- Per-million-token API prices now span more than 30x between providers’ flagship models, with DeepSeek’s V4 Pro priced near $1.32 input / $3.96 output against OpenAI’s GPT-6 Astra at $10 input / $50 output, as of November 2026.
- Every major lab now sells at least a three-tier lineup — a frontier reasoning model, a mid-tier “workhorse,” and a cheap high-volume model — rather than one flagship price.
- For most production workloads the mid-tier models (GPT-6.1 Terra, Claude Sonnet 5.5, Gemini 3.8 Flash) now deliver most of the frontier model’s coding and reasoning score at a fifth to a tenth of the cost.
- See our live, continuously updated LLM Price Index for current rates on every tracked model, since frontier pricing has moved at least once a quarter all through 2026.
| LLM API Pricing Snapshot (as of November 13, 2026) | |
|---|---|
| Cheapest flagship-class model | DeepSeek V4 Pro — $1.32 / $3.96 per million tokens (input/output, peak) |
| Most expensive flagship | OpenAI GPT-6 Astra — $10 / $50 per million tokens |
| Best price-to-performance mid-tier | GPT-6.1 Terra — $1 / $5 per million tokens |
| Longest context window in this set | Gemini 4 Carbon — 1 million tokens |
| Cheapest budget tier | DeepSeek V4.1 Flash off-peak — $0.15 / $0.60 per million tokens |
Why LLM API Prices Are So Hard to Compare
A single “price per model” number stopped being useful sometime in 2026. Every major lab now ships a tiered lineup instead of one flagship: OpenAI sells GPT-6 Astra, the mid-tier Terra, and the budget Luna model side by side; Anthropic sells Opus 5.5 alongside Sonnet 5.5; Google prices Gemini 4 Carbon well above its Flash line. Several providers also charge differently depending on prompt length, time of day, or whether a request hits a cache. DeepSeek, for example, runs peak and off-peak pricing that differs by roughly 2x, and xAI’s Grok 4.7 doubles its per-token rate once a prompt crosses 200,000 tokens. Any comparison that quotes one number per company is already out of date the moment it is published.
That said, the per-million-token input and output rate is still the right starting point, because it is the number every provider publishes and the one that scales directly with usage. The rest of this guide lines up current rates for the five providers builders ask about most, then works through what those numbers actually mean for a chatbot, a RAG pipeline, and a coding agent — the three workloads where token costs add up fastest.
Current API Pricing by Provider (November 2026)
| Provider / Model | Input ($/1M tokens) | Output ($/1M tokens) | Context window | Notes |
|---|---|---|---|---|
| OpenAI GPT-6 Astra (flagship) | $10.00 | $50.00 | 256K | Cached input $1.00; cache writes $12.50 |
| OpenAI GPT-6.1 Terra (mid-tier) | $1.00 | $5.00 | 256K | Cached input $0.05; half the price of GPT-5.6 Terra |
| Anthropic Claude Opus 5.5 | $4.00 | $20.00 | 200K | Batch API is 50% off both rates |
| Anthropic Claude Sonnet 5.5 | $2.00 | $10.00 | 200K | Cache reads $0.20; 1-hour cache writes $4.00 |
| Google Gemini 4 Carbon | $2.00 | $10.00 | 1M | Introductory rate through Nov 30, 2026; doubles to $4/$20 Dec 1 |
| Google Gemini 3.8 Flash | $0.75 | $3.75 | 1M | Rises to $1.50/$7.50 on Jan 1, 2027 |
| xAI Grok 4.7 | $2.00 | $6.00 | 500K | Doubles to $4/$12 past 200K-token prompts; cached input $0.50 |
| DeepSeek V4 Pro | $1.32 | $3.96 | 128K | Off-peak (UTC night hours) roughly half: $0.66/$1.98 |
| DeepSeek V4.1 Flash | $0.30 | $1.20 | 128K | Off-peak $0.15/$0.60; cache hits $0.006 |
These are list API rates pulled directly from each provider’s published pricing in November 2026; treat vendor-published numbers as claims, not guarantees, and check the LLM Price Index before budgeting a production workload, since several of the rates above are time-limited introductory offers.
Flagship vs. Mid-Tier: Where the Real Savings Are
The widest gap in this market right now isn’t between providers — it’s between a provider’s own flagship and mid-tier models. OpenAI’s GPT-6.1 Terra costs one-tenth of GPT-6 Astra per output token while, according to OpenAI’s own released benchmarks, retaining a meaningful share of Astra’s coding and reasoning score; Artificial Analysis placed Terra’s Intelligence Index six points above the previous-generation GPT-5.6 Terra while cutting the per-task cost to roughly a third. The same pattern holds at Anthropic, where Sonnet 5.5 costs a fifth of Opus 5.5 per output token and remains the default choice for most API integrations that don’t specifically need Opus-tier reasoning.
Google’s split is sharper still: Gemini 4 Carbon is built for multi-day autonomous coding and security workflows, not everyday chat or retrieval, and is priced and positioned accordingly. For most products, the Flash line — Gemini 3.8 Flash at $0.75/$3.75 through the end of 2026 — is the model actually worth integrating. The practical rule holds across all three labs: reach for the flagship only when a task genuinely needs the extra reasoning depth, because the mid-tier model is usually the better cost-per-task trade even before volume discounts are considered.
Three Worked Cost Examples
A customer-support chatbot. Assume a typical exchange runs 300 input tokens (system prompt, retrieved context, user message) and 150 output tokens, and the bot handles 50,000 conversations a month — 15M input tokens and 7.5M output tokens. On GPT-6.1 Terra ($1/$5), that’s $15 + $37.50 = $52.50/month. On Claude Sonnet 5.5 ($2/$10), it’s $30 + $75 = $105/month. On DeepSeek V4 Pro at off-peak rates ($0.66/$1.98), it drops to roughly $10 + $14.85 = $24.85/month. For a high-volume, latency-tolerant support bot, DeepSeek’s off-peak pricing or a Flash-tier model is the economically obvious choice.
A RAG (retrieval-augmented generation) pipeline. RAG workloads are input-heavy: a typical query might retrieve 4,000 tokens of context plus a 200-token question, generating a 300-token answer, across 20,000 queries a month — 84M input tokens, 6M output tokens. On Gemini 3.8 Flash ($0.75/$3.75), that’s $63 + $22.50 = $85.50/month, helped by Flash’s 1M-token context window if any single retrieval batch runs long. The same workload on GPT-6 Astra would run $840 + $300 = $1,140/month — a more than 13x difference for a task where the flagship’s extra reasoning adds little.
A coding agent. Agentic coding sessions are output-heavy and often re-read large chunks of a codebase as context, so caching matters more than the headline rate. A session that processes 50,000 input tokens (much of it cached repo context) and generates 15,000 output tokens, run 500 times a month by a small engineering team, totals 25M input and 7.5M output tokens. On GPT-6 Astra ($10/$50, cached input $1), with roughly 70% cache hits, input cost comes to about $10.75M-tokens-equivalent — call it $10.75 + $375 = $385.75/month; the output cost dominates, which is why coding-agent budgets should weight output price and cache-hit rate over the sticker input rate.
What Actually Drives Cost at Scale
Three factors matter more than the headline per-token price once usage grows. First, output tokens are consistently 4-5x the price of input tokens across every provider in this comparison, so workloads that generate long responses — code, long-form drafts, detailed analysis — cost far more per request than their input size suggests. Second, prompt caching can cut effective input costs by 80-95% for workloads that repeat large context blocks, which is exactly the coding-agent and RAG pattern above; a provider with a higher list price but strong cache discounts can beat a cheaper provider with no caching. Third, batch and off-peak discounts (Anthropic’s 50% batch rate, DeepSeek’s roughly 50% off-peak window) are free savings for anything that doesn’t need a synchronous response, which covers most back-office and data-pipeline use cases.
Frequently Asked Questions
Which LLM API is cheapest right now? By raw per-token rate, DeepSeek’s V4.1 Flash at off-peak pricing ($0.15 input / $0.60 output per million tokens) is the cheapest of the models tracked here as of November 2026, with V4 Pro the cheapest flagship-class option. Cheapest is workload-dependent, though — a model with poor cache support can still cost more at scale than a pricier one with strong caching.
Do these prices change often? Yes. Every provider in this comparison changed at least one tier’s pricing in the second half of 2026, and two of the rates above (Gemini 4 Carbon, Gemini 3.8 Flash) are scheduled to increase within weeks of publication. Always check a provider’s live pricing page, or the LLM Price Index, before committing a production budget.
Should I always pick the cheapest model? No. The worked examples above show mid-tier models like GPT-6.1 Terra or Sonnet 5.5 often deliver most of a flagship’s capability at a fraction of the cost, which is a better trade than the absolute cheapest option if that option’s output quality forces retries or human review.
Bottom Line
LLM API pricing in November 2026 isn’t a single number to compare — it’s a three-tier structure at every major lab, a growing set of caching and off-peak discounts, and output-token costs that dominate most real workloads. DeepSeek’s V4 line remains the cheapest way to buy raw tokens, OpenAI’s GPT-6.1 Terra and Anthropic’s Sonnet 5.5 are the two models most teams default to for production traffic, and the flagship tier — GPT-6 Astra, Claude Opus 5.5, Gemini 4 Carbon — is worth its premium only for tasks that genuinely need the extra reasoning depth. Model the actual input/output mix and cache-hit rate for your workload rather than comparing sticker prices, and recheck before shipping, because every provider here has moved its prices at least once this year.
