Claude vs GPT: what the APIs actually cost in 2026

Updated August 21, 2026 · Prices per 1M tokens, USD

Anthropic and OpenAI both ladder their models into three tiers, and as of August 2026 the matched tiers are priced surprisingly close. The differences show up when you look at your workload's input/output mix — and especially once prompt caching enters the picture.

Tier-by-tier sticker prices

*Sonnet 5 is on a promotional rate through August 31, 2026 (standard $3 / $15). Both vendors bill cached input at ~10% of the input rate and offer a 50% batch discount.

The mid-tier is where most money is spent

For a typical production chatbot (500 input / 300 output tokens per request, 30K requests a month), Sonnet 5 runs about $120/month against Terra's $138/month — near parity, with Sonnet's edge coming entirely from its cheaper output ($10 vs $12). Flip to a RAG workload (4,000 in / 500 out) and the gap narrows further, because input dominates and both charge $2.

At the bottom, OpenAI is far cheaper

The budget tiers aren't comparable: GPT-5.6 Luna at $0.20/$1.20 costs roughly a fifth of Claude Haiku 4.5 at $1/$5. If your task is classification, extraction, routing or short summaries — and a small model clears your quality bar — Luna (or Gemini 2.5 Flash-Lite, or DeepSeek V4 Flash) is the price-performance play. Anthropic's lineup currently starts higher.

At the top, it depends what "top" means

Claude Fable 5 ($10/$50) costs about 65% more than GPT-5.6 Sol ($5/$30) per token. Whether it's worth it is a quality question only your evals can answer — but note that for agentic coding workloads with huge inputs (20K in / 3K out), the monthly gap at 2K requests is roughly $600 vs $345. If Fable solves tasks in fewer iterations, the per-task cost can still favor it.

Caching changes the ranking

Both vendors bill cache hits at ~10% of input. A chatbot with a large, stable system prompt hitting 80% cache turns Sonnet 5's effective input rate into $0.56/M — at which point Sonnet undercuts several nominally cheaper models that lack published cached pricing. Run your own numbers before trusting any static ranking, including this one.

Prices compiled from vendor pricing pages, last verified August 21, 2026. Not affiliated with Anthropic or OpenAI. Verify current rates on official pages before committing.