Google bet everything on a rebuild from scratch

Gemini 3.5 Pro was supposed to arrive in June, according to the original promises from Google I/O (May 19, 2026). When Sundar Pichai told the audience "give us until next month," few expected that instead of delivering the model, Google would take a radical step — discard the entire Gemini 2.5 Pro base model and start a completely new training cycle from scratch. According to Geeky Gadgets and Tech Times, the reason was three areas where the previous architecture hit a ceiling: mathematical reasoning, SVG scene generation, and overall image quality.

The cost of such a decision is measured in hundreds of millions of dollars and months of GPU time. But Google paid it — which signals how far behind the original candidate was, or how high the competition has raised the bar.

The situation was further complicated by a wave of departures of key researchers. Noam Shazeer, co-author of the groundbreaking 2017 paper "Attention Is All You Need" and co-lead of the Gemini team, announced his departure to OpenAI on June 18. A day later, Nobel laureate John Jumper followed, leaving DeepMind after nine years for Anthropic. Together with two other senior researchers, these departures wiped roughly 225 billion dollars from Alphabet's market value on June 22 (a 5% drop in a single session).

Two million tokens: marketing or a real breakthrough?

The rebuilt Gemini 3.5 Pro is, according to leaks, expected to offer a context window of 2 million tokens — double that of Claude Opus 4.8 and GPT-5.5. In practice, this means the ability to process roughly 1.5 million words in a single prompt: an entire large codebase, a year's worth of meeting transcripts, or a multi-volume research dataset.

But there's a catch. Transformer attention scales quadratically with sequence length, so doubling the context doesn't mean double the computational cost — it's orders of magnitude higher. And researchers from Stanford and other institutions have documented the phenomenon where model performance drops significantly for information placed in the middle of very long contexts. Until independent testers publish results from long retrieval benchmarks, the 2-million-token window is more of a marketing number than a verified specification.

The model is also expected to offer a Deep Think mode (equivalent to extended thinking from competitors) for multi-step logical reasoning. The estimated price is around 15 dollars per million input tokens and 60 dollars per million output tokens — roughly ten times more expensive than Gemini 3.5 Flash, which is already generally available and represents a more sensible choice for most developers.

For Czech developers, it's significant that Gemini 3.5 Flash supports Czech at a very good level, and the same is expected from the Pro version — Google has long listed Czech among the supported languages for its AI models.

DeepSeek V4: extremely cheap, but with a legal burden

DeepSeek V4-Pro entered preview mode on April 24, 2026 — the same day OpenAI launched GPT-5.5. The timing was no coincidence. The model uses a Mixture-of-Experts (MoE) architecture: out of a total 1.6 trillion parameters, only 49 billion are active per token. The rest remain idle. This makes running the model dramatically cheaper — 0.87 dollars per million output tokens compared to roughly 25 dollars for Claude Opus 4.7 and 30 dollars for GPT-5.5.

For developers running the model on their own infrastructure (MIT license), the API cost disappears entirely. The V4-Flash model fits into 160 GB — with light quantization, it's within reach of a powerful local server.

The price difference isn't the whole story. The independent DeepSWE benchmark, which uses a more precise verifier than the widely-used SWE-bench Verified, reveals a fundamental performance gap: DeepSeek V4-Pro achieves 8% success rate (pass@1) compared to 70% for GPT-5.5 and 54% for Claude Opus 4.7. On the BenchLM leaderboard, V4-Pro ranks 29th out of 33 evaluated models and is explicitly classified as a "non-frontier model."

Security warning for European companies

For Czech and European companies considering deploying DeepSeek, it is essential to understand the legal framework in which the company operates. DeepSeek is run by the Chinese company Hangzhou DeepSeek Artificial Intelligence Co., Ltd. China's National Intelligence Law (2017), Article 7, requires all Chinese organizations and citizens to "support, assist, and cooperate with national intelligence work" — without exceptions, without a court order.

This obligation applies regardless of where DeepSeek's servers are physically located, regardless of published privacy policies, and regardless of contractual terms between DeepSeek and its users. For any sensitive data, regulated industries, or GDPR-bound environments, the hosted API is not suitable.

The solution is self-hosting: download the model weights under the MIT license and run inference on your own infrastructure. This eliminates the risk of cross-border data flows. However, the behavioral censorship embedded in the model weights (CrowdStrike documented that R1 refused to generate code related to certain politically sensitive topics in approximately 45% of tested cases) remains regardless of the deployment method.

What you must do by July 24

For every developer using DeepSeek's hosted API, there is one urgent task: update code that calls the deepseek-chat or deepseek-reasoner aliases no later than July 24, 2026, 17:59 CEST. After this deadline, these aliases will start returning errors — with no announced extension.

The migration itself is trivial (changing one parameter to deepseek-v4-pro or deepseek-v4-flash, same URL, same API key), but there's one catch: the deepseek-reasoner alias maps to V4-Flash (in "thinking" mode), not to V4-Pro. Developers who used deepseek-reasoner for demanding logical tasks and assume that swapping the alias will preserve the same level of capabilities will end up on a significantly weaker model. Those who relied on reasoning must explicitly switch to V4-Pro.

Another practical tip: since the launch of the stable version, DeepSeek will charge double the price during Beijing peak hours (9–12 and 14–18 China time). For European developers, this means that batch processing started during our nighttime will avoid peak rates.

Grok 4.5: no independent numbers yet

SpaceXAI (the entity formed by merging xAI with SpaceX in February 2026) is also entering the July convergence. Grok 4.5 has been in closed beta at SpaceX and Tesla since June 28 on a 1.5-trillion-parameter V9 model. Elon Musk internally described its performance as "close to, possibly exceeding Opus" — meaning Claude Opus, currently the highest-rated model on the Artificial Analysis Intelligence Index.

However, no independent benchmark has yet tested Grok 4.5. Even for the current Grok 4.3, there are no publicly verified numerical results. Manufacturer self-evaluation without external validation carries significant epistemic uncertainty — the AI industry has documented cases where self-reported scores differed from independently measured results by more than 20 percentage points.

What the July convergence means for Czech developers

The fact that three major model events are converging in the same week is no coincidence — it reflects the pace at which AI boundaries are shifting. For Czech development teams, several specific recommendations follow:

If you use DeepSeek API: migrate to the new aliases immediately. It's a one-line code change, but the deadline is uncompromising. For sensitive data, consider switching to self-hosting or another provider.

If you're waiting for Gemini 3.5 Pro: use Gemini 3.5 Flash in the meantime, which is fully available, fast, and supports Czech. Design your architecture so that the eventual arrival of the Pro version is a bonus — not a dependency. Google has already missed deadlines twice this year.

If you're choosing a stack for the second half of 2026: the decision window has shrunk from quarters to days. Models that today look like a clear choice may be surpassed in two weeks — and vice versa.