Disclosure: This article contains no affiliate links. Every product below is named in plain text, and every price was taken from the provider’s published pricing, checked in October 2026. Where a secondary source reported a provider’s official page, that is stated in the Sources section.
Every AI tool you pay for monthly is, underneath, renting tokens from one of a handful of model providers. If you build automations, chatbots, or agents for your business, the per-token price board is the wholesale price list behind almost everything you sell or run — and in 2026 it changed shape. OpenAI’s GPT-6 family and Anthropic’s Claude 5.5 generation both landed with headline cuts, Google is running introductory Gemini pricing that expires on a published date, and DeepSeek now charges different prices depending on the time of day.
- The 2026 price board
- The four discount levers (and one premium)
- Batch and flexible processing: a flat 50% off
- Prompt caching: up to 98% off repeated input
- Speed tiers: paying more, on purpose
- Time-of-day pricing: DeepSeek’s peak windows
- Context-length pricing: the cliff edges
- Workload 1: a customer support chatbot
- Workload 2: a content production pipeline
- Workload 3: an autonomous AI agent
- Which model for which job: the decision framework
- FAQ
- Methodology
- AWL verdict
- Sources
So this article does the arithmetic the price pages don’t. We take the current published per-1M-token rates from OpenAI, Anthropic, Google, and DeepSeek, then run them through three defined business workloads — a support chatbot, a content pipeline, and an AI agent — with every token assumption printed and every multiplication shown. The result is less tidy than a price list, and much more useful.
The short answer: At standard rates, the workhorse tier costs $2 per million input tokens and $10 per million output at both OpenAI (GPT-6.1 Sol) and Anthropic (Claude Sonnet 5.5), while the budget tier has collapsed to $0.10 / $0.50 (GPT-6 Luna, Claude Haiku 5.5). Costed out, a 10,000-conversation support bot runs $6–$120 a month depending on tier; a 100-article content pipeline runs $0.70–$28, halved again with batch pricing; and a 5,000-run AI agent costs $31.50 on the budget tier versus $540 on the workhorse tier — with prompt caching, not model choice, doing most of the saving.
The 2026 price board
All prices below are US dollars per 1 million tokens (MTok), standard API rates, as published in October 2026. “Cached input” is what you pay when part of your prompt is served from the provider’s prompt cache instead of being processed fresh — the single most important column on this page, as the workloads will show.
| Model | Input | Cached input | Output | Watch out for |
|---|---|---|---|---|
| OpenAI GPT-6 Astra | $10.00 | $1.00 | $50.00 | Flagship tier; rarely the right default |
| OpenAI GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | Requests over 272K input tokens pay higher rates |
| OpenAI GPT-6 Luna | $0.10 | $0.01 | $0.50 | Budget tier; weaker on hard reasoning |
| Claude Fable 5.1 | $10.00 | $0.25 | $50.00 | Anthropic’s most capable tier |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | Fast Mode doubles the price |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $10.00 | The default workhorse tier |
| Claude Haiku 5.5 | $0.10 | $0.01 | $0.50 | 5Ă— rate on any prompt over 100K tokens |
| Gemini 3.1 Pro Preview | $2.00 | $0.20 | $12.00 | Prompts over 200K tokens pay $4 / $18 |
| Gemini 3.8 Flash | $0.75 | $0.075 | $3.75 | Introductory rate — doubles on 1 Jan 2027 |
| Gemini 3.1 Flash-Lite | $0.25 | $0.025 | $1.50 | Google’s mid-budget tier |
| DeepSeek V4.1 Flash | $0.15 | $0.003 | $0.60 | Off-peak rate; peak hours cost exactly 2Ă— |
| DeepSeek V4 Pro | $0.66 | $0.022 | $1.98 | Off-peak rate; peak hours cost exactly 2Ă— |
Three things stand out from the board itself. First, OpenAI and Anthropic have converged: their workhorse models sit at identical $2 / $10 list prices, and their budget models at identical $0.10 / $0.50. Second, the flagship tier ($10 / $50 at both OpenAI’s Astra and Anthropic’s Fable 5.1) now costs 50–100× the budget tier per token — a gap that used to be far smaller. Third, Google and DeepSeek price time as well as tokens: Gemini 3.8 Flash’s rate is introductory and doubles on a published date, and DeepSeek charges double during weekday peak windows (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday).
One caution before the arithmetic: per-token prices are not per-task prices. Anthropic’s current tokenizer (Claude 4.7 generation and later) produces roughly 30% more tokens for the same text than its predecessor, according to Anthropic’s pricing documentation as reported by Finout. A model can cut its per-token price by a third and still bill you only slightly less per document. Always benchmark cost per finished task on your own text.
The four discount levers (and one premium)
Batch and flexible processing: a flat 50% off
All three Western providers will halve your bill if your work can wait. OpenAI’s Batch and Flex tiers price GPT-6.1 Sol at $1.00 / $5.00 instead of $2.00 / $10.00. Anthropic’s Message Batches API is exactly 50% off every model, with results returned within 24 hours and no quality difference — and batch and caching discounts stack. Google’s Batch and Flex tiers are likewise half of Standard on every Gemini language model. Any workload that runs on a schedule rather than in a live conversation — nightly content generation, document processing, data enrichment — should never be paying standard rates.
Prompt caching: up to 98% off repeated input
Caching is the bigger lever. When consecutive requests share a long prefix — a system prompt, a policy document, an agent’s tool definitions and context — the provider stores it and charges a fraction for re-reads. Cache reads cost $0.10 per MTok on GPT-6.1 Sol and Claude Sonnet 5.5 (5% of the fresh-input rate), $0.01 on the budget models, $0.075 on Gemini 3.8 Flash, and as little as $0.003 on DeepSeek V4.1 Flash off-peak — 2% of its fresh rate. The mechanics differ: Anthropic charges a write premium (1.25× the input rate for a 5-minute cache, 2× for 1-hour) and then near-nothing per read; OpenAI lists a $2.50 per MTok cache-write price on GPT-6.1 Sol; Google applies implicit caching automatically and charges storage per million tokens per hour on explicit caches ($0.50 per MTok/hour on Gemini 3.8 Flash at current rates). DeepSeek caches repeated prefixes automatically with no write fee. For any workload that re-reads the same context, caching beats switching providers.
Speed tiers: paying more, on purpose
Each provider also sells speed. OpenAI’s Fast mode and Anthropic’s Fast Mode both double the standard price (GPT-6.1 Sol Fast is $4 / $20; Claude Opus 5.5 Fast Mode is $8 / $40). Google’s Priority tier costs 1.8× Standard. These are latency purchases, not quality purchases — the same model answers faster. For a solo founder, they are almost never the right default; reserve them for genuinely user-facing, latency-critical paths.
Time-of-day pricing: DeepSeek’s peak windows
DeepSeek is the only major provider pricing by the clock. Since August 2026, its rates double during peak windows — 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. About 79% of the week is off-peak, and for US-based traffic the peak windows fall mostly overnight. A scheduled job that runs on weekends or outside those windows pays half. No other provider on the board currently does this, but it is a pricing shape worth watching.
Context-length pricing: the cliff edges
Several models charge more once a single request’s prompt crosses a threshold — and critically, crossing it re-prices the entire request, not just the extra tokens.
| Model | Threshold | Below | Above | Effect |
|---|---|---|---|---|
| GPT-6.1 Sol | 272K input tokens | $2 / $10 | $4 / $15 | Input doubles, output +50% |
| Gemini 3.1 Pro Preview | 200K prompt tokens | $2 / $12 | $4 / $18 | Input doubles, output +50% |
| Claude Haiku 5.5 | 100K prompt tokens | $0.10 / $0.50 | $0.50 / $2.50 | Whole request at 5Ă— |
| Claude Sonnet / Opus 5.5 | None up to 1M | $2 / $10, $4 / $20 | Same | Flat rate to 1M tokens |
The arithmetic of a cliff, worked once so you can feel it: on Gemini 3.1 Pro Preview, a request with a 200,000-token prompt and a 5,000-token answer costs 0.2 × $2 + 0.005 × $12 = $0.46. Add one token — a 200,001-token prompt — and the whole request bills at long-context rates: 0.2 × $4 + 0.005 × $18 = $0.89. One token nearly doubled the bill. The same shape applies on GPT-6.1 Sol above 272K input tokens: a 300K-input, 10K-output request costs $1.35 at long-context rates versus $0.70 at the short-context rate. And on Claude Haiku 5.5, a 120K-token prompt bills input at $0.50 per MTok, not $0.10 — the cheap tier quietly stops being cheap. Practical rule: cap retrieval and context-packing below your model’s threshold unless the extra material demonstrably changes the answer.
Workload 1: a customer support chatbot
Assumptions (planning model — adjust to your real logs): 10,000 conversations per month. Each conversation uses 4,000 input tokens (system prompt, policy context, retrieved help articles, and multi-turn history) and 400 output tokens. Monthly totals: 40M input tokens, 4M output tokens. In the cached scenario, 1,500 of the 4,000 input tokens per conversation are a stable, cacheable prefix (system prompt plus policy pack): 15M cached tokens, 25M fresh.
| Model | Standard rates | With prompt caching |
|---|---|---|
| GPT-6 Luna | $6.00 | $4.65 |
| Claude Haiku 5.5 | $6.00 | $4.65 |
| DeepSeek V4.1 Flash (off-peak) | $8.40 | $6.20 |
| DeepSeek V4.1 Flash (peak) | $16.80 | $12.39 |
| Gemini 3.8 Flash (2026 rate) | $45.00 | $34.88 |
| GPT-6.1 Sol | $120.00 | $91.50 |
| Claude Sonnet 5.5 | $120.00 | $91.50 |
| Gemini 3.1 Pro Preview | $128.00 | $101.00 |
Worked example, GPT-6.1 Sol with caching: 25M fresh input × $2 = $50.00, plus 15M cached × $0.10 = $1.50, plus 4M output × $10 = $40.00 — $91.50 per month. The honest read: for a support bot, model tier matters less than answer quality, because even the workhorse tier costs about 1.2 cents per conversation. The budget tier costs 0.06 cents per conversation. If a budget model resolves tickets measurably worse, the “saving” is fictional — but if it resolves them equally on your ticket mix, the workhorse tier is a 20× overspend. Note also that Gemini 3.8 Flash’s row reprices on 1 January 2027, when its standard-rate cost for this workload rises from $45 to $90.
Workload 2: a content production pipeline
Assumptions: 100 finished articles per month. Each article takes three model calls — research summarization, drafting, and an editing pass — totaling 30,000 input tokens and 8,000 output tokens per article. Monthly totals: 3M input tokens, 0.8M output tokens. Nothing here is time-critical, so every row is also shown at batch/flex rates where the provider offers them.
| Model | Standard | Batch / Flex |
|---|---|---|
| Claude Haiku 5.5 | $0.70 | $0.35 |
| GPT-6 Luna | $0.70 | $0.35 |
| DeepSeek V4 Pro (off-peak) | $3.56 | n/a — shift timing instead |
| Gemini 3.8 Flash (2026 rate) | $5.25 | $2.63 |
| GPT-6.1 Sol | $14.00 | $7.00 |
| Claude Sonnet 5.5 | $14.00 | $7.00 |
| Gemini 3.1 Pro Preview | $15.60 | $7.80 |
| Claude Opus 5.5 | $28.00 | $14.00 |
Worked example, Claude Sonnet 5.5 at batch rates: 3M × $1.00 = $3.00 input, plus 0.8M × $5.00 = $4.00 output — $7.00 per month for 100 articles, or 7 cents of model cost per article. The takeaway is uncomfortable for anyone selling “AI content” at a premium: at pipeline scale, raw model cost is now a rounding error next to human editing time. What you are actually buying with a higher tier is fewer editing passes, not the tokens. DeepSeek V4 Pro has no batch product; its discount lever is the clock, so schedule the pipeline off-peak (its peak cost for this workload is $7.13). For a full content production pipeline, see our pipeline guide.
Workload 3: an autonomous AI agent
Assumptions: 5,000 agent runs per month (research tasks, inbox triage, lead enrichment — jobs where the agent plans and calls tools). Each run makes about 20 model calls, re-reading its context at every step: 200,000 total input tokens per run, of which 90% are repeat reads that hit the prompt cache, plus 5,000 output tokens. Monthly totals: 100M fresh input, 900M cached input, 25M output. Per-request prompts stay well under 100K tokens, so Haiku 5.5’s low tier applies.
| Model | Monthly cost | Cost per run |
|---|---|---|
| GPT-6 Luna | $31.50 | $0.006 |
| Claude Haiku 5.5 | $31.50 | $0.006 |
| DeepSeek V4.1 Flash (off-peak) | $32.70 | $0.007 |
| DeepSeek V4.1 Flash (peak) | $65.40 | $0.013 |
| DeepSeek V4 Pro (off-peak) | $135.30 | $0.027 |
| Gemini 3.8 Flash (2026 rate) | $236.25 | $0.047 |
| GPT-6.1 Sol | $540.00 | $0.108 |
| Claude Sonnet 5.5 | $540.00 | $0.108 |
| Claude Opus 5.5 | $1,080.00 | $0.216 |
Worked example, GPT-6.1 Sol: 100M fresh × $2 = $200, plus 900M cached × $0.10 = $90, plus 25M output × $10 = $250 — $540. Now the counterfactual that makes the point of this whole article: the same agent without caching pays full input price on all 1 billion input tokens — $2,000 of input plus $250 of output = $2,250. Caching alone saves 76%, more than any model switch on the board. Agents are cache-shaped workloads: they re-read far more than they write. That is also why the budget tier’s tiny cached-input prices ($0.01–$0.003 per MTok) compress the field so dramatically — DeepSeek’s off-peak Flash lands within a dollar of Luna and Haiku despite a higher fresh-input rate. Gemini’s row excludes explicit-cache storage fees, which would add a modest per-hour charge on top. See also our breakdown of what AI agents really cost per run.
Which model for which job: the decision framework
- High-volume, low-stakes work (classification, routing, tagging, first-draft summaries): budget tier — GPT-6 Luna or Claude Haiku 5.5 at $0.10 / $0.50. Keep Haiku prompts under 100K tokens.
- Customer-facing quality work (support answers, sales copy, document reasoning): workhorse tier — GPT-6.1 Sol or Claude Sonnet 5.5 at $2 / $10. Identical list prices, so choose on output quality in your own tests, and use caching either way.
- Scheduled pipelines (content, enrichment, reports): whichever tier passes your quality bar, run through Batch/Flex for a flat 50% off — or DeepSeek off-peak.
- Agents: optimize for the cached-input price and cache behavior first, model second. A $0.01 cached rate on a budget model beats a smarter model re-reading context at full price.
- The hardest reasoning (complex analysis, long-horizon coding): flagship tiers (GPT-6 Astra, Claude Fable 5.1, $10 / $50; Opus 5.5 at $4 / $20 as the step below). Pay it for the runs that need it, not as a global default.
- Prototyping: Google’s free tier is genuinely useful for experiments — but Google may use free-tier content to improve its products, so keep real customer data off it.
FAQ
Which AI API is cheapest in 2026?
At list price, GPT-6 Luna and Claude Haiku 5.5 tie at $0.10 input / $0.50 output per million tokens, with cached input at $0.01. DeepSeek V4.1 Flash is close behind off-peak ($0.15 / $0.60) and has the cheapest cache reads anywhere ($0.003). “Cheapest per token” is not “cheapest per task,” though — tokenizers differ, and a model that needs two attempts costs double.
Is Claude or GPT cheaper for a chatbot?
At the tiers most chatbots use, they are priced identically: $2 / $10 workhorse, $0.10 / $0.50 budget, $0.10 cached input on the workhorse tier and $0.01 on budget. On our 10,000-conversation workload both providers bill $120 standard or $91.50 cached at the workhorse tier. Pick on answer quality and features, not price.
How much does prompt caching actually save?
On repeat-heavy workloads, more than any other lever: cached input costs 5% of fresh input on GPT-6.1 Sol and Claude Sonnet 5.5, and 2–10% elsewhere. In our agent workload, caching cut the monthly bill from $2,250 to $540 — a 76% saving — because 90% of the agent’s input was re-read context.
What is batch pricing and who should use it?
Batch (OpenAI, Anthropic) and Flex (OpenAI, Google) tiers process requests asynchronously at 50% off standard rates, with results in minutes to 24 hours depending on the provider. Anything that runs on a schedule — content pipelines, enrichment, document processing, evaluations — should use it. Live chat should not.
Why did my bill jump even though my usage didn’t?
Check three things: a prompt that crossed a context threshold (272K on GPT-6.1 Sol, 200K on Gemini 3.1 Pro, 100K on Haiku 5.5) re-prices the whole request; a scheduled price change took effect (Gemini 3.8 Flash doubles on 1 January 2027); or reasoning tokens grew — on Gemini, output prices include thinking tokens, and DeepSeek’s thinking mode is on by default, so longer reasoning bills as more output.
Can I mix models to cut costs?
Yes, and most production systems do: route each request to the cheapest model that reliably does that job — budget tier for classification and drafts, workhorse for customer-facing answers, flagship only for the hardest steps. Routing plus caching plus batch is the full playbook; each lever is worth more than chasing a marginally cheaper list price.
Methodology
What we did: We collected the published standard per-1M-token prices for the current model generations from OpenAI, Anthropic, Google, and DeepSeek in October 2026, including cached-input, batch, and context-tier rates where the providers publish them. We then defined three workloads with fixed, printed token assumptions and computed monthly costs by direct multiplication of those assumptions against the published rates. Every figure in the workload tables can be reproduced from the assumptions stated above each table.
What we did NOT do: We did not run benchmark suites, measure answer quality, or test any provider’s API for this article, and we make no claim that the models listed are equal in capability — only that these are their published prices. The workloads are planning models, not measurements of a live business: your real token counts will differ, and the right way to use this article is to substitute your own logged token counts into the same arithmetic. Prices marked as introductory or time-bound are labeled with their published change dates.
Conflict disclosure: This article contains no affiliate links and no sponsored placements. Automated Wealth Lab’s other articles do carry affiliate links for some automation tools; none of the providers priced here is an affiliate partner, and no provider reviewed or approved this article before publication.
AWL verdict
For a solo founder in 2026, the default stack is: budget tier (Luna or Haiku 5.5) for volume work, workhorse tier (GPT-6.1 Sol or Sonnet 5.5) for anything a customer reads, batch pricing for everything scheduled, and caching turned on everywhere it applies. Model choice now moves your bill by percentage points; caching, batching, and context-threshold discipline move it by multiples. The one budget line to calendar: if you build on Gemini 3.8 Flash’s introductory rate, your model cost doubles on 1 January 2027 whether your revenue does or not.
Sources
- eesel AI — GPT-6.1 Sol pricing and tiers, sourced from OpenAI’s API pricing page — https://www.eesel.ai/blog/gpt-6-1-sol (verified October 2026)
- Finout — Anthropic API pricing guide, prices verified October 20, 2026 against official Anthropic documentation — https://www.finout.io/blog/anthropic-api-pricing (verified October 2026)
- Glamdring Research — Gemini API pricing for every model, from Google’s published Gemini API pricing page, checked September 28, 2026 — https://www.glamdringresearch.com/post/gemini-api-pricing (verified October 2026)
- TokenSave — DeepSeek API pricing, from DeepSeek’s official documentation, checked October 12, 2026 — https://tokensave.app/blog/deepseek-api-pricing (verified October 2026)
