The one page with numbers on it
Model facts
Every hard figure on this site lives here, so there is exactly one page to keep current instead of a hundred. If an explainer quotes a number, it links back to this table.
Last verified 11 August 2026 against each vendor's own documentation, including a re-check of the Anthropic rows that produced one correction, noted below. Anything not verified is flagged in orange.
How to read this
Context window is how much the model can hold in mind at once, input and output together. Max output is the ceiling on a single reply. Price is per million tokens, quoted separately for what you send (input) and what you get back (output) — output almost always costs several times more.
Anthropic (Claude)
| Model | API ID | Context | Max output | Input / MTok | Output / MTok | Reliable cutoff | Training cutoff |
|---|---|---|---|---|---|---|---|
| Claude Fable 5 | claude-fable-5 |
1M | 128k | $10 | $50 | Jan 2026 | Jan 2026 |
| Claude Opus 5 | claude-opus-5 |
1M | 128k | $5 | $25 | May 2026 | May 2026 |
| Claude Sonnet 5 | claude-sonnet-5 |
1M | 128k | $2 | $10 | Jan 2026 | Jan 2026 |
| Claude Haiku 4.5 | claude-haiku-4-5 |
200k | 64k | $1 | $5 | Feb 2025 | Jul 2025 |
Correction, 11 August 2026. This page previously flagged
Sonnet 5's $2 / $10 rate as introductory, said the standard rate was
$3 / $15, and told you to check whether the promotion had ended before
relying on it. On re-checking, that is no longer true. Anthropic's pricing
documentation now states that the $2 / $10 pricing, "announced at launch as
introductory pricing through August 31, 2026, is now the standard price,"
and that "the previously scheduled increase to $3 / $15 per million
input/output tokens on September 1, 2026 will not occur." The warning has
been removed and the asterisks with it.
Sources: Anthropic models overview
and pricing documentation,
both fetched 11 August 2026. Batch processing and prompt caching change these
numbers substantially and are documented on that pricing page.
Two cutoff columns: the reliable date is how far the
model's knowledge is dependable; the training date is the outer edge
of its data. The months between are thin, half-known territory — note Haiku's
five-month gap. Explained properly in
Training cutoff.
OpenAI, Google and Meta
| Vendor | Model | API ID | Context | Max output | Input / MTok | Output / MTok |
|---|---|---|---|---|---|---|
| OpenAI | GPT-5.6 Sol | gpt-5.6-sol |
1.05M | 128k | $5 | $30 |
| OpenAI | GPT-5.6 Terra | gpt-5.6-terra |
1.05M | 128k | $2 | $12 |
| Gemini 3.6 Flash | gemini-3.6-flash |
1.05M | 64k | $1.50 | $7.50 | |
| Meta | Muse Spark 1.2 | muse-spark-1.2 |
not published | not published | $1.25 | $4.25 |
| Meta | Llama 4 / Llama 3 | — |
n/a | n/a | no first-party price | no first-party price |
Verified 11 August 2026 against each vendor's own
documentation: OpenAI's
pricing and
models pages,
Google's Gemini API pricing
and its
Gemini 3.6 Flash
model page, and Meta's
Model API pricing.
Selection: current general-purpose text models, not the full
catalogue. Each vendor sells cheaper small models and more expensive
reasoning tiers; this table exists to make the shape of the market legible,
not to be exhaustive.
Context figures: Google publishes an exact input limit of
1,048,576 tokens and an output limit of 65,536; OpenAI publishes 1.05M and
128K. The first two are the same number written two ways.
Cutoffs: OpenAI publishes a knowledge cutoff of
16 February 2026 for the GPT-5.6 family. Google's model page for Gemini 3.6
Flash does not publish one, which is why there is no cutoff column here.
Tiering: Google prices some models differently above and
below a prompt-size threshold, and offers batch, flex and priority rates that
differ substantially from the standard rate quoted. Meta publishes a cheaper
"contributor" tier. Anyone budgeting seriously should read the vendor's own
page rather than this row.
Two honest gaps in the table above
Meta doesn't fit the shape of this table, and that's the finding rather than an omission. Llama models are downloaded and run by you or by whichever host you choose, so Meta publishes no per-token price for them at all; what you pay depends entirely on where you run them. Meta's paid API sells different models (the Muse family), and its published pricing page gives rates without stating context or output limits, so those two cells say "not published" rather than carrying a number from somewhere else.
A price comparison across vendors is less meaningful than it looks. Identical per-token rates do not imply identical cost for the same job, because tokenisers differ, so the same text is a different number of tokens for different vendors. Treat the columns as order-of-magnitude guidance, not a league table. → Tokens
What changes, and how fast
Rough guide to how much you should trust a figure of a given age:
- Prices: can change with no notice. Introductory rates expire, or quietly become permanent, and an old page will not tell you which happened. Treat anything over three months old as unreliable.
- Model names and IDs: new models land every few months and old ones get retired. An ID from a year-old tutorial may simply error.
- Context windows and output limits: slower moving, and generally only go up.
- Concepts: stable for years. What a token is has not changed. This is why the explainers on this site avoid numbers and link here instead.
Concepts behind these numbers → Context windows