Best LLM APIs

The Claude API gives you the best model on the market: Claude Opus 5.5, #1 on the Artificial Analysis Intelligence Index at 58. The lineup is simple. Haiku 4.5 costs $1 / $5 per million tokens, Sonnet 5 costs $2 / $10, Opus 5.5 costs $4 / $20 and Fable 5.1 costs $10 / $50. Every current model except Haiku has a 1M-token context and 128K output.

The developer features are strong: prompt caching (reads cost as little as 2.5% of the input price on Fable), a 50% Batch discount, an effort setting to trade quality for cost, and the same model IDs on AWS, Google Cloud and Microsoft Foundry. A US-only inference option costs 10% extra.

The weak spot is the budget end. Haiku 4.5 is almost a year old, so there is no cheap Claude to rival GPT-6 Luna or Gemini 2.5 Flash-Lite. Pick it if quality on coding and agent tasks matters most. Skip it if you need the lowest cost per token at huge volume.

The OpenAI API has the best spread of price points. GPT-6 Luna costs $0.10 / $0.50 per million tokens, GPT-6 Sol costs $2 / $10, and GPT-6 Astra costs $10 / $50. All three share a context of about 1M tokens. On the Artificial Analysis index, Astra scores 53 and Sol 48, behind Claude Opus 5.5 but ahead of everything from Google.

It also has the broadest platform: Batch at half price, 90% off cached input, a Fast mode for Astra, and widely used official SDKs. Watch the long-context surcharge. Prompts over the short-context limit roughly double the input price.

Pick it if you want one vendor that covers everything from cheap classification to frontier reasoning. Skip it if you only need the single best model for coding agents, where Claude Opus 5.5 leads at less than half Astra's price.

The Gemini API has the most generous free tier of any major provider, and its paid Flash models are cheap and fast. Gemini 3.8 Flash costs $0.75 in and $3.75 out per million tokens, runs at about 283 tokens per second in Artificial Analysis tests and takes text, images, audio and video. Gemini 2.5 Flash-Lite is the cheapest option at $0.10 / $0.40.

The big gap is the top end. Gemini 3.5 Pro still has not shipped, so Google's best model scores 41 on the Artificial Analysis index, well behind Claude and GPT-6. Plan for price rises too: 3.8 Flash doubles to $1.50 / $7.50 on 1 January 2027.

Pick it if you want to prototype for free, need cheap multimodal input or use Google Cloud. Skip it if you need frontier reasoning or coding quality.

OpenRouter is not a model maker. It is a router: one API key and one OpenAI-style endpoint for 455 models from Anthropic, OpenAI, Google, Meta, SpaceXAI (formerly xAI), DeepSeek, Z.ai, Moonshot and dozens of hosts. It usually passes through the vendor's list price and charges a 5.5% fee when you buy credits by card.

That makes it the easiest way to compare models on your own prompts, switch when a new one launches, or fall back to another provider during an outage. New models appear fast: Claude Opus 5.5, GPT-6 Sol and Grok 4.7 were all listed within a day of launch. It is also an easy way to reach open models such as GLM-5.3 and Kimi K3 without running servers.

Pick it if you want flexibility, quick testing or access to open models without running servers. Skip it if you need a direct contract, strict data-residency terms or the lowest possible cost at very high volume.

SpaceXAI, the company formerly called xAI, sells Grok models through its API at low prices. The new Grok 4.7 costs $2 in and $6 out per million tokens for prompts under 200K tokens, and scores 46 on the Artificial Analysis index, close to GPT-6 Sol at a lower output price. SpaceXAI reports 71.0% on DeepSWE v1.1.

Speed is a plus: Artificial Analysis measured Grok 4.7 at about 188 tokens per second. The trade-offs are token use and context. It can use many reasoning tokens on complex tasks, which eats into the low price, and it has no Batch API. Its 500K context is half of what Claude, OpenAI and Google offer, although the older Grok 4.3 has 1M tokens at $1.25 / $2.50.

Pick it if you want a low-cost model for long coding sessions. Skip it if you need 1M-token context on the flagship, batch discounts or a long enterprise track record.

Mistral is the main European option. Its API prices are low across the range: Ministral 3 models from $0.10 per million tokens, Mistral Small 4 at $0.15 / $0.60, and Mistral Large 3 at $0.50 / $1.50. Cached input is 90% off and Batch is half price. Several models, including Mistral Small 4, are also released as open weights, so you can move them in-house later.

The weakness is quality at the top. Mistral Medium 3.5 scores 14 on the Artificial Analysis index, far behind US and Chinese flagships, so Mistral is not the choice for hard reasoning or agentic coding.

Pick it if you want a European vendor, cheap models for routine tasks, or an easy path from API to self-hosting. Skip it if you need frontier-level answers.

Meta's Model API sells its closed Muse Spark models. Muse Spark 1.3 scores 48 on the Artificial Analysis index, level with GPT-6 Sol, and runs at about 213 tokens per second, one of the fastest results for a model this capable. The standard price is $1.25 in and $4.25 out per million tokens.

Meta also offers a "Contributor" tier at just $0.10 / $0.20. The catch: Meta can use that traffic to improve its products. That is fine for public data or testing, not for customer or company secrets.

The platform is young. It has fewer SDKs, integrations and enterprise controls than OpenAI, Anthropic or Google. Pick it if you need fast, cheap, near-frontier answers and can choose the right tier for your data. Skip it if you need mature enterprise tooling or strict data guarantees.

DeepSeek's API is the cheapest serious option. DeepSeek-Flash, which runs DeepSeek V4.1 Flash, costs $0.30 in and $1.20 out per million tokens at peak and half that off-peak. Cached input can cost as little as $0.003. It supports a 1M-token context, up to 384K tokens of output, images, JSON output and tool calls, and it uses an OpenAI-compatible format.

Quality is solid, not frontier: V4.1 Flash scores 39 on the Artificial Analysis index. The bigger worry for many firms is compliance, because the service is run from China. Because the weights are MIT-licensed, you can get the same model from US hosts or run it yourself.

Pick it if you run huge volumes of routine work such as extraction, tagging or first drafts. Skip it if your data rules forbid China-based processors; use the open weights through another host instead.