Gemini 3.5 Flash: The Pricey Flash That Now Routes to 3.6

Gemini 3.5 Flash is an odd model to write a guide for, because if you call it today you are probably not getting it. Google's API documentation now marks gemini-3.5-flash as deprecated and says requests to it are automatically routed to gemini-3.6-flash. It is no longer selectable in Google AI Studio.

It is still worth understanding, for two reasons. Plenty of configuration files, agent harnesses and cost spreadsheets still name it. And the way it was priced and then replaced says a lot about how to treat the Flash line as a whole.

A Flash model priced like a Pro

Google released 3.5 Flash at I/O on 19 May 2026, and it skipped the usual -preview stage, going straight to general availability in the Gemini app, Search's AI Mode and the Gemini API. The specification was standard for the generation: 1,048,576 input tokens, 65,536 output tokens, text, image, audio, video and PDF in, text out, with thinking levels to trade quality against cost and latency.

The price was not standard. It launched at $1.50 per million input tokens and $9.00 per million output, with cached input at $0.15. As Simon Willison noted on launch day, that was three times the price of Gemini 3 Flash Preview and not far below Gemini 3.1 Pro at $2.00 and $12.00.

The per-token price understated the real gap. Willison reported that running Artificial Analysis's benchmark suite cost $1,551.60 on 3.5 Flash at high thinking against $892.28 on 3.1 Pro Preview. Appwrite's write-up, citing the same source, attributed most of that to verbosity: around 73 million output tokens across the suite against 57 million for 3.1 Pro. A model that is cheaper per token but writes more tokens can end up costing more per task. The token cost of reasoning models explains why hidden reasoning output is where that money goes.

What the extra money bought

Google's model card compared 3.5 Flash with Gemini 3 Flash and 3.1 Pro, and the agentic gains were large:

• Terminal-bench 2.1: 76.2%, against 58.0% for 3 Flash and 70.3% for 3.1 Pro
• MCP Atlas: 83.6%, against 62.0% and 78.2%
• OSWorld-Verified: 78.4%, against 65.1% and 76.2%
• Finance Agent v2: 57.9%, against 42.6% and 43.0%
• SWE-Bench Pro (public): 55.1%, against 49.6% and 54.2%

On multi-step tool use, then, a Flash model was beating Google's own Pro preview. That was the pitch, and on Google's numbers it held.

Where it fell short

The same table showed the limits. On Humanity's Last Exam 3.5 Flash scored 40.2% against Pro's 44.4%, and on ARC-AGI-2 72.1% against 77.1%. Long-context retrieval was the clearest weakness: MRCR v2 at 128K came in at 77.3%, well behind 3.1 Pro's 84.9%. At the full 1M window both models sat in the mid-twenties, which is a reminder that a 1M context window does not mean 1M tokens of reliable recall.

Its knowledge cutoff was January 2025, which is old for a model released in mid-2026. For anything involving recent library versions, it needed documentation in context to be reliable.

Deprecated: your calls already go to 3.6 Flash

On 21 July 2026, two months after launch, Google released Gemini 3.6 Flash. At launch it kept the $1.50 input price but cut output to $7.50, and Google reported it used 17% fewer output tokens on the Artificial Analysis Index. Google's pricing page now lists 3.6 Flash at an introductory $0.75 input and $3.75 output through 31 December 2026, with the rates doubling from 1 January 2027.

The deprecation is handled by silent redirection, and that is the part that deserves attention. A request naming gemini-3.5-flash does not fail; it is served by a different model. That is convenient, and it is also a behaviour change you did not deploy. Prompts tuned against 3.5 Flash's verbosity, output-length limits set for it, and evaluation baselines recorded against it all now describe a model you are not running.

If you reach it through an aggregator, check what you are billed. OpenRouter's catalogue still listed google/gemini-3.5-flash at $1.50 input and $9.00 output at the time of writing, alongside a :batch variant at half that, which is the same model on Google's asynchronous Batch API. Whether a third-party listing passes through Google's routing and its new price is something to verify on your own invoice rather than assume.

What to do if 3.5 Flash is still in your config

The fix is small, but do it deliberately rather than letting the redirect decide for you.

1. Search your code and config for the model string, including environment files and agent definitions. Redirected names hide well.
2. Choose the replacement explicitly. Google recommends gemini-3.6-flash or the newer gemini-3.8-flash. The Gemini 3.6 Flash guide covers the like-for-like replacement, and the Gemini 3.8 Flash guide covers the stronger but deliberately more token-hungry option.
3. Re-run your regression set against the model you are actually getting, and replace any baseline recorded on 3.5 Flash. Agent regression suites covers how to keep that cheap.
4. Re-check output limits and budgets. A model that writes fewer tokens may finish tasks your old limits were cutting off, and a model that writes more will hit them sooner.

The broader lesson is about the Flash line itself. Google shipped 3.5, 3.6, 3.7 and 3.8 Flash between May and September 2026, and two of those four, 3.5 and 3.7, now redirect to newer models. Treat a Flash model string as something that will move under you, and pin and test accordingly.