--- title: "LLM output prices vary 106× between vendors. Your SMB pays the ceiling. - OkamiOps Insights" description: "In September 2026, output from the priciest model in the public catalog (Claude Fable 5.1, $50 per 1M tokens) costs 106 times the cheapest one that still" url: https://okamiops.com/insights/llm-cost-spread-2026/ lang: en alternates: en: https://okamiops.com/insights/llm-cost-spread-2026/ pt-BR: https://okamiops.com/pt/insights/llm-cost-spread-2026/ de: https://okamiops.com/de/insights/llm-cost-spread-2026/ x-default: https://okamiops.com/insights/llm-cost-spread-2026/ lastmod: 2026-09-10 --- # LLM output prices vary 106× between vendors. Your SMB pays the ceiling. GATEWAY · JUL · 2026 · 6 min In September 2026, output from the priciest model in the public catalog (Claude Fable 5.1, $50 per 1M tokens) costs 106 times the cheapest one that still handles production work (Qwen3.8-Flash, $0.47). No SMB needs the top of that list to triage a support ticket or pull a field off an invoice. Run everything on one vendor and you pay the top price on 100% of requests. - **Category**: Gateway - **Date**: JUL · 2026 - **Read time**: 6 min - **Sources**: 8 ## What the numbers show We pulled the official price tables from seven vendors in September 2026 and lined everything up in one unit: dollars per 1 million output tokens. At the top sits Claude Fable 5.1 at $50. OpenAI's GPT-6 Astra ties at the same figure. Working down: Claude Opus 5 at $25, GPT-5.6 Sol at $20, Gemini 3.1 Pro at $12, Mistral Medium 3.5 at $7.50, GLM-5.3 at $4.40, Gemini 3.8 Flash at $3.75, Mistral Large 3 at $1.50, GPT-5.6 Luna at $1.20 and Qwen3.8-Flash at $0.47. The bottom end is not made of toy models. Qwen3.8-Flash scores 42 on the Artificial Analysis Intelligence Index; the top of the table scores 53. Eleven points of capability difference, 106 times the price difference. The curve is not smooth either. There are six steps between $50 and $4.40. There are four more between $4.40 and $0.47. The question that matters is not which model is best. It is which step this task needs. **Output per 1M tokens (USD, Sep 2026)** (vs Claude Fable 5.1 (priciest)) | Item | Value | | --- | --- | | anthropic · fable 5.1 | $50.00 | | anthropic · opus 5 | $25.00 | | openai · gpt-5.6 sol | $20.00 | | google · gemini 3.1 pro | $12.00 | | mistral · medium 3.5 | $7.50 | | zhipu · glm-5.3 | $4.40 | | google · gemini 3.8 flash | $3.75 | | mistral · large 3 | $1.50 | | openai · gpt-5.6 luna | $1.20 | | alibaba · qwen3.8-flash | $0.47 −99% | ## Input, output and cache are three different bills Almost every conversation about LLM cost starts and stops at the output table. That is the wrong number to start from. On a real SMB workload — RAG, extraction, triage — input token volume typically runs 5 to 10 times output volume. Input pricing weighs more than it looks. Caching changes the bill again. At Anthropic, reading a prefix already in cache costs 0.1× normal input, and 0.025× on Fable 5.1. Writing to cache costs 1.25× on the 5-minute window or 2× on the one-hour window. If your system prompt, schema and examples are stable, you pay 10% or 2.5% of input on every subsequent call. Google and OpenAI offer cache discounts in the same order of magnitude. Batch is the third discount and the most ignored: 50% at OpenAI, Anthropic and Google for anything that can wait a few hours. Monthly reporting, backfills, retroactive classification, CRM enrichment — none of that needs an answer in two seconds. And some prices move with the clock. DeepSeek charges double between 01:00 and 04:00 and between 06:00 and 10:00 UTC, Monday to Friday. That is 08:00–12:00 in Berlin, the peak of the German working day. It is 03:00–07:00 in Brasília, the middle of the night. The same model yields different savings depending on the time zone your team works in. **The discounts the output price list hides** (official price lists · Sep 2026) | | Multiplier | When it applies | | --- | --- | --- | | Cached prefix read (Anthropic) | 0.1× · 0.025× on Fable 5.1 | System prompt, schema and stable examples | | Cache write | 1,25× (5 min) · 2× (1 h) | Once per window; paid on the first call | | Batch (OpenAI, Anthropic, Google) | −50% | Reports, reprocessing, retroactive classification | | Off-peak pricing (DeepSeek) | 2× at UTC peak | Early morning in Brasília; business hours in Berlin pay double | | Input vs output volume at SMBs | 5–10× more input tokens | RAG, extraction and triage: input outweighs output | ## A real bill for a 40-person SMB Explicit assumptions, so you can redo the maths with your own numbers. 120,000 requests a month across three uses: ticket triage, document extraction and questions against an internal knowledge base. 70% light (3,000 input tokens, 400 output), 22% medium (6,000 and 900), 8% heavy (12,000 and 2,500). That gives 525.6 million input tokens and 81.4 million output tokens per month. Scenario A, everything at the top. Running all 120,000 requests on Fable 5.1 ($10 input, $50 output): $5,256 of input plus $4,068 of output. $9,324 a month. Scenario B, routed by task. Light on Qwen3.8-Flash: $56. Medium on Sonnet 5: $554. Heavy stays on Fable 5.1: $2,352. Total $2,963 — 68% less, with the frontier model still answering everything hard. Scenario C, routed plus caching. Assuming 60% of input tokens on the medium and heavy paths are a stable prefix served from cache, the bill drops to $2,118. That is 77% below scenario A, or $86,000 a year that stays in the company. **The same 120k requests/month, three ways to pay** - **US$ 9.324** — Scenario A · everything on Fable 5.1 - **US$ 2.963 −68%** — Scenario B · routed by task, frontier only for the hard cases - **US$ 2.118 −77%** — Scenario C · routed + cached prefix · $86k/year kept Nothing in that calculation depends on renegotiating a contract, changing stack or accepting a worse answer on the hard case. It depends on deciding task by task instead of vendor by company. ## Where the cheap model fails In three places, consistently. First, multi-step reasoning with state. Cheap models lose the thread on long tool-call chains, and the error usually surfaces at step seven, not step one. That is the most expensive place to find it. Second, verbosity. Artificial Analysis measured DeepSeek V4 Flash generating 240 million output tokens across its test suite, against a median of 130 million. A model three times cheaper per token that talks 1.8 times more is, in practice, only 1.7 times cheaper. List price is not cost per task. Third, language and domain. Legal German and Portuguese loaded with tax jargon do not behave like generic English. And the degradation is silent: the answer stays fluent while it starts being wrong. Only a scored case set surfaces that. Add a calendar risk. Gemini 3.8 Flash's promotional price ($0.75 input, $3.75 output) runs until 31 December 2026 and doubles afterwards. An annual budget built on that number breaks in January. ## The conditions that make the saving real Routing without four things is a gamble, not an optimisation. Automatic fallback. If the cheap provider goes down, the request moves to the next in line without human intervention. Without it, the first outage costs more than months of savings. An evaluation gate. 50 to 500 representative cases that run before any routing change. If quality drops below the threshold, the change does not ship. Per-task observability. Cost, latency and accuracy by route, not one monthly total. Without it you cannot tell which route saved money and which one only made the answer worse. Audit. Which model answered which request, with which prompt, on which date, under which data policy. It is what compliance will ask for and what the EU AI Act expects from anyone operating an AI system — and it is also what ties your data policy to the vendor's when you run on just one. ## What to do in the next two weeks One: export last month's invoice and break the spend down by task type, not by model. Most teams discover that around 70% of volume is light work paying frontier prices. Two: take the highest-volume, simplest task. Build 100 cases with expected answers. Run them on the current model and on two cheap candidates. Compare measured accuracy, not impressions. Three: settle the data policy before the price. US-only inference costs 1.1× at Anthropic. A European vendor like Mistral delivers output at $1.50 on Large 3. Data residency has a price — and it is smaller than the gap between two steps on the table. Four: put the gateway in place before swapping any model. Without a layer that routes, falls back and logs, every swap becomes a migration project. With it, it becomes one line of configuration. **Two-week plan** 1. **Split the bill by task type** — Export last month and total it by triage, extraction, Q&A. Usually ~70% of the volume is light work paying frontier prices. 2. **Measure 100 cases with ground truth** — Take the highest-volume, simplest task. Run it on the current model and two cheap candidates. Compare measured accuracy, not impressions. 3. **Decide the data policy before the price** — US-restricted inference costs 1.1× at Anthropic; Mistral delivers Large 3 output at $1.50. Residency has a price, smaller than one step of the table. 4. **Gateway before switching any model** — Without a layer that routes, falls back and logs, every switch becomes a migration project. With it, it is one line of configuration. > **What to do tomorrow**: Break last month's invoice down by task type, not by model: around 70% of volume is usually light work paying frontier prices. Then put the gateway in front — with it, swapping models is configuration, not migration. ## Sources for this analysis - **01**: [OpenAI · API pricing](https://developers.openai.com/api/docs/pricing) - **02**: [Anthropic · Claude pricing](https://claude.com/pricing) - **03**: [Google · Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing) - **04**: [DeepSeek · Models & pricing](https://api-docs.deepseek.com/quick_start/pricing) - **05**: [Alibaba Cloud · Qwen3.8-Flash price-performance](https://www.alibabacloud.com/blog/alibaba-releases-qwen3-8-flash-with-innovative-model-architecture-delivering-optimal-price-performance_603503) - **06**: [Z.ai · GLM model pricing](https://docs.z.ai/guides/overview/pricing) - **07**: [Mistral AI · Inference pricing](https://docs.mistral.ai/inference/pricing/) - **08**: [Artificial Analysis · LLM leaderboard](https://artificialanalysis.ai/leaderboards/models) ## Keep reading - **CONSULTORIA · AGO · 2026**: [What AI really costs in 2026: three bands, with the arithmetic shown](https://okamiops.com/insights/quanto-custa-implementar-ia-2026/) (6 min) - **APPSEC · ABR · 2026**: [OWASP LLM Top 10 (2025): five risks already in the incident record — and how to cover each](https://okamiops.com/insights/owasp-llm-top-10-2025/) (8 min) - **COMPLIANCE · MAR · 2026**: [The EU AI Act after the Digital Omnibus: the real calendar and a 90-day plan](https://okamiops.com/insights/eu-ai-act-timeline/) (7 min) ## No loose opinions. Every number has a source, every article ends with what to do. We write what we learn solving the problem for a client. If your situation looks like one of these, the conversation starts here.