{\"@type\":\"Organization\",\"name\":\"Qora API\",\"url\":\"https:\/\/qoraapi.com\/\"},\"license\":\"https:\/\/creativecommons.org\/licenses\/by\/4.0\/\",\"isAccessibleForFree\":true,\"datePublished\":\"2026-09-23\",\"dateModified\":\"2026-09-23\",\"version\":\"2026-09\",\"temporalCoverage\":\"2026-09\/2026-09\",\"keywords\":[\"AI API pricing\",\"LLM cost\",\"token pricing\",\"prompt caching\",\"long context pricing\",\"AI gateway\",\"LLM routing\",\"cost per workload\"],\"measurementTechnique\":\"Token-level pricing arithmetic. Each workload is defined as a fixed input token count, output token count and cache-hit rate; total cost is computed as input*(1-hit)*input_rate + input*hit*cached_rate + output*output_rate. Where a provider applies a long-context multiplier above a threshold, the multiplier is applied to the entire request.\",\"variableMeasured\":[{\"@type\":\"PropertyValue\",\"name\":\"Input price per million tokens\",\"unitText\":\"USD\"},{\"@type\":\"PropertyValue\",\"name\":\"Cached input price per million tokens\",\"unitText\":\"USD\"},{\"@type\":\"PropertyValue\",\"name\":\"Output price per million tokens\",\"unitText\":\"USD\"},{\"@type\":\"PropertyValue\",\"name\":\"Cache write multiplier\",\"unitText\":\"ratio of input price\"},{\"@type\":\"PropertyValue\",\"name\":\"Cache break-even read count\",\"unitText\":\"reads\"},{\"@type\":\"PropertyValue\",\"name\":\"Long-context threshold\",\"unitText\":\"tokens\"},{\"@type\":\"PropertyValue\",\"name\":\"Context window\",\"unitText\":\"tokens\"},{\"@type\":\"PropertyValue\",\"name\":\"Maximum output tokens\",\"unitText\":\"tokens\"}],\"distribution\":[{\"@type\":\"DataDownload\",\"name\":\"AI API cost index dataset (CSV)\",\"description\":\"22 models x 14 fields: token rates, cache economics, long-context thresholds and model limits.\",\"encodingFormat\":\"text\/csv\",\"contentUrl\":\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/ai-api-cost-index-2026.csv\",\"license\":\"https:\/\/creativecommons.org\/licenses\/by\/4.0\/\"}],\"citation\":\"Qora API. (2026). The 2026 AI API Cost Index: What 22 Models Really Cost Per Workload. https:\/\/qoraapi.com\/blog\/ai-api-cost-index-2026\/\"}<\/script>\n<script type=\"application\/ld+json\">{\"@context\":\"https:\/\/schema.org\",\"@type\":\"FAQPage\",\"@id\":\"https:\/\/qoraapi.com\/blog\/ai-api-cost-index-2026\/#faq\",\"mainEntity\":[{\"@type\":\"Question\",\"name\":\"Why do the same token counts cost different amounts on different models?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"Because providers set their own rates, and the input and output rates move independently. Output is where the divergence is widest: at the time of writing, output rates across these 22 models span from $0.40 to $75 per million tokens - a 187x range - while input rates span 125x.\"}},{\"@type\":\"Question\",\"name\":\"Should I just pick the cheapest model?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"No. The right conclusion is that the spread is large enough to justify routing different traffic to different models. Classification and extraction rarely need a frontier model; multi-step reasoning and code generation often do. The saving comes from matching, not from minimising.\"}},{\"@type\":\"Question\",\"name\":\"How much does caching actually save?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"On the workloads above, a 70% cache hit rate cuts 40-60% off the input portion of the bill. Since input is usually the larger token count but the cheaper rate, the total saving typically lands between 20% and 40%. The tables show the exact figure for each workload.\"}},{\"@type\":\"Question\",\"name\":\"Is chunking a long document always safe?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"No. It is safe when each chunk can be understood independently - extraction, summarisation, per-section classification. It breaks tasks that require cross-document reasoning, comparison, or a global view. Test on your own evaluation set before shipping it.\"}},{\"@type\":\"Question\",\"name\":\"How often should I re-run this analysis?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"Quarterly at minimum, and immediately after any provider announces a price change. Two of the models here carry promotional pricing with a known expiry date, and providers routinely retire models on short notice.\"}}]}<\/script>\n<p>Related guides: <a href=\"https:\/\/qoraapi.com\/blog\/reduce-ai-api-costs\/\">How to Reduce AI API Costs<\/a> \u00b7 <a href=\"https:\/\/qoraapi.com\/blog\/prompt-caching-guide\/\">Prompt Caching Explained<\/a> \u00b7 <a href=\"https:\/\/qoraapi.com\/blog\/batch-ai-api-processing\/\">Batch AI APIs<\/a> \u00b7 <a href=\"https:\/\/qoraapi.com\/blog\/ai-api-pricing-explained\/\">How AI API Pricing Works<\/a> \u00b7 <a href=\"https:\/\/qoraapi.com\/blog\/ai-api-cost-calculator\/\">AI API Cost Calculator<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<div class=\"wp-block-group alignfull has-background has-text-color has-global-padding is-layout-constrained wp-container-core-group-is-layout-cee23037 wp-block-group-is-layout-constrained\" style=\"background-color:#39434F;color:#FFFFFF;padding-top:var(--wp--preset--spacing--60);padding-right:var(--wp--preset--spacing--40);padding-bottom:var(--wp--preset--spacing--60);padding-left:var(--wp--preset--spacing--40)\">\n\t\n\n<figure class=\"wp-block-image aligncenter size-full is-resized has-background qora-cta-logo\" style=\"border-radius:12px;background-color:#FFFFFF;padding-top:var(--wp--preset--spacing--30);padding-right:var(--wp--preset--spacing--40);padding-bottom:var(--wp--preset--spacing--30);padding-left:var(--wp--preset--spacing--40)\"><a href=\"https:\/\/qoraapi.com\/\" target=\"_blank\" rel=\"noopener\"><img loading=\"lazy\" decoding=\"async\" width=\"900\" height=\"292\" src=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/qora-api-logo.png\" alt=\"Qora API \u2014 AI API Gateway for Developers\" class=\"wp-image-149\" style=\"width:260px\" srcset=\"https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/qora-api-logo.png 900w, https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/qora-api-logo-300x97.png 300w, https:\/\/qoraapi.com\/blog\/wp-content\/uploads\/2026\/09\/qora-api-logo-768x249.png 768w\" sizes=\"auto, (max-width: 900px) 100vw, 900px\" \/><\/a><\/figure>\n\n\t\n\n<h2 class=\"wp-block-heading has-text-align-center has-text-color has-x-large-font-size\" style=\"color:#FFFFFF;font-weight:700;line-height:1.25\">Build AI features with one clear API<\/h2>\n\n\t\n\n<p class=\"has-text-align-center has-text-color wp-block-paragraph\" style=\"color:#C7D1DA;line-height:1.6\">Qora API gives you a single, focused gateway to connect your apps, scripts and automations to AI. Start with one request.<\/p>\n\n\t\n\n<div class=\"wp-block-buttons is-content-justification-center is-layout-flex wp-container-core-buttons-is-layout-35f06ea7 wp-block-buttons-is-layout-flex\">\n\t\t\n\n<div class=\"wp-block-button\"><a class=\"wp-block-button__link has-text-color has-background wp-element-button\" href=\"https:\/\/qoraapi.com\/\" target=\"_blank\" rel=\"noopener\" style=\"border-radius:999px;color:#FFFFFF;background-color:#00B5B3;font-weight:600\">Start building &rarr;<\/a><\/div>\n\n\t\t\n\n<div class=\"wp-block-button is-style-outline is-style-outline--1\"><a class=\"wp-block-button__link has-text-color has-border-color wp-element-button\" href=\"https:\/\/qoraapi.com\/pricing\" target=\"_blank\" rel=\"noopener\" style=\"border-radius:999px;border-color:#FFFFFF;border-width:1px;color:#FFFFFF;font-weight:600\">See the model square<\/a><\/div>\n\n\t<\/div>\n\n\t\n\n<p class=\"has-text-align-center has-text-color has-small-font-size wp-block-paragraph\" style=\"color:#8FA0AE;line-height:1.5\">qoraapi.com &middot; AI API gateway for developers<\/p>\n\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>Almost every AI API pricing page answers the wrong question. It tells you the cost of a million tokens, when what you actually need to know is the cost of your workload \u2014 one support reply, one code review, one document summary \u2014 multiplied by how many of them you run. This index closes that [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"footnotes":""},"class_list":["post-343","page","type-page","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/qoraapi.com\/blog\/wp-json\/wp\/v2\/pages\/343","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/qoraapi.com\/blog\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/qoraapi.com\/blog\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/qoraapi.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/qoraapi.com\/blog\/wp-json\/wp\/v2\/comments?post=343"}],"version-history":[{"count":11,"href":"https:\/\/qoraapi.com\/blog\/wp-json\/wp\/v2\/pages\/343\/revisions"}],"predecessor-version":[{"id":393,"href":"https:\/\/qoraapi.com\/blog\/wp-json\/wp\/v2\/pages\/343\/revisions\/393"}],"wp:attachment":[{"href":"https:\/\/qoraapi.com\/blog\/wp-json\/wp\/v2\/media?parent=343"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}