Gemini 3.5 Pro: What Became of Google’s Announced Flagship and What Shipped Instead

On May 19, 2026, Sundar Pichai took the stage at Shoreline Amphitheatre and did something no one expected: he shipped Gemini 3.5 Flash first and postponed the Pro version. The audience groaned audibly. For more than a decade, Google’s playbook had been clear: launch the flagship, then distill it into cheaper variants. This time, the "cheap" variant outperformed the previous flagship on nearly every benchmark that matters to developers. And the real flagship? "Give us until next month to get it to you," Pichai said. That month came and went: Gemini 3.5 Pro had not been released by mid-October 2026. This article shows what is known about the model, what Google shipped instead, and what the Flash and Omni releases reveal about Google’s rewritten AI strategy.

Gemini 3.5 Pro had not been released by mid-October 2026; Google missed the announced June date and named no new one. Bloomberg reported on 16 July 2026 that Google was months behind schedule on 3.5 Pro because it was taking time to improve the model’s capabilities, particularly in coding; according to ten current and former employees, many inside the company feared falling behind Anthropic and OpenAI. On 21 July Google introduced three new Flash models (Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber) and wrote that 3.5 Pro was still testing with partners and would be made broadly available as soon as it was ready; in parallel, its most ambitious pre-training run yet, for Gemini 4, had started. As late as 23 September 2026, Google DeepMind’s Gemini page listed 3.5 Pro as "coming soon"; the Gemini API model list as updated on 13 October 2026 contained no gemini-3.5-pro. On 13 October 2026 Google made Gemini 4 Argon generally available as its new flagship, via the API at an introductory price of 2 US dollars per million input and 10 dollars per million output tokens until 30 November 2026, then 4 and 20 dollars. The Flash line kept moving: 3.6 Flash on 21 July, 3.7 Flash on 13 August and 3.8 Flash on 2 September 2026; since 8 October 2026 the API routes requests for Gemini 3.5 Flash to Gemini 3.6 Flash (changelog).

The Flash surprise: why Google's reversed release order matters

Gemini 3.5 Flash was not supposed to exist in this form. Historically, Flash models were compromises: faster, cheaper, and measurably weaker than their Pro siblings. Gemini 3.5 Flash broke that contract. According to Google’s published benchmarks (signed by DeepMind CTO Koray Kavukcuoglu, Jeff Dean, Oriol Vinyals and Noam Shazeer), Flash scored 76.2 percent on Terminal-Bench 2.1, the industry-standard benchmark for agentic coding, six points above Gemini 3.1 Pro’s 70.3 percent. On MCP Atlas, which measures tool-use coordination across the Model Context Protocol, Flash hit 83.6 percent. According to Google it ran four times faster than comparable frontier models; Artificial Analysis measured roughly 200 tokens per second of output speed at the end of May (The Batch). Via the API it cost $1.50 per million input tokens and $9.00 per million output tokens at launch. Since 8 October 2026 Gemini 3.5 Flash has been deprecated, and the API routes its requests to Gemini 3.6 Flash.

Google has published little about the architecture behind this leap. The model card only says that Gemini 3.5 Flash is based on Gemini 3 Flash and refers to that model’s card for details; The Batch describes it as a mixture-of-experts transformer.

What made this strategically significant was not the benchmark score alone, but the signal Google sent by releasing Flash before Pro. In most earlier Gemini generations Pro came before Flash; Gemini 1.0 had no Flash tier at all, and Gemini 2.0 began with Flash (Google, December 2024). The reversal allowed two readings: either Google was confident enough in Flash that it no longer needed Pro to establish credibility, or Pro was facing engineering challenges that required more time. In May, Pichai’s words ("I know you can’t wait to get your hands on it. Give us until next month") suggested refinement, not crisis. In hindsight the second reading was right: according to Bloomberg, 3.5 Pro was months behind schedule in July because Google wanted to improve its capabilities, particularly in coding.

What is known about Gemini 3.5 Pro

By mid-October 2026 Google had published no benchmark numbers, pricing, model card or context window specification for Gemini 3.5 Pro. The confirmed facts fit into a short list:

• Announced launch window: June 2026 ("next month" from I/O on May 19), missed
• Status in May 2026: already used internally at Google; on 21 July 2026, according to Google, testing with partners
• Focus area: shared with Flash on coding and agentic capabilities
• Everything else: unconfirmed

In May, the Flash launch showed where Pro would have to stand out: not on coding speed, which Flash already saturated, and not on price, which Flash already owned, but on complex reasoning at extreme context lengths, long-horizon task consistency and deep multimodal synthesis. That is exactly where, according to multiple independent analyses, comparative tests between Flash and 3.1 Pro exposed the weaknesses of Flash: it won on Terminal-Bench and MCP Atlas but dropped on certain long-context retrieval benchmarks and on tasks requiring extended chains of abstract reasoning. That was not a flaw but the predictable trade-off of a distilled, speed-optimized architecture.

Leaks credited Pro with a 2 million token context window, Computer Use (direct desktop and OS interaction) and a "thinking mode" with per-request control of reasoning depth. Google confirmed none of this for 3.5 Pro. Computer Use came to Flash instead: on 24 June 2026 Gemini 3.5 Flash gained support for the API’s Computer Use tool in preview. The flagship role went to Gemini 4 Argon, which Google released on 13 October 2026 with a unified input-output limit of one million tokens via the Gemini API.

Gemini Omni: the video layer Google has been missing

While the developer community focused on Flash benchmarks, Google quietly launched something potentially more consequential for mainstream adoption: Gemini Omni. Introduced by Google DeepMind CTO Koray Kavukcuoglu as a model that “can create anything from any input”, starting with video (announcement), Omni is Google's first native world model with video generation and editing as its initial output modality.

The technical distinction between Omni and Google's existing Veo 3.1 video model is not subtle — and the confusion between the two has already caused planning problems for developers. Veo 3.1 is a specialized text-to-video engine: you input text, you receive cinematic video. It has been available in preview in the Gemini API since October 2025 and generally available on Vertex AI since 17 November 2025, with established API routes, pricing tiers (according to Google’s pricing page as of October 2026: Lite from $0.05/sec, Fast from $0.10/sec, Standard from $0.40/sec, each with audio), and enterprise integrations through Vertex AI. Omni is fundamentally different: it accepts any combination of text, images, audio, and video as input, maintains an internal world model with physics-aware simulation (gravity, fluid dynamics, kinetic energy), and enables conversational editing where users refine outputs through multi-turn dialogue rather than rewriting prompts from scratch.

The first shipping variant, Gemini Omni Flash, generated 10-second clips with synchronized audio at launch and rolled out immediately to the Gemini app, Google Flow, YouTube Shorts and YouTube Create. Since 30 June 2026 it has also been in public preview in the Gemini API, via the Interactions API, for 3 to 10 second clips at 720p with conversational editing; as of 23 September 2026 the model list also named keyframe interpolation and clip extension with native audio. Veo 3.1 remained available alongside it through the Gemini API (still listed on the pricing page in October 2026), so enterprise developers faced no forced migration. For creators, Omni brought a generational leap in editability.

Character consistency across scenes is the most technically impressive capability Google claims for Omni. Users can define a character once (via text description, reference images or a digital avatar created from their own likeness) and then place that character into any scene with preserved identity across lighting changes, camera angles and environmental shifts. If this works as demonstrated, it solves the single biggest quality problem in AI video generation: the "face swap" effect where generated characters morph between cuts. Independent testing was scarce at launch, but Google’s demo footage (claymation sculptures dissolving into bubbles, sci-fi scenes with coherent physics, music-synced visuals) pointed to genuine architectural progress beyond frame-by-frame generation.

The competitive landscape: what became of the June three-way race

In late May 2026, June looked like the most densely packed release month in AI history, with three flagships converging on the same window. This is how it played out:

• Gemini 3.5 Pro: Google’s delayed flagship did not ship. More Flash models came instead, and on 13 October 2026 Gemini 4 Argon took over the flagship role.
• GPT 5.6: OpenAI started it as a limited preview on 26 June 2026 and made it generally available on 9 July in three tiers (Sol, Terra, Luna); what the leaks had predicted is covered in our article on GPT 5.6.
• Claude Opus 4.8: available since 28 May 2026, with the then SWE-bench Pro record of 69.2 percent; Anthropic followed with Opus 5 on 24 July and Opus 5.5 on 22 September 2026.

The positioning that emerged in late May held only in part. Google took the speed and cost quadrant with its Flash line, but no 3.5 Pro for reasoning depth arrived by mid-October 2026. OpenAI shipped GPT-5.6 with about one million tokens of context rather than the leaked 1.5 million. And no vendor led across all dimensions at once; the top spot changed hands every few weeks.

For enterprise procurement, this fragmentation is healthy. It breaks the "default model" trap that held many companies in 2024 and 2025, when GPT-4 was the obvious choice for nearly everything. The right question is no longer "which model should we use?" but "which model for which task?", answered by routing logic that assigns tasks based on your own tests rather than vendor promises.

What this means for the German mid-market

From our work integrating AI systems into German mid-sized companies (manufacturers, law firms, e-commerce operations), the Gemini 3.5 family sends three signals:

First: the speed/cost breakthrough is real enough to act on. Gemini 3.5 Flash cost $1.50 and $9.00 per million tokens at launch and, according to Google, ran four times faster than comparable frontier models; that shifted the economics of high-volume applications such as customer-service agents or document processing. Anyone who waited for Pro lost months. Anyone who moved high-throughput applications to Flash has been routed to 3.6 Flash automatically since 8 October 2026, because the API redirects requests for 3.5 Flash. Only teams that secure their integration with their own test set notice silent model switches like this.

Second: multimodal video is approaching production viability. At launch, Omni was a tool for consumers and creators: 10-second clips, no API, no scene extension. Since 30 June 2026 Omni Flash has been in public preview in the Gemini API, and as of 23 September 2026 also with keyframe interpolation and clip extension with audio. That is enough to build first video workflows, with the caveat of a preview that features and terms may still change. For mid-sized companies, small low-risk experiments are the right entry point, such as product clips or internal training videos.

Third: Google's ecosystem integration remains its moat. Flash shipped simultaneously into the Gemini app, AI Mode in Search, Google Antigravity, Vertex AI, the Gemini API, Android Studio, and GitHub Copilot. No competitor matches this day-one surface coverage. For companies already embedded in Google Workspace, Cloud, or Android ecosystems, the friction of adopting Gemini is near zero. The strategic implication: if you are a Google shop, Flash is probably your default before you even evaluate alternatives.

What became of the open questions

In late May 2026, four open questions hung over Gemini 3.5 Pro and Omni. This is how they played out:

• The Pro benchmark drop. It did not come: by mid-October 2026 Google had published neither a model card nor benchmarks for 3.5 Pro.
• Computer Use confirmation. It came not with Pro but on 24 June 2026, as a preview for Gemini 3.5 Flash.
• Context window verification. The leaked 2 million tokens for Pro remained unconfirmed. Gemini 4 Argon was released on 13 October 2026 with a unified input-output limit of one million tokens via the API.
• Omni API availability. On 30 June 2026 Gemini Omni Flash arrived in the Gemini API as a public preview, via the Interactions API.

Conclusion: Flash became the product, Pro stayed a promise, Omni is the platform bet

Gemini 3.5 Flash was the most consequential Google AI release of the first half of 2026. It reversed a decade of "you get what you pay for" in model tiers, proving that distillation and architectural innovation can deliver flagship capabilities at commodity prices. It was not perfect: the reasoning regressions were real, and the "laziness" complaints of earlier Flash generations were only "largely" resolved. But it was good enough, cheap enough and fast enough to become the default for a huge swath of production workloads; since 8 October 2026 Gemini 3.6 Flash has handled its requests.

Gemini 3.5 Pro was set to be judged against a higher bar than any previous Google flagship, because its cheaper sibling had already set the baseline. It had not been released by mid-October 2026. According to Bloomberg, Google was above all working on its coding capabilities, and in July the pre-training run for Gemini 4 began. On 13 October 2026 Gemini 4 Argon took over the flagship role; Google named no new date for 3.5 Pro.

Gemini Omni is the longest-term bet of the three. Video generation is not yet a core business capability for most mid-market companies, but conversational multimodal editing ("create anything from any input") points toward a future where AI is not a tool you use but an environment you work within. The figures Google cited at I/O in May 2026, expected capital expenditure of roughly $180 to $190 billion for 2026, more than 900 million monthly active users of the Gemini app and over 3.2 quadrillion tokens processed per month, show that this platform bet is backed by resources few competitors can match.

For the German mid-market, the actionable insight is simple: plan with available models, not announced ones. For anything agentic, coding or high-volume, the latest Flash generation is the obvious route; for Pro-class tasks, Gemini 4 Argon has been available since 13 October 2026; and Omni video suits low-risk experiments. Models change every few months. The infrastructure you build around them should not have to.

Frequently asked questions about Gemini 3.5 Pro