Key Takeaways
- Official per-token prices (Sep 2026): Claude Sonnet 5 $2 / $10 per million tokens (input/output), Claude Opus 5 $5 / $25, Gemini 2.5 Flash $0.30 / $2.50, GPT-5.6 Sol $4 / $20, GPT-5.6 Terra $2 / $12, and GPT-5.6 Luna $0.20 / $1.20.
- Output-to-input price ratios range from 5x to about 8.3x in this comparison; the multiplier is not universal across providers or models.
- Batch processing cuts Claude prices in half; prompt-cache hits cost 10% of the base input rate on most Claude models.
- We could not access our own API billing dashboards in this session, so we do not claim to know our exact monthly spend — that number is not verified here. What we can verify is the official rate card itself.
“How much does GPT cost per token” returns dozens of aggregator pages, but most either round the numbers or mix outdated tiers together. We pulled the current rate cards directly from Anthropic’s and Google’s own pricing documentation (checked September 2026) and laid them side by side with the tiers we actually route work through, so the comparison reflects what a real multi-model operation pays, not a single vendor’s best case.
Official Rate Card (checked Sep 2026)
| Model | Input / MTok | Output / MTok | Output ÷ Input |
|---|---|---|---|
| Claude Opus 5 | $5.00 | $25.00 | 5.0x |
| Claude Sonnet 5 | $2.00 | $10.00 | 5.0x |
| Claude Haiku 4.5 | $1.00 | $5.00 | 5.0x |
| GPT-5.6 Sol | $4.00 | $20.00 | 5.0x |
| GPT-5.6 Terra | $2.00 | $12.00 | 6.0x |
| GPT-5.6 Luna (post-cut) | $0.20 | $1.20 | 6.0x |
| Gemini 2.5 Flash | $0.30 | $2.50 | 8.3x |
Sources: the official Anthropic, OpenAI, and Google pricing documentation cited below. GPT-5.6 figures reflect the tier names we route work through in our own operation, not a full OpenAI catalog.
Direct Calculation: What One Article Actually Costs
Take a realistic content task: research and draft a ~1,500-word article, roughly 4,000 input tokens (source material, prior context, instructions) and 3,000 output tokens (the draft itself). At list price, no caching, no batching:
| Model | Input cost | Output cost | Total |
|---|---|---|---|
| Claude Opus 5 | 4,000 × $5/1M = $0.020 | 3,000 × $25/1M = $0.075 | $0.095 |
| Claude Sonnet 5 | 4,000 × $2/1M = $0.008 | 3,000 × $10/1M = $0.030 | $0.038 |
| Gemini 2.5 Flash | 4,000 × $0.30/1M = $0.0012 | 3,000 × $2.50/1M = $0.0075 | $0.0087 |
| GPT-5.6 Luna | 4,000 × $0.20/1M = $0.0008 | 3,000 × $1.20/1M = $0.0036 | $0.0044 |
Opus 5 costs roughly 21.6x more per article than GPT-5.6 Luna, and about 2.5x more than Sonnet 5 — for a single draft. That gap compounds fast at volume: 1,000 articles a month is a ~$57 difference between Sonnet 5 and Opus 5 alone, before counting research passes, revisions, or QA calls that also consume tokens. This is exactly why routing cheap, high-volume steps (search, first-pass drafts, classification) to Flash- or Luna-tier models and reserving Opus-tier models for judgment calls that are expensive to get wrong is a cost decision, not just a quality one.
Why the Output-to-Input Ratio Matters More Than the Sticker Price
The models in this comparison price output tokens at 5–8.3x the input rate. That means a task’s cost is driven far more by how much the model has to write than by how much context you feed it. For the same model, a long output can outweigh a much larger input in the bill — a distinction that sticker-price comparisons (“Model X is $Y per million tokens”) hide when they quote a single blended number.
Two Lenses
Through the buyer’s lens
The list price is the start of the estimate, not the end. Two workloads with identical token totals can differ several-fold in cost depending on how much of the total is output and how much of the input is cache-eligible. Before comparing vendors, split your own workload into input, output, and repeated-context tokens — the split moves the bill more than the vendor choice does.
Through the rate-card’s lens
Across all three vendors in our table, output stays at 5–8.3x the input rate — no vendor sells output at parity. Any comparison quoting one blended per-token number is therefore making a hidden assumption about your output share, and it flatters whichever model that assumption favors. Reading the two columns separately is the only vendor-neutral way to use a rate card.
What Would Change Our View
Prompt caching and batch processing can cut these numbers sharply — Claude’s batch API halves both input and output prices, and cached input reads cost 10% of the base rate. If a workload is repetitive enough to cache well, the effective cost gap between tiers narrows. We have not measured our own cache-hit rate in this piece; that would change the real-world total more than the list price does.
FAQ
Q. Why is GPT/Claude output priced so much higher than input?
A. The official rate cards charge different prices for input and output. They show 5–8.3x output-to-input ratios for the models compared here, but the rate cards do not by themselves establish one universal technical cause.
Q. Is Gemini Flash always cheaper than Claude or GPT?
A. At list price for equivalent token counts, yes among the tiers compared here. But “cheaper per token” isn’t the same as “cheaper per task” if a smaller model needs more retries or produces a lower-quality first draft that then needs a second, pricier pass to fix.
Q. How much does it actually cost to run an AI agent for a month?
A. It depends entirely on volume and which tier handles which step — there’s no single answer. Multiply your expected input/output tokens per task by the rate card above, by task count, per model actually used; that is the only reliable estimate.
Sources
- Claude API Pricing — Anthropic, official pricing page, checked Sep 2026
- OpenAI API Models — OpenAI, official model pricing, checked Sep 2026
- Gemini Developer API Pricing — Google, official pricing page, checked Sep 2026
Leave a Reply