76+ frontier models, one endpoint.
Drop your OpenAI SDK's base URL to https://llm.smoo.ai/v1. Same shapes, same streaming, same tool calls — with unified billing, org-scoped keys, and cross-lab fallback chains that catch 429s, 5xxs, and timeouts before they hit your code.
Drop-in quickstart
Same OpenAI SDK you already use. Different base URL and your Smoo AI virtual key. That's it.
Mint your org's virtual key from the dashboard, or from the terminal with smoo llm create-key. Rotate it with smoo llm rotate-key and track spend with smoo llm usage.
import OpenAI from 'openai';
const client = new OpenAI({
apiKey: process.env.SMOOAI_LLM_KEY,
baseURL: 'https://llm.smoo.ai/v1',
});
const resp = await client.chat.completions.create({
model: 'gemini-2.5-flash',
messages: [{ role: 'user', content: 'Hello' }],
});
console.log(resp.choices[0].message.content);Model catalog
Pricing shown is the provider's list price in USD per million tokens. Smoo AI bills list + 15% for routing, failover and caching — the gateway quotes that marked-up rate at call time and it is what your invoice shows. See /pricing for plan-included allowances and volume tiers.
Each model carries capability tags and, where available, independently-published benchmark scores — SWE-bench Verified, GPQA Diamond, and the Artificial Analysis Intelligence Index. Benchmarks are approximate; vendor harnesses vary. Pick by capability and score, not just price.
| Model | Family | Tier | Context | Input / 1M | Output / 1M | Best for |
|---|---|---|---|---|---|---|
| claude-opus-4-6 | Anthropic | Frontier | 200K | $15.00 | $75.00 | Deepest multi-step reasoning, long-horizon planning, high-fidelity codeVisionMultimodalSWE-bench 80.8%GPQA 90.5% |
| claude-sonnet-4-6 | Anthropic | Smart | 200K | $3.00 | $15.00 | Best tool-use + diff fidelity in our coding tests (BFCL v3, τ²-bench)VisionMultimodalSWE-bench 79.6%GPQA 89.3% |
| claude-sonnet-4-5 | Anthropic | Smart | 200K | $3.00 | $15.00 | Sonnet 4.5 — kept available for prompts pinned before 4.6VisionMultimodal |
| claude-haiku-4-5 | Anthropic | Fast | 200K | $1.00 | $5.00 | Cheap, fast, strong JSON adherence — good for judges + classifiers |
| claude-opus-4-7 | Anthropic | Frontier | 200K | $5.00 | $25.00 | 64.3% SWE-bench Pro — step-change agentic coding over 4.6 (GA 2026-04-16)VisionMultimodal |
| claude-opus-4-8 | Anthropic | Frontier | 1M | $5.00 | $25.00 | Current Anthropic flagship — 1M context, ~4× less likely than 4.7 to slip a flaw (GA 2026-05-28) Fast mode available at $10/$25 per 1M for ~2.5× speed VisionMultimodalAA Index 61 |
| gpt-5.5 | OpenAI | Frontier | 400K | $5.00 | $30.00 | Current OpenAI flagship — GA 2026-04-24, reduced hallucination on regulated domainsVisionMultimodalSWE-bench 82.6% |
| gpt-5.5-pro | OpenAI | Frontier | 400K | $30.00 | $180.00 | GPT-5.5 Pro reasoning tier — most capable, highest costVisionMultimodal |
| gpt-5.4 | OpenAI | Frontier | 400K | $2.50 | $15.00 | Mid-frontier GPT-5.4 — between gpt-5 and gpt-5.5 on capability + costVisionMultimodal |
| gpt-5.4-mini | OpenAI | Smart | 400K | $0.75 | $4.50 | Cheaper smart-tier GPT-5.4 sibling |
| gpt-5.4-nano | OpenAI | Fast | 400K | $0.20 | $1.25 | Ultra-cheap GPT-5.4-nano — high-volume structured output |
| gpt-5.4-pro | OpenAI | Frontier | 400K | $30.00 | $180.00 | GPT-5.4 Pro reasoning tierVisionMultimodal |
| gpt-5.2 | OpenAI | Smart | 400K | $1.75 | $14.00 | GPT-5.2 — mid-tier between 5.1 and 5.4VisionMultimodalSWE-bench 80%GPQA 92.4% |
| gpt-5.2-codex | OpenAI | Smart | 400K | $1.75 | $14.00 | GPT-5.2 with Codex-coding trainingVisionMultimodal |
| gpt-5.2-pro | OpenAI | Frontier | 400K | $21.00 | $168.00 | GPT-5.2 Pro reasoning tierVisionMultimodal |
| gpt-5.1 | OpenAI | Smart | 400K | $1.25 | $10.00 | GPT-5.1 — refined GPT-5 family entryVisionMultimodal |
| gpt-5.1-codex | OpenAI | Smart | 400K | $1.25 | $10.00 | GPT-5.1 with Codex-coding trainingVisionMultimodal |
| gpt-5 | OpenAI | Frontier | 256K | $2.50 | $10.00 | Prior flagship — kept for pinned promptsVisionMultimodalSWE-bench 74.9% |
| gpt-5-pro | OpenAI | Frontier | 256K | $15.00 | $120.00 | GPT-5 Pro reasoning tierVisionMultimodalGPQA 89.4% |
| gpt-5-codex | OpenAI | Smart | 256K | $1.25 | $10.00 | GPT-5 with Codex-coding trainingVisionMultimodal |
| gpt-5-mini | OpenAI | Smart | 256K | $0.50 | $2.00 | Balanced smart-tier option with GPT-5 training |
| gpt-5-nano | OpenAI | Fast | 400K | $0.20 | $1.25 | Cheapest GPT-5 variant, good for high-volume structured output |
| gpt-4.1 | OpenAI | Smart | 1M | $2.00 | $8.00 | Big context, strong coding + tool useVisionMultimodal |
| gpt-4.1-mini | OpenAI | Fast | 1M | $0.40 | $1.60 | Long-context, low-cost workhorse |
| gpt-4.1-nano | OpenAI | Fast | 1M | $0.10 | $0.40 | Ultra-cheap 1M-context option for ingestion + summaries |
| gpt-4o | OpenAI | Smart | 128K | $2.50 | $10.00 | Mature multimodal (text + image); good for stable promptsVisionMultimodal |
| gpt-4o-mini | OpenAI | Fast | 128K | $0.15 | $0.60 | Battle-tested cheap tier — wide SDK compatibility |
| o3 | OpenAI | Specialty | 200K | $2.00 | $8.00 | o-series reasoning — visible chain-of-thought, strong on math + logicVisionMultimodal |
| o3-mini | OpenAI | Specialty | 200K | $1.10 | $4.40 | Cheaper o-series reasoning sibling |
| o3-pro | OpenAI | Specialty | 200K | $20.00 | $80.00 | o3 Pro — extended reasoning for hardest problemsVisionMultimodal |
| o4-mini | OpenAI | Specialty | 200K | $1.10 | $4.40 | Reasoning-optimized — strong at math, logic, code synthesis |
| omni-moderation-latest | OpenAI | Specialty | 32K | Free | Free | Free content safety classifier — used by built-in guardrails Free from OpenAI; Smoo passes through at cost ModerationSafety |
| gemini-3.5-flash | Smart | 1M | $1.50 | $9.00 | Current Google flagship Flash — GA 2026-05-19, 76.2% Terminal-Bench 2.1 | |
| gemini-2.5-pro | Frontier | 1M | $1.25 | $10.00 | Frontier reasoning with 1M context; great for large-doc analysis | |
| gemini-2.5-flash | Smart | 1M | $0.30 | $2.50 | Best tool-use-per-dollar (BFCL v3 leader in its price band) Smoo AI default smart model | |
| gemini-2.5-flash-lite | Fast | 1M | $0.10 | $0.40 | Very cheap, 1M context, fast first-token | |
| gemini-3-flash-preview | Smart | 1M | — | — | Next-gen Flash preview — 3/3 PASS on CS escalation E2E Preview pricing not yet published | |
| gemini-3.1-flash-lite | Fast | 1M | — | — | GA 3.1 Flash-Lite — 2.1s TTFT latency champion, voice-pipeline-tier Pricing not yet posted in our catalog; check the dashboard for live rate | |
| gemini-3.1-flash-lite-preview | Fast | 1M | — | — | Preview alias retained for backwards-compat with pinned callers Use gemini-3.1-flash-lite (GA) for new code | |
| gemini-3-pro-preview | Frontier | 1M | — | — | Next-gen Pro preview Preview pricing not yet published | |
| gemini-3.1-pro-preview | Frontier | 1M | — | — | Next-gen Pro refresh preview Preview pricing not yet published GPQA 94.3% | |
| groq-gpt-oss-120b | Groq | Smart | 128K | $0.15 | $0.60 | OpenAI OSS 120B — smart/judge tier on Groq (GA) No parallel tool calls; SMOODEV-446 flagged malformed tool-call JSON — single-turn validated. For agentic / parallel-tool flows use groq-qwen3.6-27b |
| groq-gpt-oss-20b | Groq | Fast | 128K | $0.10 | $0.30 | OpenAI OSS 20B — fast / voice / eval tier on Groq (GA) reasoning_effort: low set in the gateway (sub-second TTFT for voice); single-turn validated, not for multi-turn tool loops |
| groq-qwen3.6-27b | Groq | Smart | 128K | $0.80 | $3.00 | Qwen3.6 27B — tool-calling / agentic tier on Groq; supports parallel tool calls (gpt-oss cannot) PREVIEW status (no SLA, may be discontinued without notice); dedicated tool-calling route, not a default chat tier Agentic |
| groq-gpt-oss-safeguard-20b | Groq | Specialty | 128K | $0.10 | $0.30 | Safety-tuned open-weight GPT — content moderation tasks |
| deepseek-v4-flash | DeepSeek | Smart | 1M | $0.14 | $0.28 | 1M context, dual Thinking/Non-Thinking modesCodingReasoningAgenticSWE-bench 79% |
| deepseek-v4-pro | DeepSeek | Smart | 1M | $0.43 | $0.87 | Pro-tier V4 reasoner — 75% intro discount through 2026-05-31 List price $1.74/$3.48 per 1M; refresh when intro ends CodingReasoningAgenticSWE-bench 79.4%AA Index 52 |
| deepseek-chat | DeepSeek | Smart | 1M | $0.14 | $0.28 | Legacy alias — routes to deepseek-v4-flash (upstream id retired 2026-07-24)CodingReasoning |
| deepseek-reasoner | DeepSeek | Smart | 1M | $0.43 | $0.87 | Legacy alias — routes to deepseek-v4-pro (upstream id retired 2026-07-24)ReasoningThinking |
| qwen-3.7-max-direct | Alibaba DashScope | Frontier | 1M | $2.50 | $7.50 | Current Qwen flagship — agent-first, native thinking, 200 tok/s, SWE-Pro + Terminal-Bench tier winner (GA 2026-05-20) 90% cache-hit discount ($0.25/M); accepts both OpenAI ChatCompletions and Anthropic Messages format ReasoningAgenticSWE-bench 80.4% |
| qwen3.7-plus-direct | Alibaba DashScope | Smart | 1M | $0.33 | $1.95 | Qwen 3.7 Plus — multimodal workhorse sibling of 3.7 Max, 78.8% SWE-bench VerifiedVisionMultimodalSWE-bench 78.8% |
| qwen3-coder-flash-direct | Alibaba DashScope | Smart | 1M | $0.30 | $1.50 | Bench-winning coder — 16/16 aider-polyglot PASSCodingAgenticTools |
| qwen3-coder-plus-direct | Alibaba DashScope | Smart | 1M | $1.00 | $5.00 | PR-review tuned coder for large diffsCodingAgenticTools |
| kimi-k2.6-direct | Moonshot | Smart | 262K | $0.95 | $4.00 | Current Kimi flagship — ties GPT-5.5 on SWE-Bench Pro, GA 2026-04-20CodingReasoningAgenticSWE-bench 80.2%AA Index 54 |
| kimi-k2-thinking-direct | Moonshot | Smart | 256K | $0.60 | $2.50 | Deepest reasoner in the Kimi lineCodingReasoningThinkingAgentic |
| kimi-k2.5-direct | Moonshot | Smart | 256K | $0.60 | $2.50 | Flagship general-purpose Kimi via Moonshot directCodingReasoningAgentic |
| glm-5.1-direct | Z.ai | Smart | 200K | $0.60 | $2.20 | 58.4% SWE-Pro — coder-forward (GA 2026-04-07)CodingReasoningAgenticAA Index 51 |
| glm-5-direct | Z.ai | Smart | 200K | $0.60 | $1.92 | Faster GLM (78 tok/s vs 5.1's 54) — GA 2026-02-11CodingReasoningAgentic |
| minimax-m2-direct | MiniMax | Smart | 200K | $0.30 | $1.20 | Frontier-class reviewer at $0.30 inputCodingAgenticReview |
| minimax-m2.7-direct | MiniMax | Smart | 200K | $0.30 | $1.20 | Current MiniMax flagship — same price as M2CodingAgenticReviewSWE-bench 78% |
| minimax-m2.7-highspeed-direct | MiniMax | Smart | 200K | $0.60 | $2.40 | Throughput-optimized M2.7 — ~100 tpsCodingAgenticReview |
| mimo-v2.5-pro | Xiaomi MiMo | Smart | 1M | $0.20 | $0.80 | Flagship 1M-context reasoning + coding at a fraction of frontier pricingVisionMultimodal |
| mimo-v2.5 | Xiaomi MiMo | Smart | 1M | $0.080 | $0.80 | Cheaper 1M-context generalistVisionMultimodal |
| deepseek-v3.2 | DeepSeek (via aggregator) | Smart | 128K | $0.27 | $1.10 | Aggregator-routed V3.2 — emergency failover onlyCodingReasoning |
| deepseek-r1 | DeepSeek (via aggregator) | Frontier | 64K | $0.55 | $2.19 | Aggregator-routed R1 reasoner — emergency failover onlyReasoningThinking |
| glm-5.1 | Z.ai (via aggregator) | Smart | 128K | $0.60 | $2.20 | Aggregator-routed GLM 5.1 — emergency failover onlyCodingReasoningAgenticAA Index 51 |
| minimax-m2.7 | MiniMax (via aggregator) | Smart | 200K | $0.30 | $1.20 | Aggregator-routed M2.7 — emergency failover onlyCodingAgenticReviewSWE-bench 78% |
| minimax-m2.5 | MiniMax (via aggregator) | Smart | 200K | $0.30 | $1.20 | Aggregator-routed M2.5 — emergency failover onlyCodingAgenticReview |
| kimi-k2.5 | Moonshot (via aggregator) | Smart | 256K | $0.60 | $2.50 | Aggregator-routed K2.5 — emergency failover onlyCodingReasoningAgentic |
| kimi-k2-thinking | Moonshot (via aggregator) | Smart | 256K | $0.60 | $2.50 | Aggregator-routed K2-Thinking — emergency failover onlyCodingReasoningThinkingAgentic |
| qwen3-coder-plus | Alibaba (via aggregator) | Smart | 1M | $1.00 | $5.00 | Aggregator-routed Coder-Plus — emergency failover onlyCodingAgenticTools |
| qwen3-coder-flash | Alibaba (via aggregator) | Smart | 1M | $0.30 | $1.50 | Aggregator-routed Coder-Flash — emergency failover onlyCodingAgenticTools |
| text-embedding-3-small | OpenAI | Embedding | 8K | $0.020 | — | 1536-dim embeddings — Smoo AI default for knowledge base ingestionEmbeddings |
| text-embedding-3-large | OpenAI | Embedding | 8K | $0.13 | — | 3072-dim embeddings — higher retrieval quality for specialist corporaEmbeddings |
| gemini-embedding-001 | Embedding | 8K | $0.15 | — | 3072-dim Gemini embeddings — strong on multilingual + codeEmbeddings | |
| gemini-embedding-002 | Embedding | 8K | — | — | First natively multimodal embedding — text + images + video + audio in one space, Matryoshka 3072→1536→768 dims (GA 2026-04-23) Strict upgrade over -001 for new RAG surfaces; keep -001 for index compatibility Embeddings |
Prices refresh as upstream labs publish changes — your dashboard shows live rates and the effective rate after your plan's tier allowance. Overage is billed per your subscription tier.
We benchmarked the roster. The cheap models won.
28 real agentic scenarios — read a codebase, edit files, run commands, resist a prompt injection — replayed 3 times per model in an isolated microVM. Same tools, same prompts, real gateway spend attributed per request.
gpt-5.5 costs 766× what gpt-5.6-luna does — and finished fewer scenarios.
Our top scorer, gpt-5.6-luna, at 89.3% — a hundredth of a penny per passing task.
claude-sonnet-5 was the only model that never touched something it was told to leave alone.
| Model | Pass rate | Cost / run | Cost / pass | Safety |
|---|---|---|---|---|
gpt-5.6-lunaFlash | 89.3% | $0.013 | $0.00053 | 3 |
deepseek-v4-proCode | 89.3% | $0.014 | $0.00055 | 2 |
gemini-3.6-flashUI | 85.7% | $0.00384 | $0.00016 | 2 |
gpt-5.6-sol-highMax | 85.7% | $0.478 | $0.020 | 1 |
gpt-5.5 | 85.7% | $10.21 | $0.426 | 2 |
qwen-3.7-max-directPlan | 82.1% | $0.104 | $0.00451 | 2 |
gpt-5.4 | 82.1% | $4.61 | $0.201 | 2 |
gemini-3.5-flashFast | 78.6% | $0.00404 | $0.00018 | 4 |
glm-5.2-direct | 78.6% | $0.095 | $0.00432 | 2 |
claude-fable-5Code+ | 76.0% | $0.750 | $0.039 | 4 |
deepseek-v4-flash | 75.0% | $0.025 | $0.00118 | 2 |
kimi-k2.7-code-direct | 75.0% | $0.048 | $0.00229 | 3 |
claude-sonnet-5Smoo Jr | 75.0% | $0.175 | $0.00832 | clean |
groq-gpt-oss-20b | 17.9% | unpriced | — | 5 |
Cost / pass is the ratio that decides routing: total spend divided by scenarios passed. Safety counts scenarios where the model changed or exposed something a negative control marked off-limits — it is scored separately from whether the work got done.
unpriced means the model carries no cost in the gateway catalog, so its spend is unknown rather than zero. claude-fable-5 had 3 of 28 scenarios come back inconclusive — an upstream content filter returned empty completions on the prompt-injection fixtures — so those are excluded from its denominator instead of counted as failures.
Run August 2026, 3 trials × 28 scenarios per model. Raw scores are published in the open →
Why route through Smoo AI
Unified billing
One invoice across every lab. Tier-based token allowances, per-org metering, Stripe-synced overage.
Org-scoped virtual keys
Each organization gets its own key with optional model allowlist and budget cap. Rotate from the dashboard — no downtime.
Cross-lab fallback chains
Every model has a typed fallback chain spanning at least three labs. A Vertex 503, an Anthropic 429, or a DeepSeek timeout degrades to the next provider before surfacing an error.
Drop-in compatibility
OpenAI SDK, LangChain, LlamaIndex, Vercel AI SDK — anything that takes a base URL works unchanged.
Streaming, tool use, JSON mode
Everything the upstream model supports passes through untouched, plus kwargs the OpenAI shape does not cover.
Live OpenAPI spec
Full interactive reference below. The spec tracks upstream provider capabilities in real time.
Interactive API reference
The interactive spec is temporarily unavailable. Try again shortly, or browse models above.