In this AI API pricing comparison as of November 2026, standard flagship rates per million input and output tokens are $10/$50 for GPT-6 Astra, $4/$20 for Claude Opus 5.5, and $2/$10 for Gemini 4 Argon. On December 1, 2026, Argon’s introductory discount ends and its price doubles to $4/$20.
Official LLM API pricing comparison table (November 2026)
As of November 1, 2026, official standard-tier API prices range from $0.10 to $10.00 per million input tokens and $0.50 to $50.00 per million output tokens across OpenAI, Anthropic, and Google.
All figures below are in USD per 1 million tokens on the standard (real-time) tier for global endpoints, checked on November 1, 2026.
| Model | Input | Cached input (cache hit) | Output | Notes |
|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | Short context (≤272K input tokens) |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | Cheaper cache than older GPT-6 Sol ($0.20) |
| GPT-6.1 Terra | $1.00 | $0.05 | $5.00 | Lower-cost mid-tier option |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | High-volume, small tier |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | Flat rate up to 1M-token context |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $10.00 | Flat rate up to 1M-token context |
| Claude Haiku 5.5 (≤100K input) | $0.10 | $0.01 | $0.50 | Prompts up to 100K tokens |
| Claude Haiku 5.5 (>100K input) | $0.50 | $0.05 | $2.50 | Prompts over 100K tokens |
| Gemini 4 Argon (through Nov 30, 2026) | $2.00 | $0.10 | $10.00 | Introductory pricing; no free tier |
| Gemini 4 Argon (starting Dec 1, 2026) | $4.00 | $0.20 | $20.00 | Standard pricing; all Argon rates double |
| Gemini 3.8 Flash (through Dec 31, 2026) | $0.75 | $0.075 | $3.75 | Rises to $1.50 / $0.15 / $7.50 on Jan 1, 2027 |
A few details sit behind those numbers:
- Google’s cheaper option isn’t a Gemini 4 model yet. If you need a lower-cost Google API today, Gemini 3.8 Flash is priced at $0.75 input and $3.75 output per million tokens through December 31, 2026, before rising to $1.50 / $0.15 / $7.50 on January 1, 2027. Google also announced Gemini 4 Carbon on October 28, 2026, but it has no public API price and is restricted to Google’s Fairwind Program.
- Older GPT-6 Sol is still listed. OpenAI’s pricing page still shows the original GPT-6 Sol at $2.00 input and $10.00 output with $0.20 cached input. GPT-6.1 Sol keeps the same base price ($2.00/$10.00) while cutting cached input in half to $0.10.
For context on what each of these models was built to do, see our breakdown of what the fall 2026 model wave changed.
What is a token, and why does output cost 5x input?
A token is a chunk of text that an AI model reads or writes, averaging about four characters or three-quarters of a word in English according to OpenAI’s Help Center. Across every model in the table above, output tokens cost five times as much as uncached input tokens.
As OpenAI’s Help Center notes, a lower price per million tokens does not necessarily produce a lower total cost, because models can tokenize the same text differently and generate different amounts of output or reasoning. Reasoning tokens are not visible as answer text, but they count toward output usage and are billed at the output rate.
Two details are worth keeping in mind:
- Every thinking token is billed as output. Google’s Gemini API pricing page states the same rule explicitly for Gemini 4 Argon: output pricing includes thinking tokens, and every token the model processes is counted and billed.
- Tokenizers vary across vendors. Anthropic notes that Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text than its earlier models. Because the same document can count as a different number of tokens on each provider, per-token list prices aren’t a one-to-one match.
How much does one real AI API task cost?
A single task on an AI API can cost anywhere from half a cent ($0.005) on small models to $0.50 on GPT-6 Astra for a long document summary, or from $2.75 to $550 per month for a 10,000-conversation support workload.
Because per-million-token tables are hard to picture in a budget meeting, I calculated two worked examples directly from the official list prices above. Anyone can verify these numbers with a calculator using the basic formula:
Cost = (input tokens × input price + cached tokens × cache price + output tokens × output price) ÷ 1,000,000
Example A: Summarize one 40-page report
Assumptions: You send 25,000 input tokens (roughly 18,000 to 19,000 English words, or a 40-page report) and get back 5,000 output tokens (about a 1,000-word summary plus the reasoning tokens billed as output). This uses the standard real-time tier with no caching.
| Model | Cost per report |
|---|---|
| GPT-6 Astra | $0.50 |
| Claude Opus 5.5 | $0.20 |
| Gemini 4 Argon (from Dec 1) | $0.20 |
| GPT-6.1 Sol | $0.10 |
| Claude Sonnet 5.5 | $0.10 |
| Gemini 4 Argon (through Nov 30) | $0.10 |
| GPT-6.1 Terra | $0.05 |
| Gemini 3.8 Flash (through Dec 31) | $0.0375 |
| GPT-6 Luna | $0.005 |
| Claude Haiku 5.5 (≤100K input) | $0.005 |
At the mid-tier ($2/$10), summarizing that 40-page report costs 10 cents. Moving down to GPT-6 Luna or Claude Haiku 5.5 drops the cost to half a cent—a 100x difference compared to GPT-6 Astra ($0.50).
Example B: Customer-support assistant (10,000 conversations a month)
Assumptions: Each conversation includes a 3,000-token fixed prefix (your system instructions and a help-center excerpt, which stays identical every time), 500 tokens of new customer text, and 400 output tokens.
- No caching: All 3,500 input tokens per conversation are billed at the base input rate.
- With caching: Assumes every read of the 3,000-token prefix is a cache hit, and leaves out the small initial cache-write charges.
- Batch (non-urgent work): Uses the 50% batch discount without caching. This only applies to work that can wait (such as tagging 10,000 support tickets overnight), not live customer chat.
| Model | No caching | With caching | Batch (non-urgent work) |
|---|---|---|---|
| GPT-6 Astra | $550 | $280 | $275 |
| Claude Opus 5.5 | $220 | $106 | $110 |
| Gemini 4 Argon (from Dec 1) | $220 | $106 | $110 |
| GPT-6.1 Sol | $110 | $53 | $55 |
| Claude Sonnet 5.5 | $110 | $53 | $55 |
| Gemini 4 Argon (through Nov 30) | $110 | $53 | $55 |
| GPT-6.1 Terra | $55 | $26.50 | $27.50 |
| Gemini 3.8 Flash (through Dec 31) | $41.25 | $21 | $20.63 |
| GPT-6 Luna | $5.50 | $2.80 | $2.75 |
| Claude Haiku 5.5 (≤100K input) | $5.50 | $2.80 | $2.75 |
Three things stand out from this table:
- Prompt caching cuts this bill roughly in half across every model tier because most of the input is repeated instructions.
- Model tier moves the bill 100x, from $5.50 a month on Luna or Haiku 5.5 to $550 on GPT-6 Astra without caching.
- These examples hold token counts constant to isolate unit price; in real runs, token counts differ by model.
Why can identical AI API list prices lead to an 8x difference in your bill?
Two models with identical per-token prices can produce wildly different bills because one model may generate several times more reasoning and output tokens to finish the same job.
Data from the independent tracker Artificial Analysis (checked November 1, 2026) shows this on its 9-benchmark Intelligence Index v5.0 suite. Its figures for Gemini 4 Argon already use the $4/$20 standard rate.
| Model (setting tested by AA) | Output tokens per task | Cost per task | AA Intelligence Index |
|---|---|---|---|
| GPT-6 Astra (Max) | 52K | $6.24 | 57 |
| GPT-6.1 Sol (Max) | 68K | $1.27 | 55 |
| Claude Opus 5.5 (Max) | 173K | $10.04 | 58 |
| Claude Sonnet 5.5 (Max) | 285K | $10.02 | 54 |
| Gemini 4 Argon (High) | 116K | $7.68 | 50 |
Look at Claude Sonnet 5.5 and GPT-6.1 Sol. Their Claude Sonnet 5.5 API pricing and GPT-6.1 Sol pricing are identical: $2.00 input, $0.10 cached input, and $10.00 output. Yet on Artificial Analysis’s suite, Sonnet 5.5 averaged $10.02 per task while GPT-6.1 Sol averaged $1.27 per task—nearly an 8x difference—because Sonnet 5.5 generated roughly 4x as many output tokens (285K vs. 68K, most of them reasoning tokens). GPT-6 Astra ($10/$50) even cost less per task ($6.24) than Sonnet 5.5 because it used only 52K output tokens.
Important caveat: These are hard, agentic benchmark tasks at maximum reasoning settings; everyday tasks use far fewer tokens. Treat this as a relative signal, not a forecast of your bill.
What changes on December 1 when Gemini 4 Argon’s introductory price ends?
On December 1, 2026, Gemini 4 Argon’s API pricing doubles across every category, moving from mid-tier pricing ($2.00 input / $10.00 output) to flagship pricing ($4.00 input / $20.00 output).
When Google opened Gemini 4 Argon to developers via the API on October 13, 2026, it launched with an introductory rate through November 30, 2026. Here is the before-and-after breakdown per 1 million tokens:
- Standard input: $2.00 through Nov 30 → $4.00 starting Dec 1
- Standard output (including thinking tokens): $10.00 through Nov 30 → $20.00 starting Dec 1
- Cached input (95% off base): $0.10 through Nov 30 → $0.20 starting Dec 1
- Batch API tier: $1.00 input / $5.00 output through Nov 30 → $2.00 input / $10.00 output starting Dec 1
Through November 30, the Gemini 4 Argon price matches GPT-6.1 Sol and Claude Sonnet 5.5 ($2/$10). From December 1, it matches Claude Opus 5.5 ($4/$20) and stays 2.5x cheaper per token than GPT-6 Astra ($10/$50). In Example B above (no caching), that doubles the monthly Argon bill from $110 to $220—so build any 2027 budget around the $4/$20 standard rate.
How prompt caching, batch pricing, and context limits change your bill
Prompt caching reduces repeated input token costs by 90% to 95%, while batch APIs cut both input and output rates by 50% for tasks that do not need an immediate reply. Long prompts, fast modes, and regional settings can add surcharges in the opposite direction.
Prompt caching (90% to 95% off repeated input)
If your workflow sends the same system prompt or reference document with every call, prompt caching discounts those repeated input tokens:
- Cache hit rates: GPT-6.1 Sol, GPT-6.1 Terra, Claude Opus 5.5, Claude Sonnet 5.5, and Gemini 4 Argon discount cached input hits by 95% (e.g., $2.00 drops to $0.10). GPT-6 Astra, GPT-6 Luna, Claude Haiku 5.5, and Gemini 3.8 Flash discount cache hits by 90%.
- Cache write fees: OpenAI charges $2.50 per million tokens for a 30-minute cache write on GPT-6.1 Sol. Anthropic charges 1.25x base input for a 5-minute cache write and 2x base input for a 1-hour cache write on Opus 5.5 and Sonnet 5.5.
- Implicit vs. explicit caching: Gemini 4 Argon supports implicit caching only; explicit caching is not supported.
Batch API (50% off for work that can wait)
For overnight classification, evaluations, or report generation that can wait:
- OpenAI Batch and Flex: 50% off standard prices (Sol $1/$5; Astra $5/$25; Luna $0.05/$0.25).
- Anthropic Batch API: 50% off (Opus 5.5 $2/$10; Sonnet 5.5 $1/$5).
- Google Batch tier: 50% off (Argon is $1/$5 through Nov 30, then $2/$10 from Dec 1).
Long-context surcharges, fast modes, and regional multipliers
- Long-context pricing: Anthropic includes the full 1M-token context window at standard prices on Claude Opus 5.5 and Sonnet 5.5. Claude Haiku 5.5 jumps from $0.10/$0.50 to $0.50/$2.50 when prompts exceed 100K tokens. OpenAI’s standard rates cover “short context” up to 272K input tokens; above 272K input tokens, GPT-6.1 Sol rises to $4.00/$15.00 and GPT-6 Astra rises to $20.00/$75.00.
- Fast modes: Priority speed doubles the price on OpenAI (Sol $4/$20; Astra $20/$100) and Anthropic (Opus 5.5 fast mode is $8/$40).
- Data residency: OpenAI adds +10% on regional endpoints for models released on or after March 5, 2026, and Anthropic adds a 1.1x multiplier for US-only inference (
inference_geo "us").
Which AI API is the cheapest and most cost-effective?
By per-token list price as of November 2026, the cheapest AI APIs in this lineup are GPT-6 Luna and Claude Haiku 5.5 (for prompts up to 100K tokens), tied at $0.10 per million input tokens and $0.50 per million output tokens. Which API is most cost-effective depends on your workload:
- High-volume classification, routing, and simple extraction: GPT-6 Luna or Claude Haiku 5.5 ($0.10/$0.50 up to 100K input tokens).
- Budget-conscious general tasks: Gemini 3.8 Flash ($0.75/$3.75 through Dec 31, 2026) and GPT-6.1 Terra ($1.00/$5.00).
- Everyday business work, coding, and writing: GPT-6.1 Sol and Claude Sonnet 5.5 ($2.00/$10.00, $0.10 cached), plus Gemini 4 Argon before its December 1 price change. Check reasoning token volume on your own prompts, as discussed in the cost-per-task section above.
- Very long documents (over 272K input tokens): Claude Sonnet 5.5 ($2/$10) and Claude Opus 5.5 ($4/$20) keep flat pricing up to 1M tokens without OpenAI’s >272K surcharge.
- Hardest flagship reasoning: Claude Opus 5.5 and Gemini 4 Argon (at its $4/$20 standard rate) list 2.5x below GPT-6 Astra ($10/$50) per token.
If your team can host models or use third-party inference providers, open-weight families from DeepSeek, Alibaba’s Qwen, and Meta are another option; see our guide to this year’s open-weight models.
Is the Claude API or OpenAI API cheaper?
At the mid-tier and small tier, Anthropic’s Claude API and OpenAI’s API have identical list prices, while at the flagship tier Claude Opus 5.5 is 60% cheaper per token than GPT-6 Astra.
Here is the tier-by-tier AI API cost comparison as of November 1, 2026:
- Flagship tier: Claude Opus 5.5 ($4.00/$0.20/$20.00) is 2.5x cheaper on list price than GPT-6 Astra ($10.00/$1.00/$50.00).
- Mid-tier: Claude Sonnet 5.5 and GPT-6.1 Sol tie at $2.00 input, $0.10 cached input, and $10.00 output, while OpenAI also offers GPT-6.1 Terra at $1.00/$5.00.
- Small tier: Claude Haiku 5.5 (up to 100K input tokens) and GPT-6 Luna tie at $0.10 input, $0.01 cached input, and $0.50 output (Haiku 5.5 rises to $0.50/$2.50 above 100K tokens).
In practice, your total bill depends on two caveats covered earlier: differences in tokenizer counts and reasoning output volume, and OpenAI’s long-context surcharge above 272K input tokens versus Claude 5.5’s flat 1M-token context.
API pricing vs. a $20/month chat subscription
API pricing is pay-per-token billing for developers and teams connecting AI models to apps or workflows, which is separate from a $20/month chat subscription such as ChatGPT Plus, Claude Pro, or Google AI Pro. Chat subscriptions charge a flat monthly fee per person with usage limits. Use a subscription when individuals are chatting in a browser; use the API when software processes work automatically at scale.
What I’d actually do before committing a budget
Don’t pick an API provider from a price table alone. To estimate your real bill:
- Run 20 to 50 representative tasks from your actual work (with sensitive data removed) across two or three candidate models, logging input, cached, and output tokens. Our guide on how to choose an AI model has a 30-minute test you can follow.
- Measure cost per successful task, as OpenAI’s GPT-6 model guide recommends, and include cache writes and long-context rates when estimating workflow cost.
- Enable prompt caching and batching wherever your prompts repeat or can wait overnight.
- Budget at standard prices, not promotional ones—use the December 1 rate ($4/$20) for Gemini 4 Argon and note the January 1, 2027 rate ($1.50/$7.50) for Gemini 3.8 Flash.
FAQ
How much do AI APIs cost?
As of November 2026, major AI APIs cost between $0.10 and $10.00 per million input tokens and between $0.50 and $50.00 per million output tokens on standard real-time tiers. Small models like GPT-6 Luna and Claude Haiku 5.5 (up to 100K tokens) cost $0.10/$0.50, mid-tier models like GPT-6.1 Sol and Claude Sonnet 5.5 cost $2.00/$10.00, and flagships cost $4.00/$20.00 (Claude Opus 5.5, and Gemini 4 Argon from Dec 1) up to $10.00/$50.00 (GPT-6 Astra).
Which AI API is the cheapest?
GPT-6 Luna and Claude Haiku 5.5 (for prompts up to 100K input tokens) are tied for the lowest list price among the major closed providers as of November 2026, at $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.50 per million output tokens. Using each provider’s Batch API for non-urgent work cuts those rates in half to $0.05 input and $0.25 output.
Which AI API is most cost-effective?
The most cost-effective AI API depends on how many output tokens a model uses on your specific task. While GPT-6.1 Sol and Claude Sonnet 5.5 share the same $2.00/$10.00 list price, independent testing by Artificial Analysis on hard reasoning tasks showed GPT-6.1 Sol averaging $1.27 per task versus $10.02 for Sonnet 5.5 because Sol generated about one-fourth as many output tokens. For prompts over 272K tokens, Claude 5.5 models avoid OpenAI’s long-context surcharge.
Is the Claude API or OpenAI API cheaper?
At the flagship tier, Anthropic’s Claude Opus 5.5 ($4.00/$20.00) is 2.5x cheaper per token than OpenAI’s GPT-6 Astra ($10.00/$50.00). At the mid-tier (Claude Sonnet 5.5 vs. GPT-6.1 Sol at $2.00/$10.00) and small tier (Claude Haiku 5.5 up to 100K tokens vs. GPT-6 Luna at $0.10/$0.50), their per-token list prices are identical, though OpenAI also offers GPT-6.1 Terra in between at $1.00/$5.00. Real bills depend on prompt length and reasoning token volume.
Does Gemini 4 Argon have a free tier?
No, Gemini 4 Argon is not available on the Gemini API free tier. Every token Argon processes—including internal thinking tokens—is counted and billed. Through November 30, 2026, Argon is billed at an introductory rate of $2.00 per million input tokens and $10.00 per million output tokens ($0.10 cached input). Starting December 1, 2026, its standard rate of $4.00 input, $20.00 output, and $0.20 cached input takes effect.
Prices last checked November 1, 2026. Next review: December 1, 2026.
Sources
- OpenAI, API Pricing, accessed November 1, 2026
- OpenAI, A model guide for the GPT-6 family, October 2, 2026
- OpenAI Help Center, Understanding and counting tokens, accessed November 1, 2026
- Anthropic, Pricing — Claude Platform Docs, accessed November 1, 2026
- Google, Gemini Developer API Pricing, last updated October 30, 2026 (accessed November 1, 2026)
- Google, Gemini 4 Argon: Built to solve your hardest problems, October 13, 2026
- Artificial Analysis, GPT-6 Astra vs GPT-6.1 Sol, Claude Sonnet 5.5 vs GPT-6.1 Sol, and Gemini 4 Argon vs Claude Opus 5.5, accessed November 1, 2026
