OpenAI vs Anthropic vs Gemini Pricing: 2026 Per-Token Cost Comparison

TL;DR — who is cheapest?

There is no single cheapest provider — it depends on the tier. Anthropic Claude Sonnet 5 currently holds the frontier value crown at $2/$10 with 1M context and no long-context surcharge. Google Gemini 3.1 Pro ties it on input price ($2 ≤200K) and matches the 1M window. OpenAI GPT-5.6 Sol is the quality flagship, promo-priced at $4/$20 through at least November 2026 (standard $5/$30). At the budget tier, GPT-5.6 Luna ($0.20/$1.20) undercuts everything by a wide margin.

Frontier tier: GPT-5.6 vs Claude Sonnet 5 / Opus 5 vs Gemini 3.1 Pro

The frontier tier is where quality matters most and price gaps are real. At August 2026 rates, here are the per-million-token prices for the value picks at the frontier from each provider.

Anthropic repriced aggressively: Sonnet 5's $2/$10 launch price — already half the $3/$15 the Sonnet line carried since 3.5 — has been made permanent, and both Sonnet 5 and Opus 5 ($5/$25) carry 1M-token contexts billed at standard rates. Google's Gemini 3.1 Pro (still "Preview") matches Sonnet on input up to 200K but roughly doubles above that threshold. OpenAI's GPT-5.6 Sol is the most expensive per token even at its promo price, which runs through at least November 21, 2026 — but it's positioned as the quality flagship, alongside Fable 5 ($10/$50), Anthropic's maximum-capability model. If Claude produces a correct answer in one pass where a cheaper model needs two, the effective cost per correct answer can favor the pricier per-token rate. Always benchmark on your own eval, not on published model cards.

Budget tier: GPT-5.6 Luna vs Claude Haiku 4.5 vs Gemini 3.7 Flash

The budget tier is where most of your traffic should land after model routing. These models are 10-50x cheaper than the frontier tier and handle classification, extraction, summarization, and FAQ with adequate quality.

GPT-5.6 Luna is the absolute cheapest at $0.20/$1.20 — roughly 4× under Gemini 3.7 Flash even before Flash's promo expires on December 31, 2026 (its price then doubles to $1.50/$7.50). Haiku 4.5 sits highest on price but often edges the others on raw response quality for the tier, and keeps a 200K window (the other two run 1M). Between budget and frontier, OpenAI's GPT-5.6 Terra ($2/$12) covers mid-tier routing. The right choice depends on what your downstream pipeline expects.

Prompt caching: the 80-90% discount you should already be using

All three providers now offer prompt caching — if your prompt shares a prefix with a recent call (a system prompt, a shared document, few-shot examples), the cached prefix tokens are billed at a steep discount. For agent and RAG workloads that resend the same system prompt every turn, this is the single biggest cost lever after model selection.

The caveat is that each provider's cache has a minimum prefix length (1024-2048 tokens) and a TTL (5-60 minutes). Short prompts see no benefit; long, repeated prefixes see huge savings. Anthropic also charges a small write premium for the initial caching, but the net is still a large win on multi-turn agent traffic.

Batch API: 50% off for non-real-time workloads

If your workload tolerates latency (eval runs, content generation, data labeling, document processing), the batch APIs halve your cost. All three providers offer this with a 24-hour SLA.

The batch API stacks with prompt caching — a batch job with a shared system prompt gets both the 50% batch discount and the 80-90% cache discount on the prefix. For eval harnesses running thousands of test prompts with a shared instruction, this can cut eval cost by 80%+. See the cost optimization playbook for where batch fits in the full stack.

Long context: who handles 200K-2M tokens?

Context length no longer splits the providers the way it used to: all three now reach 1 million tokens. What splits them is the surcharge. Anthropic bills Sonnet 5 and Opus 5 at standard rates all the way to 1M. Google roughly doubles Gemini 3.1 Pro above 200K ($2→$4 input). OpenAI doubles input (1.5× output) above 272K. One caveat: Anthropic notes its newer tokenizer emits ~30% more tokens for the same text, so effective Claude costs run higher than nominal rates suggest.

Beware that long-context quality degrades on all models — the "lost in the middle" problem. Just because a model accepts 1M tokens does not mean it attends to all of them equally. Run your eval at the context lengths you actually use; do not assume a 1M context window gives you 1M tokens of reliable recall.

The case for multi-provider

Relying on a single provider is both a reliability risk (outages, rate limits) and a cost risk (you cannot benefit from price competition). A gateway like LiteLLM or Portkey lets you route by cost, fall back across providers on errors, and A/B-test quality continuously. The typical production setup uses GPT-5.6 Luna or Gemini 3.7 Flash for bulk traffic, Claude Sonnet 5 or Gemini 3.1 Pro for quality-sensitive traffic, and Claude Sonnet 5 again for long-context work — all behind a single gateway endpoint.

FAQ

Which is cheaper: OpenAI, Anthropic, or Gemini?

It depends on the tier. Frontier: Claude Sonnet 5 is cheapest ($2.00/$10.00, no long-context surcharge), tied on input by Gemini 3.1 Pro ($2.00/$12.00 ≤200K), with GPT-5.6 Sol promo-priced at $4.00/$20.00 (standard $5/$30). Premium frontier: Claude Opus 5 ($5/$25) and Fable 5 ($10/$50). Budget: GPT-5.6 Luna ($0.20/$1.20) is cheapest, then Gemini 3.7 Flash ($0.75/$3.75 promo, doubling in 2027), then Claude Haiku 4.5 ($1.00/$5.00). For long contexts above 200K, Claude Sonnet 5 is the cheapest frontier option.

How much does prompt caching save?

Prompt caching cuts the cost of cached prefix tokens by 80-90% (OpenAI: 90% off, Anthropic: 90% off on reads plus a 25% write premium, Google: 80-90% off plus an hourly storage fee). For workloads with repeated system prompts (agents, RAG), blended input cost drops 40-60%. Requires a minimum cacheable prefix (1024-2048 tokens).

How much does the batch API save?

50% off list price, across all three providers, with a 24-hour completion SLA. It stacks with prompt caching for compounded savings. Use it for evals, content generation, and any non-real-time workload.

Should I use multiple providers?

Yes — for reliability and cost optimization. A gateway (LiteLLM, Portkey) lets you route by cost, fall back on outages, and benchmark quality continuously. The typical setup: budget model for bulk, frontier model for quality, Claude Sonnet 5 for long context.

Related Deep Dives

• LLM Cost Optimization Playbook — the full 5-lever stack that builds on these prices
• Model Routing — how to send traffic to the cheapest viable tier
• Semantic Caching — eliminates the call entirely on repeated prompts
• LiteLLM vs Portkey vs OpenRouter — the gateways that enable multi-provider routing