LLM Token Pricing
Checked —. Bought directly from the model makers, output tokens cost from $— per million (—) to $— (—) across — current models. Paid plans for one person run $— to $— a month, and team seats start at $— per user. Through clouds and aggregators, — of — listings matched to a direct model charge the same per token as the maker.
Cost per task
Describe one unit of work in tokens, say which models you treat as equal in quality, and this ranks every place you can buy them by what that work costs. Pick a sample task or enter your own numbers.
Loading…
Measured token usage by model
Output tokens (reasoning included) each model used to run the whole benchmark suite, and its score on that run. The median model is 1.00×; with scaling on, a model at 2.00× is costed at twice the output tokens you entered.
Loading…
What this ranking assumes
- Measured token usage
- With scaling on, the output tokens you enter are what the median model writes; each model is costed at that times its own output on the benchmark relative to the median, so a verbose reasoning model is charged for the tokens it really writes. Models the benchmark did not measure keep your number and are flagged. Input tokens are not scaled. The benchmark's mix of tasks may not match yours, and usage depends on the reasoning setting it measured.
- Minimum score
- A score above 0 keeps only models whose benchmark score reaches it and drops models the benchmark did not score. Combined with "All tiers", this is a measured version of "equal quality".
- Equal quality
- It assumes every model in the comparison does the task equally well, and with the same number of tokens. "One model, every channel" is the only true like-for-like comparison. A tier is where each maker positions a model in its own lineup, not a quality score, so a tier mixes models of very different ability; for a real decision, use "Models I pick" with the models your own evaluation found equivalent. Reasoning models often write several times more output tokens for the same task, so measure a sample of your own work and enter real token counts.
- Context window
- A listing drops out when one call's prompt plus its share of the output would not fit its recorded context window. Where no window is recorded, prompts over 128,000 tokens are flagged.
- What a task costs
- Uncached input × input price, plus cached input × cached price, plus output × output price, at list prices. Batch discounts apply only when you tick batch processing and the channel offers one.
- Long prompts
- When one call's prompt (input tokens ÷ calls per task) passes a vendor's long-context threshold, the whole task is priced at the long-context rate. Rows whose long-context terms could not be read are flagged for prompts over 100,000 tokens.
- Left out
- Cache-write fees, per-call tool fees such as web search, seat and subscription fees, data-residency surcharges, off-peak discounts and negotiated discounts. Subscriptions are not ranked because they have no per-task price; where a plan states its included usage in dollars, divide that by the cost per task to see how many tasks it covers.
- Sample tasks
- Illustrative token counts, not measurements: chat reply 2,000 in / 400 out; documents answer 12,000 in / 500 out with 40% cached; long-document summary 50,000 in / 1,500 out; classification 1,000 in / 20 out with 60% cached; coding agent task 1.5M in / 40,000 out over 30 calls with 90% cached; one very long prompt 400,000 in / 4,000 out.
Per-token API prices
US dollars per million tokens at list price: standard tier, default region, no batch discount. Input is what you send and output is what the model writes, including any reasoning tokens. Cached input is a repeated prompt prefix served from cache; cache write is the one-time cost of storing it where the vendor charges for that.
Price spread of each model bought direct
Loading…
Long entries are cut to five lines. Click a row, or press Enter on it, to read it in full.
Loading…
Resellers vs buying direct
The same model bought through a cloud marketplace, an aggregator or another of the maker's own channels, next to the price on the maker's own API. A positive difference costs more than buying direct. Committed-use discounts, enterprise agreements and provisioned throughput are negotiated and not shown. Channels that bill in their own units, such as Snowflake credits, are converted to dollars at the list rate named in each row's notes in the table above, so their difference moves with that rate.
Loading…
Open-weight models across hosts
Models such as Llama and gpt-oss have no price from their maker, so they are compared here against the cheapest channel that sells them. Hosts can run different builds of the same model (precision, context length, speed), so check each row's notes in the table above before switching.
Loading…
Plans and seats
Subscriptions, seats and credit bundles from the model makers and from products that resell their models. Prices are US list prices per month with monthly billing; the table also gives the monthly equivalent of annual billing. Personal plans usually allow the vendor to train on your conversations unless you opt out, while team and enterprise plans exclude business data by default; the Data and terms column gives each vendor's wording.
Who each plan is for, and how it bills
- Free tier
- Flat subscription: fixed fee, rate limits only
- Subscription + credits: allowance, then pay for more
- Per-seat + usage
- Usage-based: pay as you go
- Custom contract
Loading…
Long entries are cut to five lines. Click a row, or press Enter on it, to read it in full.
Loading…
History and changes
Every run is kept as a dated snapshot of what the vendors' own pages said that day, and the history is rebuilt from all of them on each run. Changes compare each snapshot with the one before it, only for provider groups researched in both, so a group that was skipped in one run shows as a coverage gap rather than as prices disappearing.
Loading…
Price over time, bought from the maker
Loading…
Change log
Caveats and method
Loading…
How to read this report
- Token
- The unit models bill by. A million tokens is roughly 750,000 words of English text.
- Input and output
- Input is the prompt, files and history you send; output is what the model writes. Reasoning ("thinking") tokens bill as output.
- Cached input
- The discounted price for prompt tokens the vendor has already cached from an earlier request with the same prefix.
- Cache write
- A surcharge some vendors add the first time a prompt prefix is stored in the cache.
- Batch discount
- The price cut for jobs submitted to run asynchronously, usually returned within 24 hours.
- Flat subscription
- A fixed fee with no per-use charges. Usage is capped only by rate or fair-use limits, which vendors rarely state in tokens. This is the closest thing to all-you-can-eat.
- Subscription + credits
- A fixed fee that includes a usage allowance (credits, premium requests or dollars of model usage). More usage is billed per use or bought in packs.
- Per-seat + usage
- A per-user fee plus metered usage billed separately.
- Usage-based
- Pay as you go with no subscription.
- Custom contract
- Priced by negotiation, usually with an annual commitment or minimum seats.
- Verified
- The figure was read from the vendor's own pricing page or documentation on the snapshot date. Rows from secondary sources are marked Unverified in the Source column.
- Not covered
- Taxes, non-US prices, negotiated and committed-use discounts, provisioned throughput, and audio, image, video and embedding prices.