Buyer's guide
Best LLM API: costs, performance and limits
Choose DeepInfra for a $0.00054 GPT-OSS 120B request example, Groq for 473 tokens/s on Artificial Analysis (AA), or Novita AI for a $0.00075 example plus 200+ models and Agent Sandbox.
| Need | Pick | Why |
|---|---|---|
| Lowest example cost | DeepInfra | $0.0005412 |
| Fastest measured output | Groq | 473 tokens/s3 |
| Shortest measured initial wait | Together AI | 0.47 s3 |
Performance
Performance
GPT-OSS 120B · AA provider snapshot · high reasoning effort · 10,000 input tokens.
Swipe to see all columns
Capabilities and serving conditions
Capabilities and serving conditions
Swipe to see all columns
| Provider | API and model ID | Serving and capacity |
|---|---|---|
| Novita AI456 | https://api.novita.ai/openai | Novita’s 200+ model catalog shares billing access with multimodal APIs, Agent Sandbox and GPU products; this GPT-OSS 120B endpoint allows up to 32K output. |
| DeepInfra12 | https://api.deepinfra.com/v1/openai | DeepInfra applies a default ceiling of 200 simultaneous requests per model. Its standard GPT-OSS 120B endpoint and Turbo are separate products. |
| Groq78 | https://api.groq.com/openai/v1 | Groq Free applies four GPT-OSS 120B quotas: 30 RPM, 1K RPD, 8K TPM and 200K TPD; Developer-plan quotas are separate. |
| Together AI91011 | https://api.together.xyz/v1 | Alongside a 200+ model catalog, Together offers provisioned-throughput contracts. Its best-effort serverless service can return 429 or 503 during high demand. |
| Fireworks AI1213 | https://api.fireworks.ai/inference/v1 | Fireworks bills cached GPT-OSS 120B input at $0.015 per 1M tokens. Dedicated GPU deployments remain subject to account-level request quotas. |
Pricing
Prices and request cost
Swipe to see all columns
| Provider | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Novita AI456 | $0.05 | $0.25 |
| DeepInfra12 | $0.037 | $0.17 |
| Groq78 | $0.15 | $0.60 |
| Together AI91011 | $0.15 | $0.60 |
| Fireworks AI1213 | $0.15 | $0.60 |
Options, strengths and limitations
Options, strengths and limitations
DeepInfra
A direct API for price-sensitive text generation on the standard endpoint.
- Input / 1M tokens
- $0.037
- Output / 1M tokens
- $0.17
- Estimated request cost
- $0.00054
Strengths
Limitations
- Low token prices come with 36 tokens/s on the standard endpoint. Turbo is a separate offering; the standard $0.037/$0.17 rates do not buy Turbo performance.3
Sources:12 · Novita AI vs DeepInfra
Novita AI
A model API platform for applications combining text with image, video or audio.
- Input / 1M tokens
- $0.05
- Output / 1M tokens
- $0.25
- Estimated request cost
- $0.00075
Strengths
Limitations
- Novita AI caps generated output at 32K, limiting long completions. Its 0.91 s first-chunk median also exceeds Groq’s 0.65 s.3
Groq
A streaming-oriented API for applications that prioritize output speed on supported models.
- Input / 1M tokens
- $0.15
- Output / 1M tokens
- $0.60
- Estimated request cost
- $0.00210
Strengths
Limitations
- Groq’s $0.15/$0.60 per 1M input/output tokens exceeds Novita AI’s $0.05/$0.25. Access is also subject to model-specific and plan-specific limits.
Sources:78 · Novita AI vs Groq
Together AI
An option for latency-sensitive starts, with serverless access and contracted throughput available separately.
- Input / 1M tokens
- $0.15
- Output / 1M tokens
- $0.60
- Estimated request cost
- $0.00210
Strengths
Limitations
- Together AI costs $0.15/$0.60 per 1M input/output tokens, compared with Novita AI’s $0.05/$0.25. Its model-specific and plan-specific controls still apply.
Sources:91011 · Novita AI vs Together AI
Fireworks AI
An API for repeat-prompt workloads that can use cached-input billing.
- Input / 1M tokens
- $0.15
- Output / 1M tokens
- $0.60
- Estimated request cost
- $0.00210
Strengths
Limitations
- Fireworks charges $0.15/$0.60 for synchronous requests, above DeepInfra’s standard rates. Dedicated GPUs do not remove its account-wide request quotas.
Sources:1213 · Novita AI vs Fireworks AI
Methodology
Methodology and sources
The five-provider GPT-OSS 120B shortlist uses direct list prices and an AA endpoint snapshot for the four measured providers. AA settings are high reasoning and 10,000 input tokens, with speed and first-chunk latency shown as 72-hour medians. The first received chunk may contain reasoning; its latency is not the wait for the first answer token. Endpoint Accuracy Index is relative to a reference endpoint, not a task pass rate. Fireworks has price and deployment data only in this snapshot.
Its request chart is an uncached synchronous budget of 10,000 input and 1,000 billed output tokens, including reasoning; it excludes tax, retries, discounts and reserved capacity. Cards are ordered by that cost rather than provider quality, while limits and catalog counts reflect each provider’s scope. Novita AI is a listed provider and the publisher; the page uses public evidence, not a private workload test.
- DeepInfra · GPT-OSS 120B deepinfra.com · Checked 2026-10-15
- DeepInfra · Rate Limits deepinfra.com · Checked 2026-10-15
- Artificial Analysis · GPT-OSS 120B Providers artificialanalysis.ai · Checked 2026-10-15
- Novita AI · GPT-OSS 120B novita.ai · Checked 2026-10-15
- Novita AI · Platform novita.ai · Checked 2026-10-15
- Novita AI · Batch docs.novita.ai · Checked 2026-10-15
- Groq · GPT-OSS 120B console.groq.com · Checked 2026-10-15
- Groq · Rate Limits console.groq.com · Checked 2026-10-15
- Together AI · GPT-OSS 120B www.together.ai · Checked 2026-10-15
- Together AI · Model Library www.together.ai · Checked 2026-10-15
- Together AI · Rate Limits docs.together.ai · Checked 2026-10-15
- Fireworks AI · GPT-OSS 120B fireworks.ai · Checked 2026-10-15
- Fireworks AI · Account Quotas docs.fireworks.ai · Checked 2026-10-15
Build with Novita Model APIs
Create an API key and connect your application.