Ask ten people which model is "best" and you'll get ten different answers built on benchmark scores. Ask which model is cheapest for the job you actually need done, and most people are guessing — because the spread in per-token pricing across vendors right now is enormous, and it doesn't move in a straight line with capability. We pulled current, published pricing directly from seven providers' own documentation to see what a million tokens really costs in mid-2026.
The Headline Numbers
Every vendor prices input and output tokens separately, and every vendor spans a flagship-to-budget range internally. Here's a representative slice — each provider's top reasoning tier alongside its cheapest general-purpose tier:
| Vendor | Model | Input / 1M | Output / 1M |
|---|---|---|---|
| OpenAI | GPT-5.5 Pro | $30.00 | $180.00 |
| OpenAI | GPT-5.4 Nano | $0.20 | $1.25 |
| Anthropic | Claude Fable 5 | $10.00 | $50.00 |
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 |
| Gemini 3.1 Pro (≤200k) | $2.00 | $12.00 | |
| Gemini 2.0 Flash | $0.10 | $0.40 | |
| xAI | Grok 4.3 | $1.25 | $2.50 |
| Mistral | Magistral Medium | $2.00 | $5.00 |
| Mistral | Ministral 3 (3B) | $0.10 | $0.10 |
| DeepSeek | DeepSeek-V4 Pro | $0.435 | $0.87 |
| DeepSeek | DeepSeek-V4 Flash | $0.14 | $0.28 |
| Z.ai | GLM-5.2 | $1.40 | $4.40 |
| Z.ai | GLM-4.7 FlashX | $0.07 | $0.40 |
| Groq | Llama 3.3 70B (hosted) | $0.59 | $0.79 |
| Groq | Llama 3.1 8B Instant | $0.05 | $0.08 |
Prices pulled directly from each vendor's published pricing documentation, per million tokens, standard (non-batch, non-priority) tier.
The spread is the story. GPT-5.5 Pro's output price ($180/M) is roughly 2,250 times DeepSeek-V4 Flash's ($0.08/M cache-hit — cheaper again if you count the cache-hit rate). Even comparing like-for-like "budget" tiers, Groq's hosted Llama 3.1 8B at $0.05/$0.08 undercuts OpenAI's cheapest flagship-family model, GPT-5.4 Nano, by 4x on input.
The Output Tax
Every vendor charges more for output than input — but the multiplier varies a lot, and it's the number that actually predicts your bill on reasoning-heavy workloads, since those models think in output tokens before they answer. Claude Opus 4.8 charges 5x for output over input. GPT-5.5 charges 6x. Gemini 3.1 Pro charges 6x. DeepSeek-V4's output multiplier is closer to 2x. That means a model that looks cheap on its input price can still be expensive in practice if your workload is reasoning- or generation-heavy rather than input-heavy — a document-summarisation task (lots of input, little output) and a code-generation task (little input, lots of output) can invert which model is actually cheaper for the same headline pricing.
The Open-Weight Discount Is Real
DeepSeek, Z.ai's GLM family, and the open models Groq hosts are not just cheaper than the Western flagship labs — they're cheaper by an order of magnitude at the budget end, and increasingly competitive on benchmarks at the flagship end too. GLM-4.7 FlashX at $0.07/$0.40 and DeepSeek-V4 Flash at $0.14/$0.28 (cache miss — a fraction of that on a cache hit) put serious inference within reach of workloads that would be uneconomical on a frontier proprietary model. The trade-off is real too: these models generally trail the top proprietary reasoning tiers on the hardest benchmark tasks, and enterprise buyers weigh data-residency and support commitments differently across a US lab, a European lab (Mistral), and Chinese-developed open models (DeepSeek, GLM) — that calculus belongs in the decision, not just the price column.
Cheapest Isn't the Same as Cheapest-to-Solve-the-Task
A $0.05/M model that needs three retries and a longer prompt to get a reliable answer can cost more per solved task than a $5/M model that gets it right first time with a tight prompt. This is the trap in comparing sticker prices alone: total cost per completed task — not per token — is what should drive model selection, and that number depends on accuracy, retry rates, and how many output tokens a model burns thinking before it answers.
How ApiSpi Handles a Market That Won't Sit Still
This is exactly why the ApiSpi LLM Gateway was built provider-agnostic from day one, in the same spirit as the specialist inference layer we covered in our piece on the inference wars. The gateway now routes across Anthropic, Google Gemini, OpenAI, Mistral, DeepSeek, Groq, Grok, and Z.ai — bring your own connector for whichever provider fits a given workload's cost and quality bar, model routing rules can send specific keywords to a specific engine, and every request is metered consistently regardless of which vendor answered it. As this table reshuffles — and at the rate this market moves, it will — the model your agents use is a config change, not a migration.