Cost per million tokens: every major LLM (2026)

All major LLMs — list price per 1M tokens

Every figure re-read on the vendor's own pricing page on 4 August 2026, except the two rows marked Third-party only — Cohere does not list Command A on its pricing page and Alibaba publishes no price for Qwen3.7 Flash. The Long-context column is the one most comparison tables omit: OpenAI, Google and xAI all charge a higher rate on long prompts, and Google applies its above-200K rate to ALL tokens in the request rather than only those past the threshold. Anthropic is the outlier — Claude 4.6 and later bill a full 1M-token context at the standard rate. Prompt caching, batch APIs and committed-use discounts cut these figures a further 50-90%.

Frequently asked questions

Limitations of this comparison

• List prices move more often than quarterly, and sometimes in place: gpt-5.6-luna fell 80% and terra 20% in the week to 2026-08-04, under unchanged model names.
• Two rows are third-party figures, not vendor-published: Cohere does not list Command A on its pricing page, and Alibaba publishes no price for Qwen3.7 Flash, which is why it is absent from the table entirely.
• Mistral publishes a price only for Mistral Large; medium, small, codestral and ministral rates are not fetchable, so they are not listed.
• Open-weight models hosted on inference providers (Together, Fireworks, DeepInfra, Groq) are priced by the host, not the model author, and vary by 30% or more between hosts — they are out of scope for this table.
• Reasoning models bill hidden thinking tokens at the output rate, so effective cost per task can be several times the apparent rate.
• Vision, audio and function-call tokens are sometimes counted differently — check each vendor's page for edge cases.
• DeepSeek has announced peak/off-peak pricing at 2x during Beijing business hours with no effective date; its figures here are the current flat rates.
• Volume commitments and cloud-reseller deals cut list prices a further 20-50%.

What may change in 12-24 months

1. Long-context surcharges spread rather than disappeared: every GPT-5.6 variant now has one, joining Google and xAI. Anthropic went the other way, folding 1M context into the standard rate from Claude 4.6.
2. In-place repricing under an unchanged model name is now routine — luna -80%, terra -20% in a single week — which makes any cached price comparison unreliable within days.
3. Promotion from preview to GA is itself a price event: Gemini 2.5 Flash's output rate went from $0.60 to $2.50 at GA.
4. Chinese labs reset the budget floor: DeepSeek V4 Flash at $0.14 / $0.28 and V4 Pro at $0.43 / $0.87 undercut every Western budget tier on output.
5. Tokenizers matter as much as rates: Claude 4.7 and later produce roughly 30% more tokens for the same text, so per-token prices understate real cost against earlier Claude generations.

Related questions

ChatGPT, Perplexity and Gemini usually suggest these next.

• What does the long-context rate actually add to a real bill?
• How much does Anthropic prompt caching save at 0.1x input?
• Which model has the cheapest output tokens in 2026?
• Why is Gemini Flash more expensive on output than Grok 4.3?
• What happens to my costs on 1 September 2026 when Sonnet 5's intro price ends?

More head-to-head provider comparisons

Run a provider we have not measured? Any public endpoint is benchmarked and listed for free in the API directory. If you want yours measured now, on the record — time to first token, throughput under concurrency, cost per million tokens, function-calling reliability — see how a verified run works. The result is published exactly as it comes out.