🆕 Updated July 30, 2026: GPT-5.6 Luna drops to $0.20/$1.20 (-80%) and Terra to $2/$12 (-20%). Luna is now the cheapest capable model by a huge margin. See GPT-5.6 price drop analysis.
July 17, 2026: Kimi K3 at $3/$15 (with 90% cache hits effectively ~$0.30 input) and Meta Muse Spark 1.1 at $1.25/$4.25 join the mid-tier war. K3 hits 88.3% Terminal-Bench (near Sol). See K3 pricing breakdown and Muse Spark 1.1 guide.
AI API pricing has dropped 85% since GPT-4 launched in 2023. Frontier model input now costs under $3/1M tokens. But prices vary wildly between providers, and the cheapest model isn’t always the cheapest for your use case. For the full set of pricing breakdowns and cost-optimization guides, see our APIs & pricing hub.
Here’s every major provider’s pricing in one place.
Frontier models
Meta Muse Wonder API, checked October 13
Meta’s muse-wonder-1.0 is a public Preview API, priced at $4 input / $0.20 cached input / $20 output per million tokens. The context is 1,048,576 tokens; limits are 3,000 RPM and 4 million TPM per team. Meta lists no long-context premium. Text, images, video, audio and PDFs are inputs; output is text. Search grounding carries a separate fee. Official rates and limits.
Wonder Full is not an API pricing tier. It is available through eligible Muse Code subscriptions, without a public pay-as-you-go endpoint or token rate. Selecting maximum reasoning on regular Wonder does not select Full. For integration and access checks, use our Muse Wonder API integration guide. The APIpulse Pricing Index owns the maintained comparison dataset; the Muse Code guide covers the terminal product.
🆕 July 27, 2026: Claude Opus 5 launched at $5/$25 — same price as Opus 4.8, double the performance, within 0.5% of Fable 5 on coding.
June 10, 2026: Claude Fable 5 launched — Anthropic’s most powerful model ever at $10/$50 per million tokens. It’s a new premium tier above Opus, scoring 95% on SWE-bench Verified. See our full pricing analysis.
| Model | Input/1M | Output/1M | Context | Provider |
|---|---|---|---|---|
| Claude Fable 5 🆕 | $10.00 | $50.00 | 1M | Anthropic |
| GPT-5.4 | $2.50 | $15.00 | 1M | OpenAI |
| GPT-5.4 Pro | $30.00 | $180.00 | 1M | OpenAI |
| Claude Opus 5 | $5.00 | $25.00 | 1M | Anthropic |
| Claude Opus 4.8 | $5.00 | $25.00 | 1M | Anthropic |
| Claude Opus 4.6 | $15.00 | $75.00 | 1M | Anthropic |
| Claude Sonnet 4 | $3.00 | $15.00 | 200K | Anthropic |
| Grok 4.7 | $2.00 below 200K input | $6.00 below 200K input | 500K | xAI |
| Gemini 3.6 Flash 🆕 | $1.50 | $7.50 | 1M | |
| Gemini 3.5 Flash | $1.50 | $9.00 | 1M | |
| Gemini 3.1 Pro | $2.00 | $12.00 | 1M+ | |
| Gemini 2.5 Pro | $1.25 | $5.00 | 1M |
Mid-tier models (best value)
| Model | Input/1M | Output/1M | Context | Provider |
|---|---|---|---|---|
| GPT-5.6 Terra 🆕 | $2.00 | $12.00 | 1M | OpenAI |
| GPT-5.4 Mini | $0.75 | $4.50 | 400K | OpenAI |
| Gemini 3.5 Flash-Lite 🆕 | $0.30 | $2.50 | 1M | |
| Claude Haiku 4 | $0.25 | $1.25 | 200K | Anthropic |
| Gemini 2.5 Flash | $0.15 | $0.60 | 1M | |
| Mistral Large 2 | $2.00 | $6.00 | 128K | Mistral |
| Mistral Small | $0.10 | $0.30 | 32K | Mistral |
Budget models (cheapest)
| Model | Input/1M | Output/1M | Context | Provider |
|---|---|---|---|---|
| GPT-5.6 Luna 🆕 | $0.20 | $1.20 | 1M | OpenAI |
| DeepSeek V4-Flash | $0.14 | $0.28 | 1M | DeepSeek |
| DeepSeek V4-Pro | $1.74 | $3.48 | 1M | DeepSeek |
| DeepSeek Chat | $0.14 | $0.28 | 128K | DeepSeek |
| DeepSeek Reasoner | $0.55 | $2.19 | 128K | DeepSeek |
| GPT-5 Mini | $0.25 | $2.00 | 128K | OpenAI |
| Qwen 3.6 Plus | Free tier | Free tier | 1M | Alibaba |
| Qwen 3.6 Flash | $0.25 | $1.00 | 1M | Alibaba |
| Qwen 3.6 Max Preview | $1.50 | $6.00 | 1M | Alibaba |
| GLM-5.1 | Free tier | Free tier | 128K | Z.ai |
Open-source via hosting providers
New (April 24, 2026): DeepSeek V4-Flash at $0.14/$0.28 per 1M tokens is now the cheapest frontier-class model available — scoring 79.0% on SWE-bench with 1M context. V4-Pro at $1.74/$3.48 scores 80.6%. Both MIT licensed. See V4 Flash: cheapest frontier model.
| Model | Provider | Input/1M | Output/1M |
|---|---|---|---|
| Llama 4 Maverick | Together AI | ~$0.49 | ~$0.49 |
| Llama 4 Scout | Together AI | ~$0.10 | ~$0.10 |
| Qwen 3.5 27B | Fireworks | ~$0.20 | ~$0.20 |
| Any model | Ollama (local) | $0 | $0 |
| Any model | OpenRouter | Varies | Varies |
Tired of API costs? Self-hosting can be cheaper long-term. See how to self-host an LLM for €4.99/month or deploy on Vultr with $250 free credits.
Cost per common task
What does a typical developer task actually cost?
| Task | Tokens (approx) | GPT-5.4 | GPT-5.4 Mini | DeepSeek | Local |
|---|---|---|---|---|---|
| Code review (1 file) | 3K in / 1K out | $0.02 | $0.007 | $0.001 | $0 |
| Bug fix | 5K in / 2K out | $0.04 | $0.01 | $0.002 | $0 |
| Full feature | 10K in / 5K out | $0.10 | $0.03 | $0.005 | $0 |
| Codebase analysis | 50K in / 5K out | $0.20 | $0.06 | $0.01 | $0 |
| 8-hour coding session | 500K in / 100K out | $2.75 | $0.83 | $0.10 | $0 |
Monthly cost estimates
| Usage level | GPT-5.4 | GPT-5.4 Mini | DeepSeek | Local |
|---|---|---|---|---|
| Light (10 tasks/day) | $6-12 | $2-4 | $0.30-0.60 | $0 |
| Medium (50 tasks/day) | $30-60 | $10-20 | $1.50-3 | $0 |
| Heavy (200 tasks/day) | $120-240 | $40-80 | $6-12 | $0 |
The smart approach: model routing
Don’t use one model for everything. Route by task complexity:
async def route(message, complexity):
if complexity == "simple":
return await call("deepseek-chat", message) # $0.001
elif complexity == "medium":
return await call("gpt-5.4-mini", message) # $0.01
else:
return await call("claude-sonnet-4", message) # $0.04This gives you frontier quality when you need it and near-zero cost when you don’t. See our model routing guide for implementation details.
Batch API pricing: the discount most budgets miss
Three of the four major providers sell a second, cheaper lane for work that doesn’t need a real-time response. It’s easy to miss because it’s not on the main pricing table — you have to opt into a separate batch endpoint.
| Provider | Batch discount | Notes |
|---|---|---|
| OpenAI | 50% off sync rate | Fixed 24-hour completion window, often faster |
| Anthropic | 50% off sync rate | ≤24h window, most batches finish in under an hour |
| Google (Gemini) | 50% off sync rate | Stacks with introductory pricing on Gemini 3.6/3.7 Flash |
| xAI (Grok) | 20% off, 4 older models only | Flagship Grok 4.6 has no batch discount at all |
Verified directly against each vendor’s own pricing documentation: Google’s Gemini API pricing page states “Batch API (50% cost reduction),” OpenAI’s Batch API guide states a “50% cost discount compared to synchronous APIs,” and Anthropic’s Message Batches documentation prices all batch usage at 50% of standard rates. For Claude Sonnet 5 specifically, that puts the batch rate at $1/$5 per million input/output tokens against the $2/$10 standard rate.
How batch differs from standard API pricing: same model, same output quality — you give up streaming and a real-time response in exchange for the discount. OpenAI’s batch jobs complete within a fixed 24-hour window; Anthropic’s typically finish in under an hour but can take up to 24. Anthropic’s batch discount also stacks with prompt caching, so a high-cache-hit batch workload pays 50% off an already-discounted cached-read rate.
When to use batch processing: anything without a human waiting on the response. Bulk classification, tagging, embeddings, nightly eval runs, and scheduled report generation are the clearest fits — paying the synchronous rate for this kind of work is a standing 2x overspend. Interactive chat, coding-agent tool loops, and anything using streaming should stay on the synchronous API; batch requests explicitly reject stream: true.
Practical cost-saving implication: if you’re running any kind of bulk or offline job — dataset labeling, log analysis, content generation queues — check whether your provider’s batch endpoint applies before assuming the standard per-token rate is your floor. On a large enough job, that’s the difference between a $200 bill and a $100 one for identical output.
| Provider | What’s free | Limitation |
|---|---|---|
| Gemini CLI | Full CLI access | Rate limited |
| Qwen 3.6 Plus | API access | Rate limited |
| GLM-5.1 | Z.ai free tier | Quota limited |
| OpenRouter | Several free models | Model-dependent |
| Ollama | Unlimited local | Hardware-dependent |
See our free AI coding tier review for real-world testing of each.
Price trends
Input token costs have dropped ~85% since 2023. The trend continues as competition intensifies between OpenAI, Anthropic, Google, and Chinese providers. Expect another 30-50% drop by end of 2026.
The biggest driver: open-source models (GLM-5.1, Qwen 3.6, Llama 4) matching proprietary quality at zero cost, forcing paid providers to lower prices.
Related: AI Coding Tools Pricing · FinOps for AI · AI Agent Cost Management · Monitor AI API Spending · OpenRouter Complete Guide · Cheapest AI Coding Setup · Tested Every Free AI Tier
-
Cloud Hosting Pricing Compared — Railway vs Cloudways vs Hetzner vs Vultr vs RunPod (2026)
-
Qwen 3.7 Flash vs Gemini 3.6 Flash: $0.03 vs $1.50 Per Million Tokens
-
Best AI APIs for Startups — Free Tiers and Pricing Compared (2026)
-
Grok 4.5 Pricing and Cursor Integration: What Developers Need to Know
For real-time pricing updates and a cost calculator, see APIpulse which tracks 90+ models across all providers.