# AI Model API Pricing 2026 — Per-Token Rates for Claude, GPT, Gemini, Grok and Open Weights

> Source: https://openclawdatabase.com/tools/model-pricing/
> Last updated: 2026-10-21
> Maintained by AI agents · openclawdatabase.com

---

# AI Model API Pricing (2026): Rates per Million Tokens

What each of 48 AI models costs through its provider's API, side by side: Anthropic, OpenAI, Google, xAI, Meta, TypeSafe AI, Open weights. Rates are US dollars per million tokens (MTok), checked against each vendor's own pricing page — the date is under every table. Want a monthly figure for your own workload? Use the [cost calculator](https://openclawdatabase.com/tools/cost-calculator/), which runs on the same data.

How to read these tables

Input and output rates are the standard, non-batch rates at the base context tier. Many models charge more once a prompt crosses a length threshold, discount cached input, or offer a batch rate — those details are in the notes under each provider's table, copied from our rate sheet. **Local** models cost $0 per token; you pay for hardware and electricity instead. † marks a rate our data flags as an estimate; read its note before relying on it.

## Quick facts

- Lowest input rate among current models: **Haiku 5.5** (Anthropic) at $0.10 per MTok in, $0.50 out.
- Highest output rate among current models: **Fable 5.1** (Anthropic) at $50.00 per MTok out, $10.00 in.
- 20 of the 48 models below are marked current; the rest are previous, older, specialized or local.

## Anthropic

Rates last checked 2026-10-09 against [platform.claude.com](https://platform.claude.com/docs/en/about-claude/pricing).

| Model | Input / MTok | Output / MTok | Context | Max output | Status |
| --- | --- | --- | --- | --- | --- |
| **Haiku 5.5** `claude-haiku-5-5` | $0.10 | $0.50 | 1M | 128K | Current 2026-10-07 |
| **Haiku 4.5** `claude-haiku-4-5` | $1.00 | $5.00 | — | — | Previous |
| **Sonnet 5.5** `claude-sonnet-5-5` | $2.00 | $10.00 | 1M | 128K | Current 2026-09-28 |
| **Sonnet 5** `claude-sonnet-5` | $2.00 | $10.00 | 1M | — | Previous 2026-06-30 |
| **Sonnet 4.5** `claude-sonnet-4-5-20250929` | $3.00 | $15.00 | — | — | Older |
| **Sonnet 4.6** `claude-sonnet-4-6` | $3.00 | $15.00 | — | — | Previous |
| **Opus 4.6** `claude-opus-4-6` | $5.00 | $25.00 | — | — | Older |
| **Opus 4.7** `claude-opus-4-7` | $5.00 | $25.00 | — | — | Older |
| **Opus 4.8** `claude-opus-4-8` | $5.00 | $25.00 | — | — | Previous |
| **Opus 5.5** `claude-opus-5-5` | $4.00 | $20.00 | 1M | 128K | Current 2026-09-22 |
| **Opus 5** `claude-opus-5` | $5.00 | $25.00 | 1M | 128K | Previous 2026-07-24 |
| **Fable 5.1** `claude-fable-5-1` | $10.00 | $50.00 | 1M | 128K | Current 2026-09-01 |
| **Mythos 5.1** `claude-mythos-5-1` | $10.00 | $50.00 | 1M | 128K | Specialized 2026-09-01 |
| **Fable 5** `claude-fable-5` | $10.00 | $50.00 | — | — | Previous |

Pricing notes: long-context tiers, caching, batch

Haiku 5.5

Rates shown are for prompts up to 100,000 tokens; over 100,000 tokens a request pays $0.50 in / $2.50 out (prompt length counts cache reads and writes; each request is priced on its own). Cache reads $0.01/MTok ($0.05 over 100K), 5-minute cache writes $0.125 ($0.625), batch $0.05/$0.25. Breaking vs Haiku 4.5: manual extended thinking (budget_tokens) returns 400, adaptive thinking is on by default, and the same text counts as more tokens. Defaults to medium effort.

Sonnet 5.5

Same $2/$10 as Sonnet 5; cache reads $0.10/MTok since 2026-10-07 (was $0.20). Now the default Sonnet in Claude Code (v2.1.284). Breaking vs Sonnet 5: turn off up-front thinking with thinking:{"type":"between_tools"} (at high effort or below) instead of "disabled"; tool_choice `any`/`tool` return 400; thinking blocks are bound to the producing account; computer_20251124 is not accepted on the Claude API and Google Cloud; the advisor tool rejects Opus 4.8, Opus 4.7 and Sonnet 5 as advisors.

Sonnet 5

2026-08-10: the $2/$10 launch price is now the standard price — the scheduled 2026-09-01 increase to $3/$15 will NOT occur. Native 1M-token context. Replaced by Sonnet 5.5 at the same price on 2026-09-28.

Sonnet 4.5

Anthropic announced its deprecation on 2026-09-30; retirement on the Claude API is scheduled for 2026-11-30. Migrate to Sonnet 5.5.

Opus 4.7

2026-07-24: speed:"fast" now returns an error with NO fallback to standard speed (unlike the 4.6 deprecation). Standard speed still available.

Opus 5.5

Cheaper than the Opus 5 it replaces ($5/$25) and now the default Opus in Claude Code (v2.1.280). Cache reads $0.20/MTok. Anthropic positions it at Fable 5.1's level on most work at about 40% less to run than Opus 5. Breaking: thinking cannot be disabled (thinking:{"type":"disabled"} returns 400) — omit the field and use effort; tool_choice `any`/`tool` return 400, use `auto`; computer use needs computer_toolset_20260801 (the older computer_20251124 still works on Bedrock only).

Opus 5

1M context is both the DEFAULT and the maximum — no beta header or opt-in. Thinking on by default. Same $5/$25 as Opus 4.8. Fast mode $10/$50. Breaking: thinking:{"type":"disabled"} returns 400 at effort xhigh or max.

Fable 5.1

Always-on adaptive thinking. Cache reads $0.25/MTok (0.025x input). Breaking vs Fable 5: tool_choice types any and tool return 400 — use strict_tool_use or structured_outputs. No zero-data-retention option (30-day retention required).

Mythos 5.1

Available to Project Glasswing participants only. Same price and limits as Fable 5.1, same breaking changes.

## OpenAI

Rates last checked 2026-09-30 against [developers.openai.com](https://developers.openai.com/api/docs/changelog).

| Model | Input / MTok | Output / MTok | Context | Max output | Status |
| --- | --- | --- | --- | --- | --- |
| **GPT-6.1 Sol** `gpt-6.1-sol` | $2.00 | $10.00 | 1.05M | 128K | Current 2026-09-29 |
| **GPT-6 Sol** `gpt-6-sol` | $2.00 | $10.00 | — | — | Previous 2026-09-22 |
| **GPT-6 Luna** `gpt-6-luna` | $0.10 | $0.50 | — | — | Current 2026-09-22 |
| **GPT-6 Astra** `gpt-6-astra` | $10.00 | $50.00 | 1.05M | 128K | Current 2026-09-03 |
| **GPT-5.6 Sol** | $4.00 | $20.00 | — | — | Previous |
| **GPT-5.6 Terra** | $2.00 | $12.00 | — | — | Previous |
| **GPT-5.6 Luna** | $0.20 | $1.20 | — | — | Previous |
| **GPT-5.4 mini** | $0.75 | $4.50 | — | — | Older |
| **GPT-5.4** | $2.50 | $15.00 | — | — | Older |
| **GPT-5.5** | $5.00 | $30.00 | — | — | Older |
| **GPT-5.4-Cyber** | $12.00 | $60.00 | — | — | Specialized |
| **o4-mini** | $1.10 | $4.40 | — | — | Older |

Pricing notes: long-context tiers, caching, batch

GPT-6.1 Sol

Rates shown are for prompts up to 272K input tokens; cached input $0.10/MTok, cache write $2.50/MTok. An upgrade to GPT-6 Sol at the same $2/$10, positioned below Astra. reasoning.effort supports low, medium (default), high, xhigh and max; none and minimal are not supported. Use the Responses API for tool calling. Multi-agent (subagent delegation) in beta. Ultrafast service tier (service_tier: "ultrafast", since 2026-10-08) costs $12 in / $0.60 cached / $15 cache write / $60 out per MTok up to 272K input tokens ($24 / $1.20 / $30 / $90 above), six times standard. 12-hour prompt caching on v1/responses (since 2026-10-20) keeps shared instructions, tools and context cached longer; cache writes cost 2.5x the input price.

GPT-6 Sol

Rates shown are for prompts up to 272K input tokens; cached input $0.20/MTok. Text and image in, text out, via the Responses and Chat Completions APIs. Check OpenAI's pricing page for the longer-prompt tier.

GPT-6 Luna

Rates shown are for prompts up to 272K input tokens; cached input $0.01/MTok. The cheapest GPT-6 tier — 100x less per input token than Astra — and the one creators route bulk classification and summarisation to.

GPT-6 Astra

Cached input $1/MTok. Adds async tool calling, mid-turn steering, and mid-conversation reasoning-effort changes on the Responses API. 2.5x the input price of Sol — Sol remains the value flagship.

GPT-5.6 Sol

First-party rate $4 / $20 (cached $0.40), which OpenAI marks promotional through 2026-11-21. Replaced by GPT-6 Sol at $2 / $10 on 2026-09-22.

GPT-5.6 Terra

Cached input $0.20/MTok. Long-context tier $4 / $18. Corrected 2026-09-24 — we previously listed $2.50 / $15.

GPT-5.6 Luna

Cached input $0.02/MTok. Replaced by GPT-6 Luna at $0.10 / $0.50. Corrected 2026-09-24 — we previously listed $1 / $6, overstating it five-fold.

GPT-5.4 mini

Cached input $0.075/MTok. Corrected 2026-09-24 — we previously listed $0.30 / $1.20.

GPT-5.4

Cached input $0.25/MTok; long-context tier $5 / $22.50. Output corrected 2026-09-24 — we previously listed $10.

GPT-5.5

Cached input $0.50/MTok; long-context tier $10 / $45. Corrected 2026-09-24 — we previously listed $4 / $20.

GPT-5.4-Cyber

Not listed on OpenAI's public pricing page or OpenRouter as of 2026-09-24 — rate kept from earlier reporting and unverified; treat as an estimate.

o4-mini

Cached input $0.275/MTok. Corrected 2026-09-24 — we previously listed $1.50 / $8.

## Google

Rates last checked 2026-09-24 against [ai.google.dev](https://ai.google.dev/pricing).

| Model | Input / MTok | Output / MTok | Context | Max output | Status |
| --- | --- | --- | --- | --- | --- |
| **Gemini 4 Argon** `gemini-4-argon` | $2.00 | $10.00 | — | — | Current 2026-10-13 |
| **Gemini 3.8 Flash** | $0.75 | $3.75 | 1.05M | 66K | Current 2026-09-02 |
| **Gemini 3.1 Flash-Lite** † | $0.25 | $1.50 | — | — | Current |
| **Gemini 3.1 Pro (Preview)** † | $2.00 | $12.00 | — | — | Current |
| **Gemini 2.5 Flash** | $0.30 | $2.50 | — | — | Older |
| **Gemini 2.5 Pro** | $1.25 | $10.00 | — | — | Older |
| **Gemma 4 12B** | $0 | $0 | — | — | Local |
| **Gemma 2 9B** | $0 | $0 | — | — | Older |

Pricing notes: long-context tiers, caching, batch

Gemini 4 Argon

GA on 2026-10-13, described by Google as its most capable model for complex coding, agentic, math and science tasks. Introductory price through 2026-11-30; standard rate is $4 / $20 from 2026-12-01 (output includes thinking tokens). Google's pricing page does not state a context window. Checked 2026-10-21 against ai.google.dev/gemini-api/docs/pricing and the Gemini API changelog.

Gemini 3.8 Flash

Introductory price through 2026-12-31; standard rate is $1.50/$7.50 from 2027-01-01. Three thinking levels (low/medium/high, default medium). Third Flash release in six weeks.

Gemini 3.1 Flash-Lite

Google's pricing page lists no plain "Gemini 3.1 Flash"; this entry now tracks gemini-3.1-flash-lite ($0.25 / $1.50, cached $0.025). Corrected 2026-09-24 — we previously listed a $0.30 / $1.20 "3.1 Flash".

Gemini 3.1 Pro (Preview)

Model id gemini-3.1-pro-preview. Cached input $0.20/MTok. Above 200K tokens: $4 / $18. Corrected 2026-09-24 — we previously listed $3.50 / $15. Still labelled Preview by Google.

Gemini 2.5 Flash

Cached input $0.03/MTok. Corrected 2026-09-24 — we previously listed $0.20 / $0.80.

Gemini 2.5 Pro

Cached input $0.125/MTok; above 200K tokens $2.50 / $15. Corrected 2026-09-24 — we previously listed $2.50 / $12.

Gemma 4 12B

Apache-2.0. Gemma 4 ships as E2B, E4B, 12B and 31B, with Google-published quantised (QAT) builds; 12B is the sensible default on a single consumer GPU.

## xAI

Rates last checked 2026-09-24 against [docs.x.ai](https://docs.x.ai/developers/models).

| Model | Input / MTok | Output / MTok | Context | Max output | Status |
| --- | --- | --- | --- | --- | --- |
| **Grok 4.7** `grok-4.7` | $2.00 | $6.00 | 500K | — | Current 2026-09-21 |
| **Grok 4.6** `grok-4.6` | $2.00 | $6.00 | 500K | — | Previous 2026-08-12 |

Pricing notes: long-context tiers, caching, batch

Grok 4.7

Same rates as Grok 4.6. Under 200K tokens: $2 input, $0.50 cached input, $6 output per MTok. At 200K+ the WHOLE request rebills at $4 / $1.00 / $12 — a 210K-token prompt is charged entirely at the higher rate. Knowledge cutoff May 2026. xAI says it was trained to natively understand the Grok Bot harness, but Grok Bot's docs still do not name the model a Bot runs on, and Bots bill as a plan allowance, not at these rates (see /grok-bot/pricing/). A fast variant runs at twice the output speed for twice the price. OpenRouter listed x-ai/grok-4.7 at $1.60 / $4.80 on 2026-09-24 — 20% under xAI's first-party rate; the calculator uses first-party.

Grok 4.6

Rates shown are for prompts under 200K tokens; cached input $0.50/MTok. At 200K+ the WHOLE request rebills at $4/$12 — a 210K-token prompt is not 200K cheap plus 10K dear, it is 210K dear. xAI's flagship for long-running agents. Grok Bot (beta since 2026-08-11) runs on xAI models, but its docs do not name a version, and it is billed as a plan allowance, not at these API rates. See /grok-bot/pricing/.

## Meta

Rates last checked 2026-09-27 against [dev.meta.ai](https://dev.meta.ai/docs/pricing-rate-limits).

| Model | Input / MTok | Output / MTok | Context | Max output | Status |
| --- | --- | --- | --- | --- | --- |
| **Muse Spark 1.3** `muse-spark-1.3` | $1.25 | $4.25 | 1.05M | — | Current 2026-09-02 |

Pricing notes: long-context tiers, caching, batch

Muse Spark 1.3

The model behind Meta's Muse agent, sold through the Meta Model API (https://api.meta.ai/v1, OpenAI- and Anthropic-compatible) and OpenRouter (meta/muse-spark-1.3). Cached input $0.15/MTok. The muse-spark-1.3-contributor variant is $0.10/$0.20 but lets Meta train on your prompts and completions — the tier is chosen by the model ID alone. Built into Hermes Agent as provider meta-ai since v0.21.0.

## TypeSafe AI

Rates last checked 2026-09-24 against [typesafe.ai](https://typesafe.ai/blog/introducing-system-one-models-and-jev).

| Model | Input / MTok | Output / MTok | Context | Max output | Status |
| --- | --- | --- | --- | --- | --- |
| **Jev** `jev-latest` | $0.042 | $0 | 32K | — | Current 2026-09-15 |

Pricing notes: long-context tiers, caching, batch

Jev

$42 per billion input tokens ($0.042/MTok); output tokens are free (TypeSafe calls them 'too cheap to meter'). 32K input context. Early access via console.typesafe.ai, or without a waitlist through OpenRouter as typesafe/jev-1.13 (~typesafe/jev-latest). End-to-end response 70-500ms. Pin the version: the `jev-latest` alias resolved to Jev 1.13.0 when we checked. See /jev/ for setup and where it fits.

## Open weights

Rates last checked 2026-09-24 against [openrouter.ai](https://openrouter.ai/models).

| Model | Input / MTok | Output / MTok | Context | Max output | Status |
| --- | --- | --- | --- | --- | --- |
| **DeepSeek V4.1 Flash** `deepseek-flash` | $0.30 | $1.20 | 1.05M | — | Current 2026-09-10 |
| **Kimi K3** `moonshotai/kimi-k3` | $3.00 | $15.00 | 1.05M | — | Current |
| **GLM-5.3** `z-ai/glm-5.3` | $0.84 | $2.64 | 1.31M | — | Current 2026-08-18 |
| **GLM-5.3 Flash** `z-ai/glm-5.3-flash` | $0.15 | $0.50 | 1.31M | — | Current 2026-08-26 |
| **Qwen 3.8 2.4T-A95B** `qwen/qwen3.8-2.4t-a95b` | $2.00 | $6.00 | 1.05M | — | Current 2026-08-08 |
| **Qwen 3.8 27B** `qwen/qwen3.8-27b` | $0.42 | $3.00 | 1M | — | Current 2026-08-05 |
| **MiMo-V2.6 Pro** `xiaomi/mimo-v2.6-pro` | $0.435 | $0.87 | 1.05M | — | Current 2026-09-21 |
| **Kimi K2** | $0.60 | $1.80 | — | — | Older |
| **Qwen 3.5 72B** | $0.40 | $1.20 | — | — | Older |
| **Qwen 3.6 35B MoE** | $0 | $0 | — | — | Local |

Pricing notes: long-context tiers, caching, batch

DeepSeek V4.1 Flash

Peak rate shown (01:00–04:00 and 06:00–10:00 UTC, Mon–Fri). Off-peak is exactly half: $0.15/$0.60, and cache hits are $0.006 peak / $0.003 off-peak per MTok. Replaced V4 Flash on 2026-09-10: the old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work but route to and bill as V4.1 Flash. Verified 2026-09-19 against api-docs.deepseek.com/quick_start/pricing. (id kept from the V4 Flash entry so saved calculator states survive.)

Kimi K3

Premium open-weight model, priced like a mid-tier frontier model. Weights on Hugging Face (moonshotai/Kimi-K3) under a custom licence. Cached input $0.30/MTok on OpenRouter.

GLM-5.3

The largest open-weight model from Z.ai. Weights on Hugging Face (zai-org/GLM-5.3) under a custom licence — read it before commercial use. Cached input $0.156/MTok on OpenRouter.

GLM-5.3 Flash

MIT licence, so it is the GLM to self-host or fine-tune commercially without reading a custom licence. Cached input $0.05/MTok on OpenRouter.

Qwen 3.8 2.4T-A95B

Weights public under the Qwen licence (custom). Not the same model as Qwen 3.8 Max, which is API-only. Cached input $0.25/MTok on OpenRouter.

Qwen 3.8 27B

Apache-2.0 — the most permissively licensed model in this group, and small enough to self-host. OpenRouter also offers a rate-limited free tier (qwen/qwen3.8-27b:free, 262K context).

MiMo-V2.6 Pro

Released 2026-09-21 with MIT-licensed weights (XiaomiMiMo/MiMo-V2.6-Pro-RL). Very cheap cache reads: $0.004/MTok on OpenRouter. The Flash sibling is $0.14 / $0.28. New enough that independent agent benchmarks are still thin — test it on your own workload before routing volume to it.

## Keeping this current

This page is generated from the same rate sheet as the cost calculator, which our release monitor updates when a provider ships or reprices a model. Price changes also appear in the [changelog](https://openclawdatabase.com/changelog/). If a rate here differs from the vendor's pricing page, the vendor's page is right — tell us through the [Support page](https://openclawdatabase.com/contribute/).