Transparent LLM pricing across providers.

How to read this table

• Input and Output are USD per 1 million tokens, list price from each provider.
• Context is the maximum prompt + completion length the provider advertises.
• Prices are indicative — collected on the date below, before any volume discounts, cached-input pricing, or SchneeAI margin. Use the Cost Calculator to model real requests end-to-end.
• Providers change prices often. Verify against the provider’s pricing page before signing an SLA.

Frontier & premium tier

The models teams reach for when quality is the priority.

Value tier

Fast, cheap, and good enough for most classification, summarization, and routing work.

Reasoning tier

Models that emit chain-of-thought before answering. Higher latency, higher cost, materially better on math, code, and multi-step reasoning.

Open-source tier

Open-weight models routed through SchneeAI’s provider partners. Prices vary by hosting provider.

Observations

• Premium convergence. GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro cluster within a 2-3× band on output tokens. Choice is increasingly driven by capability and ecosystem fit, not unit price.
• The value tier is shockingly cheap. At $0.075 / 1M input, Gemini 1.5 Flash costs less than a cent to summarize a 10K-token document. Most workloads should default here and escalate only when quality demands it.
• Reasoning models are expensive — but per-task, not per-token. o1’s $60 / 1M output looks scary, but for a single hard coding or math problem it’s cheaper than an engineer’s time. Don’t apply it to bulk workloads.
• Context windows are a pricing axis. Gemini 1.5 Pro’s 2M context lets you skip a retrieval stack — but at 2M tokens of input you’re paying $2.50 per request. Long context is a tool, not a freebie.
• Open-source models closed the gap. DeepSeek V3 at $0.27 / $1.10 is competitive with GPT-4o-mini while benchmarking closer to GPT-4o-class on many tasks. Worth routing to when the workload allows.
• Cached input pricing changes the math. Anthropic and Google both offer prompt caching at roughly 10% of the normal input price for cache hits. If you re-use a system prompt or long context, this can dwarf the headline number.

What SchneeAI adds on top

Per-token provider prices are the floor. The Gateway charges that price plus a transparent margin, and in return you get:

• Unified auth and routing — same Bearer JWT, same /v1/chat/completions for every model.
• Budget enforcement — per-feature, per-tenant caps that fail closed.
• Structured usage — service / tenant / user / model / tokens / cost, queryable via /v1/usage.
• Vault — encrypted raw prompts and outputs, with separate retention.
• PII scanning — 17 categories before the request leaves your boundary.
• Audit trail — every request, every prompt change, every policy update.

For most teams, the margin is cheaper than building and operating the equivalent controls in-house. Run the numbers with the build-vs-buy TCO widget.

Estimate real requests

Headlines lie. A model that’s 3× cheaper per token can be 10× more expensive for your workload if it requires a longer prompt or more retries.

Use the Cost Calculator to:

1. Paste your actual prompt.
2. Pick a model (or auto).
3. See per-request and 30-day projected cost.
4. Compare 3-5 models side by side.

Update policy

• Cadence: monthly, first week of the month.
• Last verified: 2026-07-09.
• Source: each provider’s public pricing page on the date above.
• Changelog: major price changes (≥25%) noted on the changelog.

If you spot a price that’s drifted, email [email protected] — we’ll verify and update within one business day.