The Lab
LLM Cost Comparator
Per-million-token pricing across current Anthropic, OpenAI, and Google models, read directly off each vendor's own pricing page — not an aggregator. Static data, because published pricing changes a few times a year rather than per visitor; a live Worker call here would be overengineering, not correctness.
Prices as of August 11, 2026 · 16 models · sources: Anthropic , OpenAI , Google
Sorted cheapest blended cost first. Blended $/MTok assumes 3 input tokens per 1 output token, the ratio the public comparators standardised on. It is a comparison aid, not a forecast — a coding agent runs far more input-heavy than that, a drafting tool far less. Use the calculator for a real workload.
Per-model footnotes (12)
- GPT-5.6 Luna: Cheapest model in this table. Short-context rate; long context is $0.40 in / $1.80 out. Cut from $1.00/$6.00 since the previous refresh — an 80% reduction.
- Claude Haiku 4.5: Uses the previous tokenizer — see the tokenizer note below the table.
- Gemini 3.6 Flash: Output price includes thinking tokens. Context storage billed separately at $1.00 per 1M tokens per hour.
- Gemini 3.5 Flash: Output price includes thinking tokens.
- Claude Sonnet 5: The $2/$10 launch price is now the standard price — the increase to $3/$15 previously scheduled for Sep 1, 2026 will not happen. This table said otherwise until 2026-08-11.
- GPT-5.6 Terra: Short-context rate. Long context is $4 in / $0.40 cached / $18 out. Cut from $2.50/$15 since this table's previous refresh.
- Gemini 3.1 Pro Preview: Output price includes thinking tokens. Cached input now has a published per-token read rate; context storage is still billed separately at $1.00 per 1M tokens per hour.
- Claude Sonnet 4.6: Uses the previous tokenizer — see the tokenizer note below the table.
- Claude Opus 5: Fast mode (research preview) is priced separately at $10 / $50 and is not available with the Batch API.
- GPT-5.6 Sol: Short-context rate. Long context is $10 in / $1 cached / $45 out. Cache writes $6.25/MTok.
- Claude Fable 5: 5-minute cache writes $12.50/MTok, 1-hour $20/MTok.
- Claude Mythos 5: Limited availability.
Price a workload
Pick a shape or enter your own volume. Cache hit rate applies to input tokens only, at each model's published cached rate — for agent workloads it moves the total more than the model does.
What this table can't tell you
A dollar per token is not a dollar per task
Anthropic's Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text than Claude Sonnet 4.6 and earlier. Every vendor tokenizes differently, and none of them tokenize your text the same way. Two models at an identical $/MTok can therefore differ by a third on the same input, and no comparison table — including this one — can show that. Price the workload, not the token.
Cached input is where the real money is
A cache read costs a tenth of standard input at Anthropic and Google and a tenth at OpenAI too. For an agent replaying a long system prompt on every turn, the cache-hit rate moves the bill more than the choice of model does. The calculator below takes a hit rate for that reason.
Batch is roughly half price and rarely considered
Anthropic and Google publish batch rates at about 50% of standard, shown here where published. If the work is not interactive — evaluations, backfills, bulk classification — the cheapest model at standard rates is frequently more expensive than a better model run as a batch.
This table shows list prices only
Enterprise agreements, committed-use discounts, and provisioned throughput all change the number, and none of them are public. Regional or data-residency endpoints add roughly 10% at all three vendors. Treat this as the starting point of a negotiation, not its conclusion.