
On 12 November 2026 at 18:00 UTC, Google DeepMind published Gemini 4, its fourth-generation flagship foundation model, codenamed Hydrogen, and opened it the same day through the Gemini API and Vertex AI. By the company's own figures, the model scores 92.4 on GPQA Diamond and 78.2 on SWE-bench Verified, doubles the context window from one million to two million tokens, and cuts inference cost per million tokens by roughly 45%. It is the first Gemini launch where efficiency, rather than parameter
At 18:00 UTC on 12 November 2026, Google DeepMind published Gemini 4, its fourth-generation flagship foundation model, codenamed Hydrogen, on its official blog alongside a roughly 40-page technical report. Total parameter count, training token volume and data mix were not disclosed.
Three points carried the announcement. On architecture, Gemini 4 drops the dense-plus-MoE hybrid of Gemini 3 for fully sparse expert routing, activating about one twelfth of its parameters per token, against one eighth in Gemini 3. DeepMind says this is where the Hydrogen codename comes from: the lightest possible dispatch for the largest usable capacity. On context, the window doubles from one million to two million tokens, with a stated 98.6% needle-in-a-haystack recall at full length, plus native video timestamp grounding that answers questions such as what appeared at minute 37, second 12. On cost and
Gemini 3 shipped in November 2025, making Gemini 4 a near-exactly one-year successor. Over that year the industry's main storyline shifted from whose model is smarter to whose model runs cheaper. The reason is straightforward: as enterprise customers moved large models from pilots into production systems during 2026, spending changed from one-off licences to持续的 per-token billing, and inference cost appeared directly on finance department spreadsheets.
The technical preconditions were already in place. Sparse expert routing, hybrid attention combining sliding windows with global layers, and KV-cache compression all matured across 2025 and 2026, making it viable to grow total parameters while
Liu Gong's take: the decisive number in Gemini 4 Hydrogen is not a benchmark score but an invoice line. Within six months, if Google prices the Flash variant below $0.30 per million tokens, a meaningful share of multimodal agent teams will move their inference backend onto the Gemini API. Within twelve months, a two-million-token context will stop being a selling point and become table stakes, and evaluation will move from whether a document fits to how much multi-hop reasoning survives once it does.
Start with the numbers most readers will skip. Google states a 45% cut in inference cost per million tokens against Gemini 3 Pro, while list prices fall from $2.00 to $1.25 for input and from $8.00 to $5.00 for output, both a 37.5% reduction. The seven-to-eight percentage points between the cost curve and the price cut are the real commercial move: Google did not hand the entire efficiency gain to customers, keeping roughly one sixth on its own books to offset margin pressure in cloud AI. On a single call it is invisible; across billions of daily calls it is money that lands directly in reported earnings.
From a technical standpoint, the Hydrogen codename points at routing rather than parameters. Activating about one twelfth of parameters per token, down from one eighth, means each accelerator card can serve more requests. But the familiar weakness of sparse routing has not disappeared: expert load skews more easily across long sequences, and the more routing decisions a sequence requires, the wider the tail-latency variance becomes. Google quotes median first-token latency falling from 0.70 to 0.42 seconds; a good median is not a good P99, and P99 is what actually stalls enterprise systems. The 98.6% recall at two million tokens proves a single point can be retrieved, not that cross-document coreference or temporal causality hold up, and the technical report offers no comparable figures for those harder cases.
On the business side, the repricing rewrites the anchor for the whole API market rather than Google's own rate card. At the Pro tier, $5.00 per million output tokens pulls well clear of the previous flagship's $8.00. For teams doing long-document analysis, repository-scale refactoring or video archive retrieval, the per-task cost shifts from a decision to a default. Liu Gong expects at least one major vendor to follow with an output-side cut within two quarters, because output tokens are the hardest part of the cost structure to compress and whoever moves first captures agent-scenario volume.
For the industry, the immediate effect of a two-million-token window is that the boundary of RAG retreats. Teams that spent half a year on chunking, embedding and reranking can now collapse part of that stack into load-the-whole-thing-plus-one-rerank. But Liu Gong's observation from industrial inspection is that real inputs there are not long texts: they are tens of hours of multi-camera, multi-station footage with timestamps. Two million tokens hold roughly a dozen hours of compressed video features, enough for shift-level review, not week-level traceability. Production lines typically run 30 to 60 units per minute and demand end-to-end verdicts at hundred-millisecond scale, so 0.42 seconds of first-token latency still cannot close an online control loop. Hydrogen's first landing point in factories is likely offline re-judgement and defect root-cause analysis, not in-line detection.
- Context window doubles from one million tokens in Gemini 3 to two million, with a stated 98.6% needle-in-a-haystack recall at full length.
- Inference cost per million tokens falls about 45% versus Gemini 3 Pro, while API prices drop from $2.00 to $1.25 for input and from $8.00 to $5.00 for output, both 37.5%, leaving roughly eight percentage points of margin between cost and price.
- GPQA Diamond reaches 92.4 against 87.1 previously, a 5.3-point gain; SWE-bench Verified reaches 78.2 against 71.5, a 6.7-point gain; Google notes external reproduction is still pending.
- Average activated parameters per token fall from one eighth to about one twelfth, putting per-token compute at roughly
For the industry, the first three months will be dominated by price transmission. At $5.00 per million output tokens, Gemini 4 Pro forces mid-tier model vendors into a binary choice: match the cut and surrender margin, or retreat into vertical custom delivery. Over twelve months, long context becomes an entry ticket rather than a differentiator, evaluation communities shift from needle recall to multi-hop reasoning and long-range consistency, and the phrase it fits stops carrying a narrative. Competitive weight moves further onto two engineering metrics: cost per token and P99 latency.
For ordinary users, the visible change within three months is better handling of long documents and
Competitors stayed quiet. Neither OpenAI nor Anthropic issued a formal response on launch day, while researchers at Anthropic focused their social commentary on expert load balancing in sparse routing, framing it as a harder engineering problem than raw parameter count. Developer reaction was more direct: the highest-traffic threads on Hacker News and X centred on pricing and the two-million-token window, with one developer working through the arithmetic that feeding a two-million-token document in whole and billing output at list price still lands in the double-digit dollars per task, which points to enterprise rather than personal use as the first market. Third-party evaluator Artificial
Questions worth exploring:
- Expert load skew in fully sparse routing at long context: what a 40% median latency gain costs at P99
- The eight points between a 45% cost cut and a 37.5% price cut: how Google balances cloud AI margin against share
- Where a two-million-token window displaces the RAG stack, and where retrieval pipelines become more necessary rather than less
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from DeepMind Blog (published 2026-11-12 18:00:50). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.
