LLM prices / Reference
AI model reference · 71 models · docs checked Oct 24, 2026
AI model reference: limits and features, sourced
The facts you check before you ship: context window, max output, inputs, tool calling, JSON schema output, caching, batch, cutoff, and the rate limit on your tier. Every value links to the provider page it came from. Blank means the docs don't say.
Short answerOf 29 current models, the largest context window is 1,050,000 tokens (GPT-5.5 Pro) and the longest single reply is 384,000 tokens (DeepSeek V4 Pro). 29 support tool calling and 27 document schema-checked structured output. Rate limits: by provider and tier, with a fit checker.
Every model, side by side
Filter to what your app needs. Click a column to sort. Underlined values link to their source.
29 models shown
Context, max output, inputs and tool calling: each provider's model pages (also used on our price pages). Caching and batch: computed from the prices on each provider's pricing page (cache-read price as a share of the input price; batch input price vs standard). Cutoff, output types, structured output and reasoning: the model pages and feature guides, checked Oct 24, 2026.
Rate limits by tier, and will your traffic fit?
Pick your provider, model and tier, then your expected traffic. It checks each published limit and tells you which one you hit first.
Anthropic
RPM, ITPM and OTPM per model class, token bucket. Cached input reads do not count toward ITPM on current models. Separate Message Batches limits (RPM + requests in processing queue; batch_queue here counts requests, not tokens). Tiers are Start, Build, Scale, Custom; new orgs may start in a lower Evaluation tier. Monthly spend caps: Start: $500 USD; Build: $1,000 USD; Scale: $200,000 USD; Custom: No monthly spend cap; limits arranged with account team. No per-model numbers published for: Custom. Source: https://platform.claude.com/docs/en/api/rate-limits, checked Oct 24, 2026.
| Model (as the page names it) | StartAssigned automatically based on usage history and account standing | BuildAssigned automatically based on usage history and account standing | ScaleAssigned automatically based on usage history and account standing |
|---|---|---|---|
| Claude Fable 5.x (one combined limit for Claude Fable 5.1 and Claude Fable 5) | 1K RPM 500K Input TPM 100K Output TPM | 2K RPM 1.5M Input TPM 300K Output TPM | 4K RPM 4M Input TPM 800K Output TPM |
| Claude Opus 5.5 | 1K RPM 2M Input TPM 400K Output TPM | 5K RPM 5M Input TPM 1M Output TPM | 10K RPM 10M Input TPM 2M Output TPM |
| Claude Opus 5 | 1K RPM 2M Input TPM 400K Output TPM | 5K RPM 5M Input TPM 1M Output TPM | 10K RPM 10M Input TPM 2M Output TPM |
| Claude Opus 4.x (one combined limit for Opus 4.8, 4.7, 4.6 and 4.5) | 1K RPM 2M Input TPM 400K Output TPM | 5K RPM 5M Input TPM 1M Output TPM | 10K RPM 10M Input TPM 2M Output TPM |
| Claude Sonnet 5.5 | 1K RPM 2M Input TPM 400K Output TPM | 5K RPM 5M Input TPM 1M Output TPM | 10K RPM 10M Input TPM 2M Output TPM |
| Claude Sonnet 5 | 1K RPM 2M Input TPM 400K Output TPM | 5K RPM 5M Input TPM 1M Output TPM | 10K RPM 10M Input TPM 2M Output TPM |
| Claude Sonnet 4.x (one combined limit for Sonnet 4.6 and Sonnet 4.5) | 1K RPM 2M Input TPM 400K Output TPM | 5K RPM 5M Input TPM 1M Output TPM | 10K RPM 10M Input TPM 2M Output TPM |
| Claude Haiku 5.5 | 1K RPM 2M Input TPM 400K Output TPM | 5K RPM 5M Input TPM 1M Output TPM | 10K RPM 10M Input TPM 2M Output TPM |
| Claude Haiku 4.5 | 1K RPM 2M Input TPM 400K Output TPM | 5K RPM 5M Input TPM 1M Output TPM | 10K RPM 10M Input TPM 2M Output TPM |
| Message Batches API (shared across all models; max 100,000 requests per batch) | 1K RPM 200K Batch queue | 2K RPM 300K Batch queue | 4K RPM 500K Batch queue |
OpenAI
RPM, RPD, TPM, TPD (and IPM for images), per organization and project, per model. Batch queue limit = input tokens queued per model. Long context requests have separate limits on some models. Per-model numbers come from each model page; the Free tier has no per-model numbers published. Monthly spend caps: Free: $100 / month; Build: $500 / month; Launch: $5,000 / month; Grow: $200,000 / month. No per-model numbers published for: Free. Source: https://developers.openai.com/api/docs/guides/rate-limits, checked Oct 24, 2026.
| Model (as the page names it) | Build$5 in total credit purchases | Launch$100 in total credit purchases | Grow$500 in total credit purchases |
|---|---|---|---|
| gpt-6-astra | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-6.1-sol | 5K RPM 1M TPM | 10K RPM 4M TPM | 15K RPM 40M TPM |
| gpt-6-luna | 5K RPM 2M TPM 20M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 180M TPM 15B Batch queue |
| gpt-6-sol | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-5.6-sol | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-5.6-terra | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-5.6-luna | 5K RPM 2M TPM 20M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 180M TPM 15B Batch queue |
| gpt-5.5 | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-5.5 (long context, 272K input tokens) | 500 RPM 1M TPM 40M Batch queue | 2K RPM 10M TPM 200M Batch queue | 8K RPM 20M TPM 2B Batch queue |
| gpt-5.5-pro | 500 RPM 200K TPM 1M Batch queue | 1K RPM 1M TPM 20M Batch queue | 2K RPM 4M TPM 1.5B Batch queue |
| gpt-5.4 | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-5.4 (long context, 272K input tokens) | 500 RPM 1M TPM 40M Batch queue | 2K RPM 10M TPM 200M Batch queue | 8K RPM 20M TPM 2B Batch queue |
| gpt-5.4-pro | 500 RPM 100K TPM 1.35M Batch queue | 1K RPM 400K TPM 200M Batch queue | 1.5K RPM 4M TPM 15B Batch queue |
| gpt-5.4-pro (long context, 272K input tokens) | 50 RPM 100K TPM 20M Batch queue | 200 RPM 1M TPM 100M Batch queue | 800 RPM 2M TPM 1B Batch queue |
| gpt-5.4-mini | 5K RPM 2M TPM 20M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 180M TPM 15B Batch queue |
| gpt-5.4-nano | 5K RPM 2M TPM 20M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 180M TPM 15B Batch queue |
| gpt-5.2 | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-5.2-pro | 5K RPM 450K TPM 1.35M Batch queue | 10K RPM 2M TPM 200M Batch queue | 10K RPM 30M TPM 5B Batch queue |
| gpt-5.2-pro (long context, 128k input tokens) | 250 RPM 500K TPM 20M Batch queue | 1K RPM 5M TPM 100M Batch queue | 4K RPM 10M TPM 1B Batch queue |
| gpt-5.1 | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-5 | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-5-mini | 5K RPM 2M TPM 20M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 180M TPM 15B Batch queue |
| gpt-5-nano | 5K RPM 2M TPM 20M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 180M TPM 15B Batch queue |
| gpt-5-pro | 5K RPM 450K TPM 1.35M Batch queue | 10K RPM 2M TPM 200M Batch queue | 10K RPM 30M TPM 5B Batch queue |
| gpt-4.1 | 5K RPM 450K TPM 1.35M Batch queue | 10K RPM 2M TPM 200M Batch queue | 10K RPM 30M TPM 5B Batch queue |
| gpt-4.1 (long context, 128k input tokens) | 250 RPM 500K TPM 20M Batch queue | 1K RPM 5M TPM 100M Batch queue | 4K RPM 10M TPM 1B Batch queue |
| gpt-4.1-mini | 5K RPM 2M TPM 20M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 150M TPM 15B Batch queue |
| gpt-4.1-mini (long context, 128k input tokens) | 500 RPM 1M TPM 40M Batch queue | 2K RPM 10M TPM 200M Batch queue | 8K RPM 20M TPM 2B Batch queue |
| gpt-4.1-nano | 5K RPM 2M TPM 20M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 150M TPM 15B Batch queue |
| gpt-4.1-nano (long context, 128k input tokens) | 500 RPM 1M TPM 40M Batch queue | 2K RPM 10M TPM 200M Batch queue | 8K RPM 20M TPM 2B Batch queue |
| gpt-4o | 5K RPM 450K TPM 1.35M Batch queue | 10K RPM 2M TPM 200M Batch queue | 10K RPM 30M TPM 5B Batch queue |
| gpt-4o-mini | 5K RPM 2M TPM 20M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 150M TPM 15B Batch queue |
| o1 | 5K RPM 450K TPM 1.35M Batch queue | 10K RPM 2M TPM 200M Batch queue | 10K RPM 30M TPM 5B Batch queue |
| o1-pro | 5K RPM 450K TPM 1.35M Batch queue | 10K RPM 2M TPM 200M Batch queue | 10K RPM 30M TPM 5B Batch queue |
| o3-pro | 5K RPM 450K TPM 1.35M Batch queue | 10K RPM 2M TPM 200M Batch queue | 10K RPM 30M TPM 5B Batch queue |
| o3 | 5K RPM 450K TPM 1.35M Batch queue | 10K RPM 2M TPM 200M Batch queue | 10K RPM 30M TPM 5B Batch queue |
| o4-mini | 2K RPM 2M TPM 2M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 150M TPM 15B Batch queue |
| o3-mini | 2K RPM 200K TPM 2M Batch queue | 10K RPM 10M TPM 1B Batch queue | 30K RPM 150M TPM 15B Batch queue |
| gpt-3.5-turbo | 5K RPM 2M TPM 5M Batch queue | 10K RPM 10M TPM 1B Batch queue | 10K RPM 50M TPM 10B Batch queue |
| gpt-5.6-cyber | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| chat-latest | 5K RPM 450K TPM 1.35M Batch queue | 10K RPM 2M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
| gpt-5.3-codex | 5K RPM 1M TPM 3M Batch queue | 10K RPM 4M TPM 200M Batch queue | 15K RPM 40M TPM 15B Batch queue |
RPM, TPM (input) and RPD per project; RPD resets at midnight Pacific. Per-model RPM/TPM/RPD tables are no longer published on the page, only in AI Studio. Spend-based limit per rolling 10 minutes. Published per-model numbers are batch enqueued tokens only; 100 concurrent batch requests. No per-model numbers published for: Free. Source: https://ai.google.dev/gemini-api/docs/rate-limits, checked Oct 24, 2026.
| Model (as the page names it) | Tier 1Set up and link an active billing account | Tier 2Paid $100 + 3 days from first successful payment | Tier 3Paid $1,000 + 30 days from first successful payment |
|---|---|---|---|
| Gemini 4 Argon | 5M Batch queue | 500M Batch queue | 1B Batch queue |
| Gemini 3.1 Pro Preview | 5M Batch queue | 500M Batch queue | 1B Batch queue |
| Gemini 3.5 Flash-Lite | 10M Batch queue | 500M Batch queue | 1B Batch queue |
| Gemini 3.8 Flash | 3M Batch queue | 400M Batch queue | 1B Batch queue |
| Gemini 3.1 Flash Lite | 10M Batch queue | 500M Batch queue | 1B Batch queue |
| Gemini 3.6 Flash | 3M Batch queue | 400M Batch queue | 1B Batch queue |
| Gemini 2.5 Pro | 5M Batch queue | 500M Batch queue | 1B Batch queue |
| Gemini 2.5 Flash | 3M Batch queue | 400M Batch queue | 1B Batch queue |
| Gemini 2.5 Flash Lite | 10M Batch queue | 500M Batch queue | 1B Batch queue |
DeepSeek
No RPM/TPM tiers published; a per-account concurrency limit per model (requests in flight). Over the limit returns HTTP 429. Higher concurrency on request at no extra cost. Source: https://api-docs.deepseek.com/quick_start/rate_limit, checked Oct 24, 2026.
| Model (as the page names it) | All accounts (no usage tiers) |
|---|---|
| deepseek-flash | 2.5K In flight |
| deepseek-v4-pro | 500 In flight |
Caching, batch and structured output, by provider
| Provider | Prompt caching | Batch API | Structured output |
|---|---|---|---|
| Anthropic | automatic (one top-level cache_control field) or explicit cache_control breakpoints on content blocks, up to 4. Reads: 10% of input price; 2.5% on Claude Fable 5.1, 5% on Claude Opus 5.5 and Claude Sonnet 5.5. Writes cost 1.25x input (5 min) or 2x (1 hour). Minimum cacheable prompt is 512 to 4,096 tokens depending on model Lifetime: 5 minutes by default (refreshed free on each hit); 1 hour optional at 2x input price. source | 50% off input and output; most batches finish within 1 hour; results available when done or after 24 hours, batches expire if not done in 24 hours. source | JSON outputs via output_config.format (JSON schema, constrained decoding) and strict tool use (strict: true). source |
| OpenAI | automatic (implicit breakpoints); GPT-5.6 and later also support explicit breakpoints (prompt_cache_breakpoint, up to 4 cache writes). Reads: GPT-5.6 and later: 10% of input (5% on GPT-6.1 Sol), writes 1.25x input (30 min) or 2.5x (12 h). Earlier models: model-specific cached-input rate, no write charge. 1,024 tokens for GPT-5.6 and later; varies by request settings on earlier models Lifetime: GPT-5.6 and later: minimum 30 minutes (default) or 12 hours (Responses API). Earlier: in_memory about 5 to 10 minutes idle up to 1 hour, or 24h extended retention on supported models. source | 50% off; within 24 hours (completion_window 24h only). source | Structured Outputs: JSON schema via text.format json_schema with strict: true (Responses) or response_format (Chat Completions), and strict function calling. source |
| automatic (implicit caching on Gemini 2.5 and newer); explicit cache objects on supported Gemini 3 and 2.5 models via generateContent only, not on Gemini 4 Argon or the Interactions API. Reads: per-model context caching price on the pricing page, e.g. Gemini 4 Argon $0.10 per 1M vs $2.00 input; explicit caches also pay hourly storage. 2,048 to 4,096 tokens for implicit caching depending on model Lifetime: explicit caches default to 1 hour, settable; implicit cache lifetime not stated. source | 50% of standard cost; target 24 hours, often much faster. source | JSON schema via response_format (type text, mime_type application/json, schema). source | |
| DeepSeek | automatic (context caching on disk, on by default, best effort). Reads: cache-hit input price, e.g. deepseek-flash $0.006 vs $0.30 per 1M (peak); deepseek-v4-pro $0.044 vs $1.32 (peak). Lifetime: cleared automatically when unused, usually within a few hours to a few days. source | No batch API in the docs. | JSON Output via response_format {type: json_object} (valid JSON, no schema); strict tool calling with JSON schema in beta (base_url /beta, strict: true). source |
What the docs don't say
Where a cell is blank, we looked and the provider doesn't publish it. Specifically:
- Google: per-model RPM/TPM/RPD tables are not published on ai.google.dev/gemini-api/docs/rate-limits (only viewable in AI Studio while signed in); only batch enqueued token limits per model are recorded.
- Google: Gemini 3 Flash Preview (gemini-3-flash-preview) has no row in the batch enqueued tokens tables.
- Google: knowledge cutoff is only stated on the Gemini 2.5 model pages; Gemini 4 Argon, 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash Lite, 3.1 Pro Preview and 3 Flash Preview pages give no cutoff (null).
- Google: implicit cache minimum tokens are listed only for 4 Argon, 3.8 Flash, 3.6 Flash, 3.1 Pro Preview, 2.5 Flash, 2.5 Pro.
- DeepSeek: docs state no knowledge cutoff for deepseek-flash or deepseek-v4-pro (null).
- DeepSeek: no Batch API is documented (batch null).
- DeepSeek: structured_output true means JSON Output mode (json_object, no schema enforcement); schema-enforced output only via beta strict tool calling.
- DeepSeek: output_modalities ["text"] taken from the chat completion API reference (message content is text); the models page does not list output modalities explicitly.
- OpenAI: gpt-5.2-pro and gpt-5.4-pro model pages do not list structured_outputs under supported features (and have no unsupported list), so structured_output is null.
- OpenAI: chat-latest page has no "Reasoning token support" line but it is a moving alias, so reasoning is null; other models without that line are recorded as reasoning false.
- OpenAI: usage tiers are now Free, Build, Launch, Grow (not Tier 1-5). Model pages publish Build/Launch/Grow rows only; no Free tier per-model limits.
- OpenAI: gpt-6.1-sol page publishes no batch queue limit (null). Rows for deprecated and shut-down models (gpt-4.1-nano, o1, o1-pro, o3-mini, o4-mini, gpt-3.5-turbo etc.) are included as their pages still publish them.
- Anthropic: tiers are now Start, Build, Scale, Custom (not Tier 1-4), with no published dollar qualification; Claude Haiku 3.5 rows skipped (not in our list).
- Anthropic: prompt caching min_tokens varies by model, so provider-level min_tokens is null and min_tokens_by_model is given.
Method
Every value was read from the provider's own documentation (Anthropic, OpenAI, Google Gemini API, DeepSeek) on Oct 24, 2026, never from memory or third-party lists. Each cell links to the exact page. Limits change often and vary by account; the provider console shows your real numbers.
Up next
Now price it: You know which models fit. The picker prices them on your job and volume and gives 3 picks with the monthly cost.Pick an AI model for your jobLLM API pricing for 71 models per 1M tokens, checked at the source. Pick for me: your job, volume and budget give 3 models with the monthly cost. Filters, CSV.Read it next →