Imagine giving the same difficult research folder to three colleagues. The first reads every page, compares the footnotes, studies the graphs, and returns a careful argument. The second moves quickly, calls whatever tools are needed, and turns the folder into a working dashboard before lunch. The third handles a thousand similar folders at once, extracting the same twelve fields from each one without exhausting the budget.
Those three colleagues are a useful way to understand Google’s current Gemini family. Gemini 3.1 Pro is the first: the deliberate analyst. Gemini 3.7 Flash is the fast builder and agent. Gemini 3.5 Flash-Lite is the inexpensive production line. The word Pro does not automatically make the first model best for every job; it tells you which kind of compromise Google originally designed it to make.
The reasoning flagship that aged in public
Google released Gemini 3.1 Pro in Preview on February 19, 2026. Its headline improvement was not a new chat interface or another bundle of Google services. It was a large increase in core reasoning. On ARC-AGI-2—puzzles that require discovering unfamiliar visual rules rather than recalling facts—Google reported 77.1%, compared with 31.1% for Gemini 3 Pro. It also recorded 94.3% on GPQA Diamond and 44.4% on Humanity’s Last Exam without tools.
Artificial Analysis reached a similar broad conclusion from its own evaluation suite. At launch, 3.1 Pro scored 57 on its Intelligence Index, took first place, and led six of the ten component evaluations. Artificial Analysis measured roughly 114 output tokens per second: not Flash-fast, but unusually quick for a top reasoning model at that moment. It also praised the economics relative to contemporary flagships, because Google kept the familiar price of $2 per million input tokens and $12 per million output tokens for requests at or below 200,000 tokens.
That February context matters. A benchmark score is a photograph stamped with a date, not a permanent championship belt. Six months of model releases have shifted the field. Gemini 3.1 Pro has not become less intelligent, but competitors and Google’s own Flash models have become better at long-horizon coding, computer use, and enterprise workflows.
Where 3.1 Pro still earns its place
The model is at its best when a problem is difficult before it becomes repetitive. It can receive text, images, audio, video, and PDF files in one prompt, with up to one million input tokens and 64,000 output tokens. A scientist can place a paper beside microscopy images. An investigator can compare a recorded interview with a timeline and financial report. A product team can inspect a screen recording alongside the relevant source code.
This is analysis, not native media generation. Gemini 3.1 Pro produces text. Veo creates video, Nano Banana creates images, and Lyria creates music. Google may connect those systems inside a product, but describing all of them as abilities of the 3.1 Pro model is like crediting a conductor with personally playing every instrument in the orchestra.
The million-token window also deserves a realistic explanation. It is an enormous worktable, not a photographic memory. Google’s own MRCR test gives 3.1 Pro an 84.9% average at 128k but only 26.3% on the pointwise 1M setting. If the answer depends on one sentence buried near page 4,000, simply uploading everything and hoping is poor information architecture. Narrow the corpus, label the sources, ask for page-level evidence, and verify important claims.
What happened to Gemini 3.5 Pro?
The succession story has become more revealing than a normal missed launch date. On May 19, Google announced Gemini 3.5 Flash and said 3.5 Pro was already being used internally and would roll out “next month.” June ended without a release. On July 16, Bloomberg reported that the model was months behind schedule because Google was trying to improve its capabilities, particularly coding, and that a late-June training-data refresh had not met expectations.
Five days later, Google launched Gemini 3.6 Flash and 3.5 Flash-Lite. Its language about Pro had changed: 3.5 Pro was “testing with partners” and would become broadly available “as soon as it’s ready.” In the same announcement, Google said it had already begun its most ambitious pretraining run for Gemini 4. By August 14, the public Pro page still showed Gemini 3.1 Pro under a small promise that 3.5 Pro was “coming soon.”
There is now a credible industry report from SemiAnalysis that 3.5 Pro has been quietly canceled. The surrounding facts make the report plausible: the June window was missed, the model was reportedly struggling to meet coding goals, two newer Flash generations shipped, Gemini 4 training began, and there is still no public 3.5 Pro endpoint. But plausible is not proven. Google’s latest public statements have not announced a cancellation.
The most accurate conclusion is therefore conditional. The original Gemini 3.5 Pro plan has plainly slipped. The model carrying that exact name may be canceled, renamed, folded into Gemini 4, or held until a substantial rework is competitive. Users should not buy a subscription or design a system around an unreleased promise. If Google ships it, evaluate the actual endpoint; until then, “coming soon” is marketing, not availability.
The ecosystem did not wait
While Pro remained in the workshop, the Flash line took over much of Google’s practical roadmap. Gemini 3.7 Flash, released August 13, is now generally available and is Google’s recommended migration target even for 3.1 Pro workloads involving code generation, multimodal reasoning, multi-step agents, and design adherence. It supports a million-token context, 64k output, built-in tools, and adjustable low, medium, or high thinking effort.
The performance table explains why Google calls it a workhorse. It records 43.6% on FrontierCode 1.1, 65.3% on DeepSWE v1.1, 85.8% on Terminal-Bench 2.1, 30.4% on AutomationBench, and 34.0% on GDP.pdf. Its Artificial Analysis Intelligence Index is 56—one point below 3.1 Pro’s launch score, though cross-date scores should be compared cautiously—and independent testing cited by Google places output around 340 tokens per second. Through December 31, 2026, it costs $0.75 per million input tokens and $3.75 per million output tokens; the standard $1.50/$7.50 rate begins in January.
Gemini 3.5 Flash-Lite serves a different layer of the stack. It is stable, accepts the same broad set of input media, offers a 1,048,576-token input limit, and costs $0.30/$2.50. Google cites roughly 350 output tokens per second. Its natural jobs are document parsing, classification, receipt extraction, search subagents, and generating many candidates in parallel. Google reports 54% on Terminal-Bench 2.1 versus 31% for 3.1 Flash-Lite, 72.2% versus 60.1% on long-context MRCR, and 1140 versus 642 on GDPval-AA v2. That does not turn Lite into Pro; it makes Lite a remarkably capable conveyor belt.
Which Gemini should you actually use?
Use Gemini 3.1 Pro when your own evaluation shows that its deeper reasoning produces better scientific analysis, research synthesis, or handling of unfamiliar problems. Keep it if a carefully validated workflow depends on its particular behavior. It remains a serious model, not an antique.
Use Gemini 3.7 Flash as the default for new coding agents, web and UI generation, tool-heavy automation, and high-volume knowledge work. Its combination of current agent performance, speed, stable availability, and discounted price is difficult for 3.1 Pro to answer.
Use Gemini 3.5 Flash-Lite behind the scenes, where one orchestrating model delegates many small, well-specified tasks. A cheap model extracting invoice dates ten thousand times can create more business value than a flagship composing one magnificent paragraph.
And wait for evidence before treating Gemini 3.5 Pro as anything at all. The lesson of its delay is not that Google has stopped building powerful models. It is that model names are promises, while endpoints are products.
