Gemini 2.5 Flash pricing: 1M context at budget rates

Where it sits in the catalog

On the history-carrying chatbot above, Gemini 2.5 Flash runs $71/mo. The same workload costs $490/mo on GPT-5.4 (6.9×) and $628/mo on Claude Sonnet 4.6 (8.8×). That is not a marginal discount — it is a different order of magnitude from the frontier tier, which changes what you can afford to build rather than just trimming a line item.

1M context at budget pricing

The 1000K window is the headline feature and the reason this model shows up in architecture decisions rather than only cost ones. Long-context and retrieval are substitutes: if you can fit the corpus in the prompt, you can sometimes delete the retrieval pipeline — and with it the extra LLM calls, the vector database, and the reranking step that a RAG architecture drags along.

That trade only works when the per-token price is low enough to make stuffing context affordable, which is precisely the combination here. Worth modelling both shapes before committing to a retrieval stack.

The tokenizer bills slightly under baseline

Gemini’s SentencePiece tokenizer produces roughly 3% fewer tokens than o200k_base for typical English — a ratio of 0.97. It is a small effect next to the price difference, but it points the same direction, and it is the only model in the catalog that bills below the baseline rather than above it.

Caching without a write premium

Cached reads come back at $0.03/1M, or 10% of standard input, and there is no charge to warm the cache. Combined with the low base rate, the absolute savings per cache hit are small simply because the starting price is already low — at these rates, caching is a refinement rather than the main lever.